A learning repository for understanding post-training reasoning data: what it is, how it is built, how it is verified, how it enters training, and how to audit it.
Awesome LLM Reasoning Data is designed as a field-learning repository, not just a paper list. If you want to understand large-model post-training reasoning data, you should be able to start here, learn the core vocabulary, follow a reading path, inspect representative papers, open paper cards, and gradually build a mental model of the whole area.
The repository is organized around one practical question:
When a model becomes better at reasoning after post-training, what data record, feedback signal, verifier, reward, environment, or judge actually made that possible?
To answer that, the repo combines four layers:
- 🧭 Learning guides that explain concepts and reading paths.
- 📚 Paper maps that organize work by subfield rather than by publication date.
- 🗂️ Cards that summarize data objects, construction recipes, verifiers, risks, and links.
- 🔎 Searchable structured metadata so readers can filter by verifier type, training use, curation level, and artifact availability.
Companion paper: A Primer in Post-Training Reasoning Data.
Project website: Awesome-LLM-Reasoning-Data.github.io.
Ask the Atlas: source-grounded AI assistant · launch pending.
| Date | Update | Why it matters |
|---|---|---|
| 2026-06-15 | Promoted the atlas to 165 verified entries, 87 linked cards, and 53 L5 audit-ready cards. | The front page can now be attractive without hiding uncertainty or inventing unverified links. |
| 2026-06-15 | Completed two artifact-verification rounds: 41 code, 27 data, 20 Hugging Face, and 25 project links are pinned. | Readers can jump from a paper to reusable artifacts when official sources exist. |
| 2026-06-15 | Rebuilt the searchable site, paper atlas, exports, QA reports, contribution workflow, and release notes from structured metadata. | Counts and public reports are reproducible instead of hand-maintained. |
The atlas is intentionally conservative about metadata. If a paper/code/data/project link is not verified locally, it stays in reports/needs_search.md instead of being promoted into the verified lists.
Most Awesome lists stop at "which papers exist." This atlas is organized around the extra questions that make post-training reasoning data reusable: what data object is released, what verifies it, how it enters training, and how a reader should audit the claim.
Build the shared vocabulary before opening dense primary papers.
| Track | Subfields | Best for | Entries | Jump |
|---|---|---|---|---|
| 🧭 Foundations & Primers | 🧭 Post-training surveys 🧠 Reasoning LLM surveys 📦 Data documentation / datasheets 🧪 RLHF / reward-model surveys 🌐 Agent data / tool-use surveys 🧯 Contamination / evaluation surveys |
beginners building the field map before primary papers | 84 | Papers |
Compare the actual records: demonstrations, preferences, verifiable outcomes, process labels, rollout traces, agent episodes, and rubrics.
| Track | Subfields | Best for | Entries | Jump |
|---|---|---|---|---|
| 🧱 Instruction / Demo / Rationale | 🧱 Instruction tuning / SFT data 🧑🏫 Human demonstrations 🤖 Synthetic instruction data 🧠 Chain-of-thought / rationale data 🔁 Self-training / STaR ✂️ Long/short CoT distillation |
demonstration, SFT, CoT, rationale, and teacher-trace data | 58 | Papers |
| 🤝 Preference & Reward Feedback | 🤝 Human preference data / RLHF ⚖️ DPO / preference optimization 🎚️ Scalar reward / ORM data 🤖 RLAIF / synthetic feedback 🧪 Reward-model benchmarks 🧾 Rubric-conditioned rewards |
RLHF, DPO, reward modeling, rubric rewards, and AI feedback | 73 | Papers |
| 🧮 Programmatic Verification | 📐 Math answer-verifiable data 🧮 Math RLVR datasets 💻 Code execution / unit-test data 🧾 Formal proof / Lean / theorem proving 🧪 Verifier robustness and answer extraction 🧰 Programmatic benchmarks |
math, code, proof, and answer-verifiable reasoning data | 94 | Papers |
| 🪜 Process / Trace Supervision | 🪜 Human step-level labels 🧪 Process reward models 🔁 Rollout-value supervision 🛠️ Automatic process supervision ❌ First-error localization 📊 PRM benchmarks and evaluation |
step-level labels, PRMs, rollout values, and first-error signals | 25 | Papers |
| 🔁 Rollout / Search / TTC Trace | 🎲 Multiple rollouts / best-of-N 🌳 Search trees / MCTS 🔎 Rejection sampling traces 🧠 Self-consistency / repeated sampling ⏱️ Test-time compute logs ✂️ Long2short / distill-from-search |
search-generated candidates, best-of-N, pass@k, and test-time compute traces | 39 | Papers |
| 🌐 Environment & Agent Trajectories | 🛠️ Tool-use data 🌍 Web/browser agents 📱 App/mobile agents 🖥️ OS/desktop agents 🧑💻 SWE/repository agents 🔁 Replayable trajectory data 🧰 Agent benchmarks and terminal predicates |
tool, web, OS, app, SWE, and replayable environment data | 95 | Papers |
| ⚖️ Judgment / Rubric / Domain Expert | ⚖️ LLM-as-judge data 🧑⚖️ Human/expert judgment 🩺 Medical reasoning / health rubrics 🛡️ Safety reasoning data 🧾 Factuality / grounding ⚖️ Legal reasoning 🏦 Financial reasoning 🧪 Rubric reward models |
LLM judges, expert rubrics, factuality, safety, medical, legal, and finance reasoning | 83 | Papers |
Follow the lifecycle from construction recipes to training use, scaling, benchmarks, frontier disclosures, and failure audits.
| Track | Subfields | Best for | Entries | Jump |
|---|---|---|---|---|
| 🏗️ Construction & Open Releases | 🧱 Prompt sourcing ✍️ Teacher trace generation 🔎 Rejection sampling / search-generated data 🔁 Self-play / self-improvement 🧪 Filtering and verifier refresh 🏗️ Open reasoning data releases 🧬 Data lineage and release metadata |
building, filtering, releasing, and reproducing reasoning datasets | 108 | Papers |
| 🎯 Training Usage & Objectives | 🧱 SFT / instruction tuning 📚 Distillation ⚖️ Preference optimization 🎚️ Reward modeling / ORM 🪜 PRM / process supervision 🏋️ RLVR / verifier RL 🌐 Agent training 🧪 Evaluation / reranking / audit |
how data enters SFT, DPO, RM, PRM, RLVR, agents, evaluation, and audit | 97 | Papers |
| 📈 Scaling / RLVR / TTC | 📈 Data scaling 🔁 Data reuse and uniqueness ⏱️ Test-time compute 🎲 pass@k / sampling budget 🧪 Verifier scaling 🏋️ RLVR optimization scaling 🔍 Scaling attribution |
data scale, RLVR, verifier scaling, pass@k, and inference budget claims | 90 | Papers |
| 🧰 Benchmarks & Evaluation | 📐 Math benchmarks 💻 Code benchmarks 🧾 Proof benchmarks 🌐 Agent benchmarks ⚖️ Rubric/domain benchmarks 🧪 Reward-model benchmarks 🧯 Live / contamination-resistant benchmarks |
evaluation surfaces and reusable feedback contracts | 109 | Papers |
| 🚀 Frontier Disclosure Ledger | 🚀 DeepSeek-R1 family 🌙 Kimi reasoning reports 🐉 Qwen reasoning/math/code reports 🧠 Magistral / Phi / Nemotron style reports 🧪 RLVR recipe reports 🧬 What is disclosed vs hidden |
reading frontier reports as partial data-recipe disclosures | 40 | Papers |
| 🧯 Audit & Failure Modes | 🧯 Benchmark contamination 🔍 Search-time contamination 🧬 Hidden lineage / teacher leakage 🎮 Reward hacking 🧪 Verifier gaming ⚖️ LLM-as-judge attacks 🧨 Spurious rewards 📉 Reproducibility failures |
leakage, contamination, verifier gaming, judge attacks, and reproducibility failures | 68 | Papers |
Only entries with verified official primary links are listed here. If a subfield still lacks verified entries, it is explicitly marked as needs search instead of receiving invented links. The data and feedback hints tell you whether a paper is the right kind of post-training reasoning-data work to open next.
beginners building the field map before primary papers · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 🧭 Post-training surveys | field-level maps of post-training, reasoning models, and data-centric LLM practice | A Comprehensive Survey of Reward Models: Taxonomy, Applications, Challenges, and Future (2025, arXiv) data: taxonomy of reward-model data sources, objectives, applications, evaluations, and cha…; feedback: reward model as proxy objective for downstream post-training. A Survey on Human Preference Learning for Large Language Models (2024, arXiv) data: preference-centered taxonomy over feedback data, preference modeling, preference usag…; feedback: human preference signal transformed into reward, preference loss, or evalua… Reinforcement Learning for LLM Post-Training: A Survey (2024, arXiv) data: technical survey comparing RLHF and RLVR policy-gradient style post-training methods.…; feedback: learned preference rewards, verifiable rewards, and policy-gradient objecti… |
survey taxonomy hides concrete data objects |
| 🧠 Reasoning LLM surveys | reasoning-model lineages, claims, and recurring evaluation patterns | A Survey of Reasoning with Foundation Models (2025, arXiv) data: survey taxonomy and literature map.; literature survey.; feedback: metadata pending Reasoning with Large Language Models, a Survey (2024, arXiv) data: survey taxonomy and literature map.; literature survey.; feedback: metadata pending |
model-centric framing obscures data and verifier details |
| 📦 Data documentation / datasheets | datasheets, data statements, lineage, license, and release metadata | Data statements for natural language processing (2018, TACL) · Card data: survey background; feedback: metadata pending Datasheets for datasets (2018, arXiv) · Card data: survey background; feedback: metadata pending |
reusable data lacks provenance or consent context |
| 🧪 RLHF / reward-model surveys | background linking preference data, reward models, and reasoning rewards | A Comprehensive Survey of Reward Models: Taxonomy, Applications, Challenges, and Future (2025, arXiv) data: taxonomy of reward-model data sources, objectives, applications, evaluations, and cha…; feedback: reward model as proxy objective for downstream post-training. A Survey of Reinforcement Learning from Human Feedback (2023, TMLR) data: survey taxonomy over feedback collection, reward modeling, and policy optimization.;…; feedback: learned reward model from human feedback. |
generic alignment lessons are over-applied to verifiable reasoning |
| 🌐 Agent data / tool-use surveys | orientation for tools, web tasks, OS tasks, and repository agents | A Survey on Evaluation of LLM-based Agents (2025, arXiv) data: survey taxonomy for agent evaluation tasks and environments.; process: task, environm…; feedback: environmental and benchmark evaluators summarized by the survey. |
agent traces are treated as transcripts rather than replayable episodes |
| 🧯 Contamination / evaluation surveys | reproducibility, contamination, model collapse, and benchmark refresh | LiveBench: A challenging, contamination-free benchmark for large language models (2024, arXiv) · Card data: answer level; feedback: programmatic, mixed Language Model Developers Should Report Train-Test Overlap (2024, arXiv) data: overlap and reporting analysis.; process: training corpus, evaluation set, overlap es…; feedback: overlap analysis rather than a reward model. |
benchmark deltas are accepted without overlap checks |
demonstration, SFT, CoT, rationale, and teacher-trace data · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 🧱 Instruction tuning / SFT data | instruction-response examples, demonstrations, and target behavior records | Magicoder: Empowering code generation with OSS-instruct (2024, ICML) · Card data: instruction-response coding example, often linked to a code reference or task scaffol…; feedback: coding benchmark pass rates and optional executable checks. Tulu 3: Pushing frontiers in open language model post-training (2024, arXiv) · Card data: instruction-response examples, preference pairs, verifiable task outputs, and model-e…; feedback: mixture of preference labels, reward models, and verifiable rewards dependi… Self-Instruct: Aligning language models with self-generated instructions (2023, ACL) · Card data: answer level; feedback: mixed OpenMathInstruct-2: Accelerating AI for math with massive open-source instruction data (2024, ICLR) · Card data: problem-solution pair with natural-language mathematical reasoning and final answer.;…; feedback: answer checks and benchmark evaluation over math tasks. |
prompt sources and mixture weights are hidden |
| 🧑🏫 Human demonstrations | human-written solutions, explanations, rationales, and expert demonstrations | Needs verified primary-source additions; see needs_search. | human trace policy and expertise are unclear |
| 🤖 Synthetic instruction data | self-instruct, teacher-generated tasks, and synthetic instruction mixtures | Orca: Progressive learning from complex explanation traces of GPT-4 (2023, arXiv) · Card data: instruction response with detailed explanation, intermediate reasoning, and final ans…; feedback: downstream reasoning, exam, and benchmark evaluation rather than a single a… KodCode: A Diverse, Challenging, and Verifiable Synthetic Dataset for Coding (2025, ACL Findings) · Card data: question-solution-test triplet; process: problem, solution, unit tests; code executio…; feedback: test-based self-verification |
synthetic prompts collapse diversity or inherit teacher artifacts |
| 🧠 Chain-of-thought / rationale data | rationales, CoT traces, self-consistency, and reasoning-style supervision | Self-consistency improves chain of thought reasoning in language models (2023, ICLR) · Card data: multiple rationales and final answers for the same prompt.; process: sampling tempera…; feedback: majority or marginalization over sampled answers. |
trace style is mistaken for faithful reasoning |
| 🔁 Self-training / STaR | bootstrapped traces, self-training, critique loops, and filtered self-improvement | STaR: Bootstrapping reasoning with reasoning (2022, NeurIPS) · Card data: answer level; feedback: mixed Qwen2.5-Math technical report: Toward mathematical expert model via self-improvement (2024, arXiv) · Card data: math solution, final answer, optional tool/code execution trace, and reward-model sco…; feedback: math answer checks, reward model signals, and benchmark evaluations. |
feedback loop repeats hidden errors or shortcuts |
| ✂️ Long/short CoT distillation | teacher long traces, distilled short traces, and reasoning compression | DeepSeek-R1 (2025, arXiv) · Card data: answer level; feedback: mixed |
distillation loses uncertainty and failed attempts |
RLHF, DPO, reward modeling, rubric rewards, and AI feedback · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 🤝 Human preference data / RLHF | human comparison data, helpful/harmless feedback, and RLHF reward targets | A Survey of Reinforcement Learning from Human Feedback (2023, TMLR) data: survey taxonomy over feedback collection, reward modeling, and policy optimization.;…; feedback: learned reward model from human feedback. Training language models to follow instructions with human feedback (2022, NeurIPS) · Card data: pairwise preference; scalar reward; feedback: judgment required A Survey on Human Preference Learning for Large Language Models (2024, arXiv) data: preference-centered taxonomy over feedback data, preference modeling, preference usag…; feedback: human preference signal transformed into reward, preference loss, or evalua… A Comprehensive Survey of Reward Models: Taxonomy, Applications, Challenges, and Future (2025, arXiv) data: taxonomy of reward-model data sources, objectives, applications, evaluations, and cha…; feedback: reward model as proxy objective for downstream post-training. |
annotator assumptions and disagreement are hidden |
| ⚖️ DPO / preference optimization | pairwise data used directly for preference optimization | Direct preference optimization: Your language model is secretly a reward model (2023, NeurIPS) · Card data: pairwise preference; feedback: judgment required Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs (2024, arXiv) data: step-wise preference pairs; process: reasoning step, preferred continuation, rejected…; feedback: step-wise preference optimization objective |
preference pairs are reused outside collection context |
| 🎚️ Scalar reward / ORM data | outcome reward labels, scalar scores, and trained reward-model targets | Qwen2.5-Math technical report: Toward mathematical expert model via self-improvement (2024, arXiv) · Card data: math solution, final answer, optional tool/code execution trace, and reward-model sco…; feedback: math answer checks, reward model signals, and benchmark evaluations. |
scalar reward hides why an answer is better |
| 🤖 RLAIF / synthetic feedback | model-generated preferences, critiques, and constitutional feedback | Constitutional AI: Harmlessness from AI feedback (2022, arXiv preprint) · Card data: original answer, self-critique, revised answer, preference pair, reward-model score.;…; feedback: AI preference model trained from comparisons guided by constitutional princ… UltraFeedback: Boosting language models with high-quality feedback (2023, ICML) · Card data: instruction, candidate responses, fine-grained ratings, textual critiques, and derive…; feedback: AI-generated scalar and textual feedback over response quality dimensions. |
synthetic judge behavior is treated as human preference |
| 🧪 Reward-model benchmarks | rewardbench-style evaluation data and reward-model stress tests | RewardBench: Evaluating Reward Models for Language Modeling (2024, NeurIPS) · Card data: pairwise or scalar reward decisions; process: prompt, chosen/rejected response, rewar…; feedback: reward model or judge |
benchmark preference does not predict downstream training value |
| 🧾 Rubric-conditioned rewards | rubric scores, critique-plus-score records, and domain-specific reward signals | Prometheus 2: An open source language model specialized in evaluating other language models (2024, EMNLP) · Card data: rubric-conditioned scalar score, critique, or pairwise preference output.; process: i…; feedback: Prometheus 2 judge output aligned against human/proprietary-judge benchmark… Rewarding progress: Scaling automated process verifiers for LLM reasoning (2024, ICLR) · Card data: step-level process advantage score plus final answer/correctness signal.; process: pr…; feedback: Process Advantage Verifier trained to predict progress toward correct answe… |
rubric wording becomes an exploitable reward channel |
math, code, proof, and answer-verifiable reasoning data · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 📐 Math answer-verifiable data | math problems, final answers, solution traces, and answer checkers | GSM8K: Grade School Math 8K (2021, arXiv / OpenAI dataset) · Card data: natural-language solution with final numeric answer; process: question, solution, fin…; feedback: answer extraction and arithmetic correctness checks Training verifiers to solve math word problems (2021, arXiv) · Card data: answer level; scalar reward; feedback: programmatic, judgment required DeepMath-103K (2025, arXiv) · Card data: answer level; feedback: programmatic OpenMathInstruct-2: Accelerating AI for math with massive open-source instruction data (2024, ICLR) · Card data: problem-solution pair with natural-language mathematical reasoning and final answer.;…; feedback: answer checks and benchmark evaluation over math tasks. |
answer extraction and normalization inflate scores |
| 🧮 Math RLVR datasets | math records used for rejection sampling, SFT, PRMs, and RLVR | DeepSeekMath: Pushing the limits of mathematical reasoning in open language models (2024, arXiv) · Card data: natural-language mathematical solution plus final answer, sometimes sampled multiple…; feedback: answer correctness and GRPO-style reward over math tasks. |
data reuse and contamination are not reported |
| 💻 Code execution / unit-test data | code problems, unit tests, generated tests, execution logs, and repair tasks | LiveCodeBench: Holistic and contamination-free evaluation of large language models for code (2024, arXiv) · Card data: program submission or code-related output evaluated by tests or task-specific checks.…; feedback: programmatic tests and task-specific correctness checks. Evaluating large language models trained on code (2021, arXiv) · Card data: executable Python function.; process: prompt, generated code, unit-test results, samp…; feedback: HumanEval tests and pass@k evaluation. HumanEval: Hand-Written Evaluation Set (2021, arXiv / OpenAI dataset) · Card data: Python function completion; process: prompt, canonical solution, unit tests; Python e…; feedback: unit tests Measuring coding challenge competence with APPS (2021, NeurIPS) · Card data: Python code submission evaluated against test cases.; process: difficulty, prompt, st…; feedback: unit-test pass/fail signal. |
flaky or leaked tests become the reward |
| 🧾 Formal proof / Lean / theorem proving | Lean, proof scripts, tactic environments, theorem statements, and proof checkers | DeepSeek-Prover: Advancing theorem proving in LLMs (2024, arXiv) · Card data: Lean 4 theorem statement and proof script checked by Lean.; process: informal problem…; feedback: Lean kernel/checker acceptance. LeanDojo: Theorem proving with retrieval-augmented language models (2023, NeurIPS Datasets and Benchmarks) · Card data: Lean tactic sequence or proof script checked by Lean.; process: repository commit, th…; feedback: Lean checker and environment feedback. DeepSeek-Prover-V1.5: Harnessing proof assistant feedback for reinforcement learning and Monte-Carlo tree search (2024, arXiv) · Card data: Lean proof script, proof-search path, feedback signal, and verification result.; proc…; feedback: proof assistant feedback used for RL and search selection. miniF2F: A cross-system benchmark for formal olympiad-level mathematics (2021, ICLR) · Card data: formal proof accepted by a target proof assistant.; process: formal system, theorem s…; feedback: proof assistant kernel/checker acceptance. |
proof succeeds only under an undocumented environment |
| 🧪 Verifier robustness and answer extraction | false positives, false negatives, checker brittleness, and adversarial formats | AutoPSV: Automated Process-Supervised Verifier (2024, arXiv) data: step-level confidence-change annotations; process: reasoning step, verifier confidenc…; feedback: answer-trained verifier converted into process annotations |
model learns verifier quirks instead of task skill |
| 🧰 Programmatic benchmarks | evaluation sets whose scoring can become a post-training signal | SciCode: A benchmark for scientific code generation and reasoning (2024, NeurIPS Datasets and Benchmarks) · Card data: code solution evaluated with scientist-annotated tests or expected outputs.; process:…; feedback: test cases and scientist-curated gold solutions. LiveBench: A challenging, contamination-free benchmark for large language models (2024, arXiv) · Card data: answer level; feedback: programmatic, mixed PRMBench: A fine-grained and challenging benchmark for process-level reward models (2025, arXiv) · Card data: step-level labels or scores; process: step, label, error type; offline reasoning trac…; feedback: process-level reward model benchmark ProcessBench: Identifying Process Errors in Mathematical Reasoning (2024, arXiv) · Card data: step labels or first-error markers; process: reasoning step, error marker, diagnostic…; feedback: process-error detector |
benchmark scoring is reused as reward without audit |
step-level labels, PRMs, rollout values, and first-error signals · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 🪜 Human step-level labels | human-labeled intermediate steps and first-error annotations | Let's Verify Step by Step (2023, arXiv) · Card data: step-level labels and final answers; process: step, label, solution trace; offline ma…; feedback: process reward model trained from step labels |
step boundaries and label policy are ambiguous |
| 🧪 Process reward models | PRMs, process verifiers, calibration, and reward-model training | ReST-MCTS* (2024, arXiv) data: reasoning trajectory with intermediate search states; process: node state, rollout ca…; feedback: process reward guided tree search |
process reward rises while final correctness does not |
| 🔁 Rollout-value supervision | rollout values, search-derived labels, and automated progress signals | Math-Shepherd (2024, arXiv) · Card data: step-level rollout-value labels; process: reasoning step, rollout result, process rew…; feedback: rollout-derived process reward signal Rewarding progress: Scaling automated process verifiers for LLM reasoning (2024, ICLR) · Card data: step-level process advantage score plus final answer/correctness signal.; process: pr…; feedback: Process Advantage Verifier trained to predict progress toward correct answe… |
rollout policy leaks solver strength into labels |
| 🛠️ Automatic process supervision | programmatic or model-generated process labels without dense human annotation | PRIME: Process reinforcement through implicit rewards (2025, arXiv) data: rollout with implicit process reward signal; process: policy rollout, outcome label,…; feedback: implicit process rewards derived from outcome labels OmegaPRM: Improve Mathematical Reasoning in Language Models by Automated Process Supervision (2024, arXiv) · Card data: process supervision annotations; process: partial reasoning prefix, first-error signa…; feedback: automated process reward signal AutoPSV: Automated Process-Supervised Verifier (2024, arXiv) data: step-level confidence-change annotations; process: reasoning step, verifier confidenc…; feedback: answer-trained verifier converted into process annotations |
automatic labels silently inherit verifier bias |
| ❌ First-error localization | where a solution first becomes invalid and how that signal is used | Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs (2024, arXiv) data: step-wise preference pairs; process: reasoning step, preferred continuation, rejected…; feedback: step-wise preference optimization objective |
localized errors are not causally linked to correction |
| 📊 PRM benchmarks and evaluation | ProcessBench, PRMBench, Qwen PRM, and evaluation surfaces for process rewards | PRMBench: A fine-grained and challenging benchmark for process-level reward models (2025, arXiv) · Card data: step-level labels or scores; process: step, label, error type; offline reasoning trac…; feedback: process-level reward model benchmark ProcessBench: Identifying Process Errors in Mathematical Reasoning (2024, arXiv) · Card data: step labels or first-error markers; process: reasoning step, error marker, diagnostic…; feedback: process-error detector |
PRM benchmark success does not transfer to training use |
search-generated candidates, best-of-N, pass@k, and test-time compute traces · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 🎲 Multiple rollouts / best-of-N | sets of sampled attempts and selected accepted answers | Evaluating large language models trained on code (2021, arXiv) · Card data: executable Python function.; process: prompt, generated code, unit-test results, samp…; feedback: HumanEval tests and pass@k evaluation. |
only accepted traces are visible |
| 🌳 Search trees / MCTS | tree search, MCTS, verifier-guided search, and path selection | ReST-MCTS* (2024, arXiv) data: reasoning trajectory with intermediate search states; process: node state, rollout ca…; feedback: process reward guided tree search DeepSeek-Prover-V1.5: Harnessing proof assistant feedback for reinforcement learning and Monte-Carlo tree search (2024, arXiv) · Card data: Lean proof script, proof-search path, feedback signal, and verification result.; proc…; feedback: proof assistant feedback used for RL and search selection. |
tree policy or value model is hidden |
| 🔎 Rejection sampling traces | accepted and rejected candidates produced during filtering | Introducing SWE-bench Verified (2024, OpenAI / SWE-bench report) · Card data: patch diff applied to a repository plus test execution results.; process: repository,…; feedback: post-patch unit tests plus human filtering of task validity. |
rejected examples are not released |
| 🧠 Self-consistency / repeated sampling | vote-based or agreement-based reasoning from repeated samples | Self-consistency improves chain of thought reasoning in language models (2023, ICLR) · Card data: multiple rationales and final answers for the same prompt.; process: sampling tempera…; feedback: majority or marginalization over sampled answers. DeepSeekMath: Pushing the limits of mathematical reasoning in open language models (2024, arXiv) · Card data: natural-language mathematical solution plus final answer, sometimes sampled multiple…; feedback: answer correctness and GRPO-style reward over math tasks. |
sampling budget is not comparable |
| ⏱️ Test-time compute logs | thinking budgets, inference-time scaling, and runtime search traces | s1: Simple Test-Time Scaling (2025, arXiv) · Card data: answer level; feedback: mixed MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention (2025, arXiv preprint arXiv:2506.13585) · Card data: reasoning output, code/tool result, or agent task output; process: reasoning output,…; feedback: programmatic, environment, and benchmark feedback OpenCodeReasoning-II: A Simple Test Time Scaling Approach via Self-Critique (2025, arXiv) · Card data: question-solution-critique triple; process: solution, critique, language/runtime labe…; feedback: tests and critique model signals TTRL: Test-Time Reinforcement Learning (2025, arXiv preprint arXiv:2504.16084) · Card data: candidate response with reward/adaptation signal; process: unlabeled input, rollout,…; feedback: task-specific or learned reward used during adaptation |
training and inference budget effects are conflated |
| ✂️ Long2short / distill-from-search | using long search traces to train shorter or cheaper behavior | Needs verified primary-source additions; see needs_search. | teacher search artifacts become hidden data lineage |
tool, web, OS, app, SWE, and replayable environment data · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 🛠️ Tool-use data | tool calls, function signatures, API banks, and tool-use traces | Toolformer: Language models can teach themselves to use tools (2023, NeurIPS) · Card data: text sequence with inserted API call and tool result markup.; process: candidate call…; feedback: language-model likelihood improvement after including tool result. ToolLLM: Facilitating large language models to master 16000+ real-world APIs (2023, ICLR) · Card data: tool-call trajectory plus final response; process: API call, arguments, tool response…; feedback: tool response validity and task success checks Qwen2.5-Math technical report: Toward mathematical expert model via self-improvement (2024, arXiv) · Card data: math solution, final answer, optional tool/code execution trace, and reward-model sco…; feedback: math answer checks, reward model signals, and benchmark evaluations. |
tool schemas change or hide execution failures |
| 🌍 Web/browser agents | web tasks, browser state, navigation traces, and page observations | WebArena: A realistic web environment for building autonomous agents (2023, ICLR) · Card data: environment interaction trajectory; process: observation, action, state; browser-acce…; feedback: task-specific success evaluator BrowserGym: A gym environment for web agents (2024, arXiv) · Card data: browser trajectory; process: DOM/state observation, action, reward/result; gym-style…; feedback: environment task evaluator DeepSeekMath: Pushing the limits of mathematical reasoning in open language models (2024, arXiv) · Card data: natural-language mathematical solution plus final answer, sometimes sampled multiple…; feedback: answer correctness and GRPO-style reward over math tasks. GPQA (2023, arXiv) · Card data: multiple-choice answer with optional rationale and expert label.; process: domain, qu…; feedback: expert-authored answer key and validation protocol. |
web state is not replayable after collection |
| 📱 App/mobile agents | mobile apps, app-world tasks, UI actions, and user simulators | AppWorld: A controllable world of apps and people for benchmarking interactive coding agents (2024, arXiv) · Card data: API/tool action trajectory; process: tool call, state transition, observation; simula…; feedback: programmatic environment assertions AndroidWorld: A dynamic benchmarking environment for autonomous agents (2024, arXiv) · Card data: Android action trajectory; process: screen observation, UI action, state transition;…; feedback: task-specific success evaluator |
UI state and app versions are not pinned |
| 🖥️ OS/desktop agents | desktop/OS tasks, filesystem state, shell actions, and multi-app workflows | OSWorld: Benchmarking multimodal agents for open-ended tasks in real computer environments (2024, NeurIPS) · Card data: GUI/OS action trajectory; process: observation, action, environment state; desktop op…; feedback: task completion evaluator |
hidden environment state makes episodes non-reproducible |
| 🧑💻 SWE/repository agents | GitHub issues, code patches, tests, commits, and repository repair episodes | SWE-bench: Can language models resolve real-world GitHub issues? (2023, ICLR) · Card data: full episode; state action level; feedback: environmental, programmatic Introducing SWE-bench Verified (2024, OpenAI / SWE-bench report) · Card data: patch diff applied to a repository plus test execution results.; process: repository,…; feedback: post-patch unit tests plus human filtering of task validity. R2E-Gym (2025, arXiv) · Card data: full episode; state action level; feedback: environmental, programmatic SWE-Gym (2025, arXiv) · Card data: full episode; state action level; feedback: environmental, programmatic |
repository commit, tests, and scaffold are not pinned |
| 🔁 Replayable trajectory data | state-action-observation schemas, terminal predicates, and failure traces | ReAct: Synergizing reasoning and acting in language models (2023, ICLR) · Card data: trajectory containing reasoning note, action, observation, and final answer or task c…; feedback: environment success, answer correctness, or task-specific evaluation. |
success transcript cannot be replayed or audited |
| 🧰 Agent benchmarks and terminal predicates | agent evaluation suites, task resets, terminal predicates, and success/failure labels | miniF2F: A cross-system benchmark for formal olympiad-level mathematics (2021, ICLR) · Card data: formal proof accepted by a target proof assistant.; process: formal system, theorem s…; feedback: proof assistant kernel/checker acceptance. PRMBench: A fine-grained and challenging benchmark for process-level reward models (2025, arXiv) · Card data: step-level labels or scores; process: step, label, error type; offline reasoning trac…; feedback: process-level reward model benchmark FinanceBench: A benchmark for financial question answering (2023, arXiv) data: answer level; feedback: judgment required, mixed TAT-QA: A question answering benchmark on a hybrid of tabular and textual content in finance (2021, ACL) data: answer level; step level; feedback: mixed |
score is reported without a replayable predicate |
LLM judges, expert rubrics, factuality, safety, medical, legal, and finance reasoning · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| ⚖️ LLM-as-judge data | model judges, preference judgments, judge prompts, and evaluator models | Judging LLM-as-a-judge with MT-Bench and Chatbot Arena (2023, NeurIPS Datasets and Benchmarks) · Card data: model response, judge score, pairwise preference, or arena battle outcome.; process:…; feedback: strong model judge and human preference comparisons. Prometheus 2: An open source language model specialized in evaluating other language models (2024, EMNLP) · Card data: rubric-conditioned scalar score, critique, or pairwise preference output.; process: i…; feedback: Prometheus 2 judge output aligned against human/proprietary-judge benchmark… One Token to Fool LLM-as-a-Judge (2025, arXiv) · Card data: scalar reward; feedback: judgment required |
judge is sensitive to style, position, or prompt attacks |
| 🧑⚖️ Human/expert judgment | human labels, expert adjudication, disagreement handling, and rubric design | Qwen2.5-Math technical report: Toward mathematical expert model via self-improvement (2024, arXiv) · Card data: math solution, final answer, optional tool/code execution trace, and reward-model sco…; feedback: math answer checks, reward model signals, and benchmark evaluations. Training language models to follow instructions with human feedback (2022, NeurIPS) · Card data: pairwise preference; scalar reward; feedback: judgment required A Survey on Human Preference Learning for Large Language Models (2024, arXiv) data: preference-centered taxonomy over feedback data, preference modeling, preference usag…; feedback: human preference signal transformed into reward, preference loss, or evalua… A Survey of Reinforcement Learning from Human Feedback (2023, TMLR) data: survey taxonomy over feedback collection, reward modeling, and policy optimization.;…; feedback: learned reward model from human feedback. |
expertise and adjudication policy are not disclosed |
| 🩺 Medical reasoning / health rubrics | health, biomedical, scientific, and evidence-grounded reasoning tasks | HealthBench (2025, arXiv) · Card data: response with rubric/judge evaluation; process: prompt, response, rubric dimension; o…; feedback: rubric-guided expert/LLM judgment GPQA (2023, arXiv) · Card data: multiple-choice answer with optional rationale and expert label.; process: domain, qu…; feedback: expert-authored answer key and validation protocol. |
rubrics are not calibrated for high-stakes error |
| 🛡️ Safety reasoning data | safety reasoning, refusals, jailbreaks, harmfulness, and guardrail data | AbstentionBench (2025, arXiv) · Card data: model response, abstention decision, and correctness/abstention judgment.; process: s…; feedback: human-validated judges and benchmark labels for abstention scenarios. Leaky Thoughts (2025, arXiv) · Card data: internal reasoning trace, final answer, and leakage/extraction outcome.; process: sen…; feedback: extraction probes and agentic evaluations. Llama-Nemotron: Efficient Reasoning Models (2025, arXiv) · Card data: answer level; feedback: mixed |
safe-looking refusals replace correct domain reasoning |
| 🧾 Factuality / grounding | claims, citations, retrieval grounding, fact checking, and evidence quality | Self-RAG: Learning to retrieve, generate, and critique through self-reflection (2023, ICLR) · Card data: generation with retrieval decisions, critique signals, and final answer.; process: qu…; feedback: critique signals and task-specific factuality/answer-quality evaluation. |
citation style masks unsupported claims |
| ⚖️ Legal reasoning | legal QA, statutes, case reasoning, contracts, and expert legal rubrics | LegalBench (2023, NeurIPS) data: answer level; feedback: judgment required, mixed |
splits leak templates or jurisdiction assumptions |
| 🏦 Financial reasoning | financial QA, tabular/text numerical reasoning, filings, and analyst-style judgments | FinanceBench: A benchmark for financial question answering (2023, arXiv) data: answer level; feedback: judgment required, mixed FinQA: A dataset of numerical reasoning over financial data (2021, EMNLP) data: answer level; step level; feedback: mixed TAT-QA: A question answering benchmark on a hybrid of tabular and textual content in finance (2021, ACL) data: answer level; step level; feedback: mixed |
splits leak templates or memorized company facts |
| 🧪 Rubric reward models | rubrics as trainable rewards and domain-conditioned reward models | RewardBench: Evaluating Reward Models for Language Modeling (2024, NeurIPS) · Card data: pairwise or scalar reward decisions; process: prompt, chosen/rejected response, rewar…; feedback: reward model or judge |
rubric scores are optimized without semantic robustness |
building, filtering, releasing, and reproducing reasoning datasets · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 🧱 Prompt sourcing | question pools, seed sources, licenses, difficulty, and base-model pass rates | OpenMathInstruct-2: Accelerating AI for math with massive open-source instruction data (2024, ICLR) · Card data: problem-solution pair with natural-language mathematical reasoning and final answer.;…; feedback: answer checks and benchmark evaluation over math tasks. KodCode: A Diverse, Challenging, and Verifiable Synthetic Dataset for Coding (2025, ACL Findings) · Card data: question-solution-test triplet; process: problem, solution, unit tests; code executio…; feedback: test-based self-verification OpenMathReasoning: A large-scale dataset of math reasoning traces (2025, arXiv) · Card data: answer level; feedback: programmatic s1: Simple Test-Time Scaling (2025, arXiv) · Card data: answer level; feedback: mixed |
prompt sources are mixed without attribution |
| ✍️ Teacher trace generation | teacher models, trace policies, sampling settings, and distillation targets | Subliminal Learning (2025, arXiv) · Card data: generated data plus downstream behavioral evaluation of the student.; process: teache…; feedback: trait probes after student training. Orca: Progressive learning from complex explanation traces of GPT-4 (2023, arXiv) · Card data: instruction response with detailed explanation, intermediate reasoning, and final ans…; feedback: downstream reasoning, exam, and benchmark evaluation rather than a single a… Phi-4-reasoning Technical Report (2025, arXiv) · Card data: answer level; feedback: mixed ToolLLM: Facilitating large language models to master 16000+ real-world APIs (2023, ICLR) · Card data: tool-call trajectory plus final response; process: API call, arguments, tool response…; feedback: tool response validity and task success checks |
teacher identity or sampling protocol is hidden |
| 🔎 Rejection sampling / search-generated data | candidate generation, search budget, filtering, and accepted/rejected examples | DeepSeek-Prover-V1.5: Harnessing proof assistant feedback for reinforcement learning and Monte-Carlo tree search (2024, arXiv) · Card data: Lean proof script, proof-search path, feedback signal, and verification result.; proc…; feedback: proof assistant feedback used for RL and search selection. Evaluating large language models trained on code (2021, arXiv) · Card data: executable Python function.; process: prompt, generated code, unit-test results, samp…; feedback: HumanEval tests and pass@k evaluation. OmegaPRM: Improve Mathematical Reasoning in Language Models by Automated Process Supervision (2024, arXiv) · Card data: process supervision annotations; process: partial reasoning prefix, first-error signa…; feedback: automated process reward signal ReARTeR: Retrieval-Augmented Reasoning with Trustworthy Process Rewarding (2025, arXiv) data: step-level preference data with process scores and explanations; process: retrieval c…; feedback: process reward model plus process explanation model |
only accepted traces are released |
| 🔁 Self-play / self-improvement | self-improvement, co-evolution, generator-verifier cycles, and curricula | Qwen2.5-Math technical report: Toward mathematical expert model via self-improvement (2024, arXiv) · Card data: math solution, final answer, optional tool/code execution trace, and reward-model sco…; feedback: math answer checks, reward model signals, and benchmark evaluations. Absolute Zero: Reinforced Self-play Reasoning with Zero Data (2025, arXiv preprint arXiv:2505.03335) · Card data: generated task, solution, and verified answer; process: proposed task, solution, veri…; feedback: executor-backed verifiable reward AutoPSV: Automated Process-Supervised Verifier (2024, arXiv) data: step-level confidence-change annotations; process: reasoning step, verifier confidenc…; feedback: answer-trained verifier converted into process annotations LeanDojo: Theorem proving with retrieval-augmented language models (2023, NeurIPS Datasets and Benchmarks) · Card data: Lean tactic sequence or proof script checked by Lean.; process: repository commit, th…; feedback: Lean checker and environment feedback. |
feedback loop amplifies hidden shortcuts |
| 🧪 Filtering and verifier refresh | answer filters, judge filters, decontamination, and verifier updates | Big-Math-RL-Verified (2025, arXiv) · Card data: math problem, answer, and verification signal; process: problem, answer, verification…; feedback: answer-level math verifier DAPO (2025, arXiv) · Card data: answer level; feedback: programmatic DeepMath-103K (2025, arXiv) · Card data: answer level; feedback: programmatic |
filter thresholds become hidden objectives |
| 🏗️ Open reasoning data releases | open datasets, code, HF releases, recipes, ablations, and reproducibility | OpenThoughts: Data recipes for reasoning models (2025, arXiv) · Card data: reasoning traces and final answers; process: question, reasoning trace, answer; offli…; feedback: filters, benchmark feedback, and recipe ablations |
dataset is open but recipe details are not |
| 🧬 Data lineage and release metadata | datasheets, splits, lineage, licensing, versioning, and known failures | Training a helpful and harmless assistant with reinforcement learning from human feedback (2022, arXiv) data: pairwise preference; scalar reward; feedback: judgment required |
reuse loses the release context |
how data enters SFT, DPO, RM, PRM, RLVR, agents, evaluation, and audit · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 🧱 SFT / instruction tuning | data used as supervised target behavior | OpenMathInstruct-2: Accelerating AI for math with massive open-source instruction data (2024, ICLR) · Card data: problem-solution pair with natural-language mathematical reasoning and final answer.;…; feedback: answer checks and benchmark evaluation over math tasks. Tulu 3: Pushing frontiers in open language model post-training (2024, arXiv) · Card data: instruction-response examples, preference pairs, verifiable task outputs, and model-e…; feedback: mixture of preference labels, reward models, and verifiable rewards dependi… Toolformer: Language models can teach themselves to use tools (2023, NeurIPS) · Card data: text sequence with inserted API call and tool result markup.; process: candidate call…; feedback: language-model likelihood improvement after including tool result. Training language models to follow instructions with human feedback (2022, NeurIPS) · Card data: pairwise preference; scalar reward; feedback: judgment required |
target text hides verifier and source assumptions |
| 📚 Distillation | teacher outputs, traces, or policies distilled into a student | Subliminal Learning (2025, arXiv) · Card data: generated data plus downstream behavioral evaluation of the student.; process: teache…; feedback: trait probes after student training. Orca: Progressive learning from complex explanation traces of GPT-4 (2023, arXiv) · Card data: instruction response with detailed explanation, intermediate reasoning, and final ans…; feedback: downstream reasoning, exam, and benchmark evaluation rather than a single a… Phi-4-reasoning Technical Report (2025, arXiv) · Card data: answer level; feedback: mixed OpenThoughts: Data recipes for reasoning models (2025, arXiv) · Card data: reasoning traces and final answers; process: question, reasoning trace, answer; offli…; feedback: filters, benchmark feedback, and recipe ablations |
teacher lineage is hidden |
| ⚖️ Preference optimization | pairwise feedback for DPO/IPO/KTO-style objectives | Step-DPO: Step-wise Preference Optimization for Long-chain Reasoning of LLMs (2024, arXiv) data: step-wise preference pairs; process: reasoning step, preferred continuation, rejected…; feedback: step-wise preference optimization objective Direct preference optimization: Your language model is secretly a reward model (2023, NeurIPS) · Card data: pairwise preference; feedback: judgment required |
pair context does not match downstream use |
| 🎚️ Reward modeling / ORM | scalar or pairwise data used to train outcome rewards | Needs verified primary-source additions; see needs_search. | reward can be overoptimized |
| 🪜 PRM / process supervision | step-level or trace-level signals used to train process rewards | Rewarding progress: Scaling automated process verifiers for LLM reasoning (2024, ICLR) · Card data: step-level process advantage score plus final answer/correctness signal.; process: pr…; feedback: Process Advantage Verifier trained to predict progress toward correct answe… Let's Verify Step by Step (2023, arXiv) · Card data: step-level labels and final answers; process: step, label, solution trace; offline ma…; feedback: process reward model trained from step labels OmegaPRM: Improve Mathematical Reasoning in Language Models by Automated Process Supervision (2024, arXiv) · Card data: process supervision annotations; process: partial reasoning prefix, first-error signa…; feedback: automated process reward signal |
PRM rewards trace style |
| 🏋️ RLVR / verifier RL | programmatic or verifier rewards used in RL | Reinforcement Learning for LLM Post-Training: A Survey (2024, arXiv) data: technical survey comparing RLHF and RLVR policy-gradient style post-training methods.…; feedback: learned preference rewards, verifiable rewards, and policy-gradient objecti… DeepSeek-Prover-V2: Advancing formal mathematical reasoning via reinforcement learning (2025, arXiv) · Card data: subgoal chain, informal reasoning trace, Lean proof, and checker result.; process: pr…; feedback: Lean verification and RL reward over formal proof success. DeepSeek-Prover-V1.5: Harnessing proof assistant feedback for reinforcement learning and Monte-Carlo tree search (2024, arXiv) · Card data: Lean proof script, proof-search path, feedback signal, and verification result.; proc…; feedback: proof assistant feedback used for RL and search selection. TTRL: Test-Time Reinforcement Learning (2025, arXiv preprint arXiv:2504.16084) · Card data: candidate response with reward/adaptation signal; process: unlabeled input, rollout,…; feedback: task-specific or learned reward used during adaptation |
verifier false positives become policy incentives |
| 🌐 Agent training | environment episodes, tool traces, or terminal rewards for agent policies | BrowserGym: A gym environment for web agents (2024, arXiv) · Card data: browser trajectory; process: DOM/state observation, action, reward/result; gym-style…; feedback: environment task evaluator AndroidWorld: A dynamic benchmarking environment for autonomous agents (2024, arXiv) · Card data: Android action trajectory; process: screen observation, UI action, state transition;…; feedback: task-specific success evaluator WebArena: A realistic web environment for building autonomous agents (2023, ICLR) · Card data: environment interaction trajectory; process: observation, action, state; browser-acce…; feedback: task-specific success evaluator LeanDojo: Theorem proving with retrieval-augmented language models (2023, NeurIPS Datasets and Benchmarks) · Card data: Lean tactic sequence or proof script checked by Lean.; process: repository commit, th…; feedback: Lean checker and environment feedback. |
environment cannot be replayed |
| 🧪 Evaluation / reranking / audit | data used for scoring, selection, reporting, or failure analysis | LiveBench: A challenging, contamination-free benchmark for large language models (2024, arXiv) · Card data: answer level; feedback: programmatic, mixed SciCode: A benchmark for scientific code generation and reasoning (2024, NeurIPS Datasets and Benchmarks) · Card data: code solution evaluated with scientist-annotated tests or expected outputs.; process:…; feedback: test cases and scientist-curated gold solutions. miniF2F: A cross-system benchmark for formal olympiad-level mathematics (2021, ICLR) · Card data: formal proof accepted by a target proof assistant.; process: formal system, theorem s…; feedback: proof assistant kernel/checker acceptance. PRMBench: A fine-grained and challenging benchmark for process-level reward models (2025, arXiv) · Card data: step-level labels or scores; process: step, label, error type; offline reasoning trac…; feedback: process-level reward model benchmark |
evaluation data becomes training data |
data scale, RLVR, verifier scaling, pass@k, and inference budget claims · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 📈 Data scaling | number, diversity, difficulty, and uniqueness of examples | OpenMathReasoning: A large-scale dataset of math reasoning traces (2025, arXiv) · Card data: answer level; feedback: programmatic DeepSeek-Prover: Advancing theorem proving in LLMs (2024, arXiv) · Card data: Lean 4 theorem statement and proof script checked by Lean.; process: informal problem…; feedback: Lean kernel/checker acceptance. UltraFeedback: Boosting language models with high-quality feedback (2023, ICML) · Card data: instruction, candidate responses, fine-grained ratings, textual critiques, and derive…; feedback: AI-generated scalar and textual feedback over response quality dimensions. Big-Math-RL-Verified (2025, arXiv) · Card data: math problem, answer, and verification signal; process: problem, answer, verification…; feedback: answer-level math verifier |
unique examples and repeated rollouts are conflated |
| 🔁 Data reuse and uniqueness | reuse counts, deduplication, repeated prompts, and train/test overlap | Language Model Developers Should Report Train-Test Overlap (2024, arXiv) data: overlap and reporting analysis.; process: training corpus, evaluation set, overlap es…; feedback: overlap analysis rather than a reward model. |
same source examples are counted as fresh data |
| ⏱️ Test-time compute | sampling, search, self-critique, thinking budgets, and inference-time scaling | s1: Simple Test-Time Scaling (2025, arXiv) · Card data: answer level; feedback: mixed MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention (2025, arXiv preprint arXiv:2506.13585) · Card data: reasoning output, code/tool result, or agent task output; process: reasoning output,…; feedback: programmatic, environment, and benchmark feedback TTRL: Test-Time Reinforcement Learning (2025, arXiv preprint arXiv:2504.16084) · Card data: candidate response with reward/adaptation signal; process: unlabeled input, rollout,…; feedback: task-specific or learned reward used during adaptation The Art of Scaling Reinforcement Learning Compute for LLMs (2025, arXiv) · Card data: training runs, reward outcomes, validation curves, and ablation results.; process: lo…; feedback: compute-performance curves and recipe ablations. |
different inference budgets are compared |
| 🎲 pass@k / sampling budget | pass@k, repeated sampling, best-of-N, and budget-aware evaluation | Evaluating large language models trained on code (2021, arXiv) · Card data: executable Python function.; process: prompt, generated code, unit-test results, samp…; feedback: HumanEval tests and pass@k evaluation. |
reported gains hide selection or budget changes |
| 🧪 Verifier scaling | how verifier strength, refresh, and coverage scale with training | Needs verified primary-source additions; see needs_search. | verifier becomes stale or easy to exploit |
| 🏋️ RLVR optimization scaling | policy optimization, reward contracts, curriculum, and rollout policy | DAPO (2025, arXiv) · Card data: answer level; feedback: programmatic DeepSeek-Prover-V2: Advancing formal mathematical reasoning via reinforcement learning (2025, arXiv) · Card data: subgoal chain, informal reasoning trace, Lean proof, and checker result.; process: pr…; feedback: Lean verification and RL reward over formal proof success. DeepSeek-R1 (2025, arXiv) · Card data: answer level; feedback: mixed Tulu 3: Pushing frontiers in open language model post-training (2024, arXiv) · Card data: instruction-response examples, preference pairs, verifiable task outputs, and model-e…; feedback: mixture of preference labels, reward models, and verifiable rewards dependi… |
optimizer/scaffold gains are mistaken for data gains |
| 🔍 Scaling attribution | separating data, verifier, optimizer, model, and inference-budget effects | Needs verified primary-source additions; see needs_search. | ablation tables do not isolate the source of improvement |
evaluation surfaces and reusable feedback contracts · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 📐 Math benchmarks | math problem sets, answer extraction, verifier compatibility, and difficulty | GSM8K: Grade School Math 8K (2021, arXiv / OpenAI dataset) · Card data: natural-language solution with final numeric answer; process: question, solution, fin…; feedback: answer extraction and arithmetic correctness checks Training verifiers to solve math word problems (2021, arXiv) · Card data: answer level; scalar reward; feedback: programmatic, judgment required Qwen2.5-Math technical report: Toward mathematical expert model via self-improvement (2024, arXiv) · Card data: math solution, final answer, optional tool/code execution trace, and reward-model sco…; feedback: math answer checks, reward model signals, and benchmark evaluations. Measuring mathematical problem solving with the MATH dataset (2021, NeurIPS Datasets and Benchmarks) · Card data: answer level; feedback: programmatic |
short-answer normalization hides reasoning errors |
| 💻 Code benchmarks | coding tasks, generated tests, hidden tests, repair tasks, and live coding | LiveCodeBench: Holistic and contamination-free evaluation of large language models for code (2024, arXiv) · Card data: program submission or code-related output evaluated by tests or task-specific checks.…; feedback: programmatic tests and task-specific correctness checks. Evaluating large language models trained on code (2021, arXiv) · Card data: executable Python function.; process: prompt, generated code, unit-test results, samp…; feedback: HumanEval tests and pass@k evaluation. HumanEval: Hand-Written Evaluation Set (2021, arXiv / OpenAI dataset) · Card data: Python function completion; process: prompt, canonical solution, unit tests; Python e…; feedback: unit tests Measuring coding challenge competence with APPS (2021, NeurIPS) · Card data: Python code submission evaluated against test cases.; process: difficulty, prompt, st…; feedback: unit-test pass/fail signal. |
unit tests are brittle, leaked, or too narrow |
| 🧾 Proof benchmarks | formal proof datasets, proof assistants, theorem statements, and checking | miniF2F: A cross-system benchmark for formal olympiad-level mathematics (2021, ICLR) · Card data: formal proof accepted by a target proof assistant.; process: formal system, theorem s…; feedback: proof assistant kernel/checker acceptance. LeanDojo: Theorem proving with retrieval-augmented language models (2023, NeurIPS Datasets and Benchmarks) · Card data: Lean tactic sequence or proof script checked by Lean.; process: repository commit, th…; feedback: Lean checker and environment feedback. |
proof checker version and imports are not pinned |
| 🌐 Agent benchmarks | web, tool, OS, app, and SWE environments with terminal predicates | Introducing SWE-bench Verified (2024, OpenAI / SWE-bench report) · Card data: patch diff applied to a repository plus test execution results.; process: repository,…; feedback: post-patch unit tests plus human filtering of task validity. SWE-bench: Can language models resolve real-world GitHub issues? (2023, ICLR) · Card data: full episode; state action level; feedback: environmental, programmatic SWE-Gym (2025, arXiv) · Card data: full episode; state action level; feedback: environmental, programmatic BrowserGym: A gym environment for web agents (2024, arXiv) · Card data: browser trajectory; process: DOM/state observation, action, reward/result; gym-style…; feedback: environment task evaluator |
benchmark episodes cannot be replayed |
| ⚖️ Rubric/domain benchmarks | medical, safety, legal, finance, science, factuality, and expert rubrics | HealthBench (2025, arXiv) · Card data: response with rubric/judge evaluation; process: prompt, response, rubric dimension; o…; feedback: rubric-guided expert/LLM judgment AbstentionBench (2025, arXiv) · Card data: model response, abstention decision, and correctness/abstention judgment.; process: s…; feedback: human-validated judges and benchmark labels for abstention scenarios. TruthfulQA (2022, ACL) · Card data: free-form generation or multiple-choice answer with truthfulness and informativeness…; feedback: human references plus automated/human scoring protocols for truthfulness an… FinanceBench: A benchmark for financial question answering (2023, arXiv) data: answer level; feedback: judgment required, mixed |
rubric or judge expertise is insufficiently disclosed |
| 🧪 Reward-model benchmarks | reward model, LLM-judge, PRM, and rubric evaluation suites | RewardBench: Evaluating Reward Models for Language Modeling (2024, NeurIPS) · Card data: pairwise or scalar reward decisions; process: prompt, chosen/rejected response, rewar…; feedback: reward model or judge Judging LLM-as-a-judge with MT-Bench and Chatbot Arena (2023, NeurIPS Datasets and Benchmarks) · Card data: model response, judge score, pairwise preference, or arena battle outcome.; process:…; feedback: strong model judge and human preference comparisons. One Token to Fool LLM-as-a-Judge (2025, arXiv) · Card data: scalar reward; feedback: judgment required PRMBench: A fine-grained and challenging benchmark for process-level reward models (2025, arXiv) · Card data: step-level labels or scores; process: step, label, error type; offline reasoning trac…; feedback: process-level reward model benchmark |
benchmark reward preference does not reflect training value |
| 🧯 Live / contamination-resistant benchmarks | live, refreshed, hidden, or contamination-aware evaluation | LiveBench: A challenging, contamination-free benchmark for large language models (2024, arXiv) · Card data: answer level; feedback: programmatic, mixed Language Model Developers Should Report Train-Test Overlap (2024, arXiv) data: overlap and reporting analysis.; process: training corpus, evaluation set, overlap es…; feedback: overlap analysis rather than a reward model. |
static benchmark becomes a training target |
reading frontier reports as partial data-recipe disclosures · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 🚀 DeepSeek-R1 family | RLVR, distillation, reasoning traces, and public recipe disclosure | DeepSeek-R1 (2025, arXiv) · Card data: answer level; feedback: mixed |
report describes outcomes but not enough data partitions |
| 🌙 Kimi reasoning reports | long-context reasoning, RL compute, and frontier inference budgets | Kimi K1.5: Scaling Reinforcement Learning with LLMs (2025, arXiv) · Card data: answer level; feedback: mixed |
test-time compute is mixed with training-data effects |
| 🐉 Qwen reasoning/math/code reports | math, code, PRM, and open-weight reasoning model families | Qwen2.5-Math technical report: Toward mathematical expert model via self-improvement (2024, arXiv) · Card data: math solution, final answer, optional tool/code execution trace, and reward-model sco…; feedback: math answer checks, reward model signals, and benchmark evaluations. Qwen3 Technical Report (2025, arXiv) · Card data: answer level; feedback: mixed |
release cards do not separate SFT, RLVR, and evaluation data |
| 🧠 Magistral / Phi / Nemotron style reports | open-weight reasoning reports with partial data and reward disclosures | Llama-Nemotron: Efficient Reasoning Models (2025, arXiv) · Card data: answer level; feedback: mixed Magistral (2025, arXiv) · Card data: answer level; feedback: mixed MiniMax-M1: Scaling Test-Time Compute Efficiently with Lightning Attention (2025, arXiv preprint arXiv:2506.13585) · Card data: reasoning output, code/tool result, or agent task output; process: reasoning output,…; feedback: programmatic, environment, and benchmark feedback Phi-4-reasoning Technical Report (2025, arXiv) · Card data: answer level; feedback: mixed |
model-card claims cannot be mapped to concrete data objects |
| 🧪 RLVR recipe reports | reports that expose reward contracts, rollout policies, or RL scaffolds | DeepSeek-Prover-V2: Advancing formal mathematical reasoning via reinforcement learning (2025, arXiv) · Card data: subgoal chain, informal reasoning trace, Lean proof, and checker result.; process: pr…; feedback: Lean verification and RL reward over formal proof success. DeepSeek-Prover-V1.5: Harnessing proof assistant feedback for reinforcement learning and Monte-Carlo tree search (2024, arXiv) · Card data: Lean proof script, proof-search path, feedback signal, and verification result.; proc…; feedback: proof assistant feedback used for RL and search selection. Scaling Behaviors of LLM Reinforcement Learning Post-Training (2025, arXiv) · Card data: problem, generated solution/answer, reward outcome, and training curve metrics.; proc…; feedback: answer-level reward for mathematical reasoning and scaling curves. Prometheus 2: An open source language model specialized in evaluating other language models (2024, EMNLP) · Card data: rubric-conditioned scalar score, critique, or pairwise preference output.; process: i…; feedback: Prometheus 2 judge output aligned against human/proprietary-judge benchmark… |
RL gains are attributed without verifier coverage |
| 🧬 What is disclosed vs hidden | data sources, filters, lineage, safety mixtures, and undisclosed partitions | Skywork Open Reasoner 1 Technical Report (2025, arXiv preprint arXiv:2505.22312) data: metadata pending; feedback: metadata pending rStar2-Agent: Agentic Reasoning Technical Report (2025, arXiv preprint arXiv:2508.20722) data: survey background; feedback: metadata pending |
opaque mixtures are reused as open recipes |
leakage, contamination, verifier gaming, judge attacks, and reproducibility failures · Full track page
| Subfield | What this subfield studies | Representative papers with official links | Key audit risk |
|---|---|---|---|
| 🧯 Benchmark contamination | train/test overlap, stale evaluations, and benchmark refresh | LiveBench: A challenging, contamination-free benchmark for large language models (2024, arXiv) · Card data: answer level; feedback: programmatic, mixed LiveCodeBench: Holistic and contamination-free evaluation of large language models for code (2024, arXiv) · Card data: program submission or code-related output evaluated by tests or task-specific checks.…; feedback: programmatic tests and task-specific correctness checks. SciCode: A benchmark for scientific code generation and reasoning (2024, NeurIPS Datasets and Benchmarks) · Card data: code solution evaluated with scientist-annotated tests or expected outputs.; process:…; feedback: test cases and scientist-curated gold solutions. PRMBench: A fine-grained and challenging benchmark for process-level reward models (2025, arXiv) · Card data: step-level labels or scores; process: step, label, error type; offline reasoning trac…; feedback: process-level reward model benchmark |
memorized items are reported as reasoning progress |
| 🔍 Search-time contamination | contamination introduced by search, tools, retrieval, or inference scaffolds | Self-RAG: Learning to retrieve, generate, and critique through self-reflection (2023, ICLR) · Card data: generation with retrieval decisions, critique signals, and final answer.; process: qu…; feedback: critique signals and task-specific factuality/answer-quality evaluation. |
test-time tool access leaks answer traces |
| 🧬 Hidden lineage / teacher leakage | teacher-model traces, synthetic data inheritance, and hidden trait transfer | Subliminal Learning (2025, arXiv) · Card data: generated data plus downstream behavioral evaluation of the student.; process: teache…; feedback: trait probes after student training. Leaky Thoughts (2025, arXiv) · Card data: internal reasoning trace, final answer, and leakage/extraction outcome.; process: sen…; feedback: extraction probes and agentic evaluations. |
student behavior inherits undisclosed teacher artifacts |
| 🎮 Reward hacking | ways reward models, tests, or judges can be optimized as shortcuts | Spurious Rewards (2025, arXiv) · Card data: scalar reward; feedback: programmatic |
reward rises while real quality falls |
| 🧪 Verifier gaming | models exploiting checkers, answer formats, or judge blind spots | DeepMath-103K (2025, arXiv) · Card data: answer level; feedback: programmatic TinyV: Reducing False Negatives in Verification Improves RL for LLM Reasoning (2025, arXiv) · Card data: candidate answer with recovered reward decision; process: original verifier verdict,…; feedback: small LLM verifier augmenting rules |
verifier-passing examples are semantically wrong |
| ⚖️ LLM-as-judge attacks | one-token attacks, position bias, verbosity bias, and prompt attacks | One Token to Fool LLM-as-a-Judge (2025, arXiv) · Card data: scalar reward; feedback: judgment required Judging LLM-as-a-judge with MT-Bench and Chatbot Arena (2023, NeurIPS Datasets and Benchmarks) · Card data: model response, judge score, pairwise preference, or arena battle outcome.; process:…; feedback: strong model judge and human preference comparisons. Prometheus 2: An open source language model specialized in evaluating other language models (2024, EMNLP) · Card data: rubric-conditioned scalar score, critique, or pairwise preference output.; process: i…; feedback: Prometheus 2 judge output aligned against human/proprietary-judge benchmark… |
judge score changes for non-semantic reasons |
| 🧨 Spurious rewards | shortcut rewards, memorization-triggered rewards, and wrong-behavior correlations | Spurious Rewards (2025, arXiv) · Card data: scalar reward; feedback: programmatic |
reward improves while model learns a shortcut |
| 📉 Reproducibility failures | decoding, evaluation, scaffold, and data reporting failures | AbstentionBench (2025, arXiv) · Card data: model response, abstention decision, and correctness/abstention judgment.; process: s…; feedback: human-validated judges and benchmark labels for abstention scenarios. HealthBench (2025, arXiv) · Card data: response with rubric/judge evaluation; process: prompt, response, rubric dimension; o…; feedback: rubric-guided expert/LLM judgment Introducing SWE-bench Verified (2024, OpenAI / SWE-bench report) · Card data: patch diff applied to a repository plus test execution results.; process: repository,…; feedback: post-patch unit tests plus human filtering of task validity. RewardBench: Evaluating Reward Models for Language Modeling (2024, NeurIPS) · Card data: pairwise or scalar reward decisions; process: prompt, chosen/rejected response, rewar…; feedback: reward model or judge |
reported gains disappear under controlled reruns |
- 📚 Main Research Tracks
- 🧭 Background / Foundations
- 🧬 Core Reasoning Data Types
- 🧱 Instruction / Demo / Rationale: Instruction / Demo / Rationale
- 🤝 Preference & Reward Feedback: Preference & Reward Feedback
- 🧮 Programmatic Verification: Programmatic Verification
- 🪜 Process / Trace Supervision: Process / Trace Supervision
- 🔁 Rollout / Search / TTC Trace: Rollout / Search / TTC Trace
- 🌐 Environment & Agent Trajectories: Environment & Agent Trajectories
- ⚖️ Judgment / Rubric / Domain Expert: Judgment / Rubric / Domain Expert
- 🛠️ Data Lifecycle
- 🏗️ Construction & Open Releases: Construction & Open Releases
- 🎯 Training Usage & Objectives: Training Usage & Objectives
- 📈 Scaling / RLVR / TTC: Scaling / RLVR / TTC
- 🧰 Benchmarks & Evaluation: Benchmarks & Evaluation
- 🚀 Frontier Disclosure Ledger: Frontier Disclosure Ledger
- 🧯 Audit & Failure Modes: Audit & Failure Modes
- 🧩 Browse by Data Object
- Prompt-answer, trace-answer, step label, rollout value, preference pair, reward record, agent trajectory, rubric record
- 🛠️ Browse by Training Use
- SFT, distillation, reward modeling, process supervision, RLVR, agent training, evaluation, audit
Post-training reasoning data is multi-axis. A math paper can be a benchmark, an SFT trace release, a PRM source, an RLVR verifier, and a contamination risk at the same time. Use these four views before deciding where a paper belongs.
| View | Question | Examples | Use it when... |
|---|---|---|---|
| 🔍 By feedback contract | Who decides correctness? | programmatic, environmental, judgment-required, mixed | you need to understand the verifier/reward/judge/environment behind a paper. |
| 📦 By data object | What is serialized? | answer, trace, step label, preference pair, reward, trajectory, rubric | you need to compare what the dataset actually stores. |
| 🛠️ By training use | How does it enter post-training? | SFT, distillation, PRM, RM, RLHF, RLVR, agent training, evaluation | you need to map papers to an engineering pipeline. |
| 🧪 By task domain | Where does it operate? | math, code, proof, tools, SWE, web, medical, safety, legal, finance | you need a domain-specific literature route. |
| Research question | Best entry |
|---|---|
| What counts as post-training reasoning data? | docs/01 + Foundations |
| How do we verify reasoning data? | Programmatic + Process supervision + Verifiers |
| How are open reasoning datasets constructed? | Construction recipes + Release cards |
| What data does RLVR actually need? | Programmatic verification + Scaling/RLVR |
| How should agent trajectories be serialized? | Agent data + docs/07 |
| How do frontier reports disclose or hide data recipes? | Frontier reports |
| How do contamination and verifier gaming affect claims? | Audit/failure modes |
| Which benchmarks are still useful for reasoning data? | Benchmarks and evaluation |
| Learning goal | What this repo gives you |
|---|---|
| 🧠 Build the mental model | Understand why reasoning data is not just prompt -> answer, but a record with traces, actions, feedback, and metadata. |
| 🧮 Understand verifiable reasoning data | Learn how math answers, code tests, theorem provers, and executable environments create training and evaluation signals. |
| 🪜 Understand process supervision | Compare outcome rewards, step labels, process reward models, rollout values, and first-error localization. |
| 🏗️ Understand data construction recipes | Track prompt sourcing, teacher generation, search, filtering, deduplication, decontamination, and release metadata. |
| 🌐 Understand agent trajectory data | Learn what must be stored for tool use, browser tasks, app worlds, OS tasks, and repository-level SWE episodes. |
| ⚖️ Understand judges and rubrics | Study rubric-conditioned evaluation, open evaluator models, reward models, human preference data, and judge attacks. |
| 📈 Understand scaling and RLVR claims | Separate data scale, verifier quality, optimization scaffold, and inference budget when reading frontier reports. |
| 🧯 Learn how to audit failures | Look for leakage, contamination, verifier gaming, reward hacking, hidden lineage, and benchmark fragility. |
| Reader | Best path through the repo |
|---|---|
| New student / newcomer | Start with the 60-second model, then read the Starter Pack and the first two docs pages. |
| Researcher entering post-training | Use the paper atlas to locate subfields, then read L5 cards before opening full papers. |
| Dataset builder | Follow the construction stack and release-card checklist before building or releasing data. |
| RLVR / verifier engineer | Use the verifier audit sections, process-supervision papers, and programmatic benchmark cards. |
| Agent researcher | Follow the agent trajectory section and compare SWE-bench, SWE-bench Verified, ReAct, Toolformer, and environment cards. |
| Reading group organizer | Use the Starter Pack and category pages as a week-by-week syllabus. |
| Open-source contributor | Add verified links, metadata, cards, or missing artifacts through the contribution workflow. |
Post-training reasoning data is the data used after pretraining to teach, reinforce, or audit reasoning behavior.
A useful sample is usually not only:
prompt -> answerbut:
task/context -> trace/actions -> answer/artifact -> verifier/reward/judge/environment -> metadataThis repo helps you compare those records across math, code, proof, agents, rubric judging, frontier model reports, scaling studies, and failure audits.
Read this repository if you want to answer questions like:
- 🧪 What exactly verified the answer: unit tests, a proof checker, a reward model, an LLM judge, a rubric, or an environment?
- 🪜 Was feedback attached to the final answer, each step, a rollout set, a state-action transition, or a full episode?
- 🧬 Which teacher, base model, prompt source, filtering rule, split, license, and contamination check produced the released data?
- 📈 Did a result improve the asymptote, the sample efficiency, the inference budget curve, or only the reported pass rate?
- 🧯 Where can the verifier fail, leak, overfit, reward-hack, or silently encode lineage artifacts?
| Section | What you will learn | Go |
|---|---|---|
| 🧭 Start Here | Zero-to-field overview and reading paths | docs/00 |
| 🎯 What You Can Learn | The repository as a learning roadmap | jump |
| 🧑💻 Who It Is For | Paths for students, builders, researchers, and auditors | jump |
| 🧠 60-second Model | The verifier-bearing sample mental model | jump |
| 🔥 Latest Updates | What changed recently in this atlas | jump |
| 🧭 Research Tracks | Browse the field like an Awesome paper atlas | jump |
| 📚 Detailed Paper Directory | Subfield-level paper links with data/feedback hints | jump |
| 🧩 Four Views | Feedback contract, data object, training use, and domain | jump |
| 🔎 Research Questions | Jump from a question to the right paper track | jump |
| 📊 Snapshot | Current verified/card/artifact coverage | jump |
| 🛣️ Learning Roadmap | Learn the field in 6 stages | jump |
| 🧭 Starter Pack | 20 papers to read first | jump |
| 🧮 Core Paper Map | The compact map from data objects to papers | jump |
| 🗺️ Category Map | Programmatic, environmental, judgment-required, scaling, audit | jump |
| 🧰 Build Data | Construction stack for reasoning datasets | jump |
| 🧪 Audit Verifiers | How to inspect rewards, judges, checkers, and rubrics | jump |
| 🌐 Agent Trajectories | State/action/replay fields for tools, web, OS, SWE | jump |
| 📈 Scaling Claims | RLVR, reuse, pass@k, test-time compute, inference budget | jump |
| 🧩 Repo Structure | How files, docs, cards, and reports fit together | jump |
| 📚 Paper Atlas | Category pages, cards, exports, searchable website | jump |
| 🌱 Roadmap | High-impact priorities for making the atlas more citable | ROADMAP |
| 🤝 Contribute | Add papers with metadata, not only links | CONTRIBUTING |
| Metric | Count |
|---|---|
| Total structured entries | 280 |
| Verified official primary links | 165 |
| Entries with paper/arXiv/venue/DOI links | 165 |
| Unique entry-linked cards | 87 |
| Card files | 89 |
| L5 audit-ready entries | 53 |
| Needs search / metadata | 115 |
| Official code links | 41 |
| Official data links | 27 |
| Hugging Face links | 20 |
| Project links | 25 |
| I want to... | Go to |
|---|---|
| 🧭 understand the field | docs/00_start_here.md |
| 📚 find papers by subfield | papers/README.md |
| 🧮 study math/code/proof data | papers/01_core_reasoning_data_types/03_programmatically_verifiable_outcome_data.md |
| 🪜 study process supervision | papers/01_core_reasoning_data_types/04_process_trace_supervision_data.md |
| 🌐 study agent trajectories | papers/01_core_reasoning_data_types/06_environment_agent_trajectory_data.md |
| 🚀 study frontier model reports | papers/02_data_lifecycle/12_frontier_reports_data_disclosure_ledger.md |
| 🔎 use the searchable atlas | live atlas |
| 📊 inspect link coverage | reports/link_coverage.md |
| 🤝 contribute a paper/card | CONTRIBUTING.md |
This repository should work like a small open course. You do not need to read every paper first. Use the route below and open papers only when a concept becomes important.
| Stage | Learn | Main resources | Output you should have |
|---|---|---|---|
| 1 | Vocabulary and mental model | 60-second version, docs/00, docs/01 | You can explain the difference between answer data, trace data, reward data, verifier data, and trajectory data. |
| 2 | Feedback contracts | docs/02, docs/06, Verifier cards | You can identify whether a work uses programmatic, environmental, judgment-required, or mixed verification. |
| 3 | Core papers | Starter Pack, papers/README.md, cards/README.md | You can locate the canonical papers for math, code, process supervision, agents, RLVR, and audit. |
| 4 | Data construction | docs/05, Release cards, Recipe cards | You can describe prompt sourcing, teacher generation, filtering, verifier pinning, and release metadata. |
| 5 | Specialized tracks | programmatic data, agents, rubrics, scaling | You can choose a subfield and follow its top papers and audit questions. |
| 6 | Audit and contribution | docs/09, reports/link_coverage.md, CONTRIBUTING.md | You can tell what is verified, what is missing, and how to improve an entry without hallucinating links. |
Read these as a learning path, not as a citation dump. The rightmost columns tell you what question each paper should answer before you move on.
| # | Paper / report | Lens | Start with this question | Card |
|---|---|---|---|---|
| 1 | Datasheets for datasets | 📋 release docs | What must be disclosed before anyone reuses a dataset? | Card |
| 2 | Data statements for natural language processing | 🧬 provenance | Which population, language, and annotation assumptions travel with the data? | Card |
| 3 | Training language models to follow instructions with human feedback | 🧑🏫 RLHF pipeline | How do demonstrations, preferences, rewards, and policy optimization separate? | Card |
| 4 | Chain-of-thought prompting elicits reasoning in large language models | 🧠 traces | When does a rationale become a reusable training object? | Card |
| 5 | Training verifiers to solve math word problems | 🧪 verifier | What exactly scores a generated solution? | Card |
| 6 | STaR: Bootstrapping reasoning with reasoning | 🔁 self-improvement | Which generated traces survive answer-based filtering? | Card |
| 7 | Self-Instruct: Aligning language models with self-generated instructions | 🏗️ synthetic data | How do generated instructions get filtered before training? | Card |
| 8 | Direct preference optimization: Your language model is secretly a reward model | ⚖️ preference data | What changes when preference pairs directly train the policy? | Card |
| 9 | Let's Verify Step by Step | 🪜 process supervision | What does step-level feedback buy over outcome-only labels? | Card |
| 10 | GSM8K: Grade School Math 8K | 🧮 answer checks | Why is a small verifiable math set still a useful sanity check? | Card |
| 11 | Measuring mathematical problem solving with the MATH dataset | 🧮 hard math | How do harder problems change trace and verifier requirements? | Card |
| 12 | HumanEval: Hand-Written Evaluation Set | 💻 unit tests | What makes executable tests a feedback contract? | Card |
| 13 | SWE-bench: Can language models resolve real-world GitHub issues? | 🌐 agent environment | What fields make repository repair a replayable episode? | Card |
| 14 | RewardBench: Evaluating Reward Models for Language Modeling | 🏅 reward eval | When does a reward model fail outside generic chat helpfulness? | Card |
| 15 | HealthBench | ⚕️ rubrics | How do high-stakes rubrics become auditable? | Card |
| 16 | LiveBench: A challenging, contamination-free benchmark for large language models | 🧯 contamination | How can benchmarks stay fresh against memorization? | Card |
| 17 | OpenThoughts: Data recipes for reasoning models | 🏗️ open recipe | Which prompt, trace, filtering, and ablation fields are disclosed? | Card |
| 18 | DeepSeek-R1 | 🚀 RLVR report | What can and cannot be inferred from a public frontier report? | Card |
| 19 | s1: Simple Test-Time Scaling | ⏱️ test-time compute | When is inference budget part of the data story? | Card |
| 20 | A Sober Look at Progress in Language Model Reasoning: Pitfalls and Paths to Reproducibility | 🔍 reproducibility | Which reported gains survive decoding and evaluation audits? | Card |
Next steps:
- Newcomer: read docs/00_start_here.md and docs/01_what_is_post_training_reasoning_data.md.
- Builder: read docs/05_construction_cookbook.md and compare release cards in cards/releases/.
- Auditor: read docs/09_audit_and_failure_modes.md and compare three L5 cards from cards/README.md.
| Cluster | Representative entries | What to inspect |
|---|---|---|
| 🧮 Programmatic math/code/proof | OpenMathInstruct-2, DeepSeek-Prover-V2, SciCode | answer checker, unit tests, proof checker, pass-rate bands, decontamination |
| 🪜 Process supervision and PRMs | Let's Verify Step by Step, Math-Shepherd, Rewarding Progress | step labels, rollout values, first-error localization, reward-model calibration |
| 🏗️ Open construction recipes | OpenThoughts, Self-RAG, Magicoder | prompt source, teacher trace, filtering rule, optimizer/scaffold, ablation fields |
| 🚀 Frontier and model reports | DeepSeek-R1, Qwen2.5-Math, Tulu 3 | disclosed data partitions, reward contract, RLVR setup, distillation path |
| 🌐 Agent and environment data | SWE-bench, SWE-bench Verified, ReAct | state/action/observation schema, terminal predicate, replayability, scaffold metadata |
| ⚖️ Judgment-required rubrics | HealthBench, RewardBench, Prometheus 2 | rubric provenance, judge prompts, adjudication, domain expertise, calibration |
| 🧯 Audit and failure modes | LiveBench, A Sober Look, TruthfulQA | leakage, contamination, verifier gaming, judge attack, hidden trait transfer |
A reasoning-data taxonomy should start from the feedback contract, not only the academic domain. The same math problem can be an SFT trace, an RLVR answer record, a PRM step record, a rejection-sampling candidate, or a contamination probe.
| Axis | Values | Reader question |
|---|---|---|
| 🧪 Verification contract | programmatic, environmental, judgment-required, mixed, unknown | Who or what says the sample is correct? |
| 🪜 Granularity | answer, step, transition, full episode, rollout set, scalar reward | Where does feedback attach? |
| 🧩 Data object | prompt-answer, trace-answer, PRM record, preference pair, trajectory, rubric record | What fields must be serialized? |
| 🧬 Lineage | human, teacher model, search, self-play, environment, synthetic mix | Where did the behavior come from? |
| 🧰 Training use | SFT, distillation, reward modeling, RLVR, agent training, evaluation, audit | How could this data enter a post-training pipeline? |
| 🧯 Risk | leakage, contamination, verifier failure, judge attack, reward hacking, license ambiguity | What can make the gain misleading? |
Use the construction stack from docs/05_construction_cookbook.md:
| Layer | Builder checklist | Common evidence |
|---|---|---|
| 1. 📥 Prompt sourcing | Describe source mix, license, split, difficulty, and base-model pass rate. | prompt pool, dedupe report, contamination checks |
| 2. ✍️ Trace writing | Say whether traces are human-written, teacher-generated, search-generated, or self-played. | teacher model, sampling temperature, rollout count |
| 3. 🔍 Search substrate | Record beam/search/MCTS/self-critique/scaffold details. | search budget, candidate count, pruning rules |
| 4. 🧪 Verifier layer | Pin the checker, judge, environment, rubric, or reward model. | tests, proof checker version, judge prompt, rubric |
| 5. 🧹 Filtering | Keep pass/fail bands, rejection reasons, and ambiguous cases. | filter script, verifier logs, disagreement set |
| 6. 🏋️ Optimizer/scaffold | State whether data is used for SFT, distillation, RLVR, PRM, or agent training. | loss, reward, rollout policy, curriculum |
| 7. 🧬 Release metadata | Preserve attribution, lineage, splits, license, and known failure modes. | card, datasheet, citation, changelog |
Minimal release question: Could a different team reproduce the data object, rerun the verifier, and explain what would fail if the verifier were wrong?
A verifier is not just a score. It is a data-producing instrument.
| Verifier type | Audit focus | Red flag |
|---|---|---|
| 🧮 Answer checker | canonicalization, tolerance, symbolic equivalence | formatting hacks count as reasoning gains |
| 💻 Unit tests | hidden tests, flaky tests, generated tests, coverage | model learns test style rather than task skill |
| 🧾 Proof checker | version, imports, tactic environment, timeout | proof succeeds only under an undocumented environment |
| 🪜 PRM | step boundary, label policy, calibration, rollout values | reward rises while final correctness falls |
| ⚖️ Rubric judge | rubric source, domain expertise, adjudication, prompt | judge is sensitive to wording or verbosity |
| 🧑⚖️ LLM-as-judge | model version, prompt, decoding, attack suite | one token or style cue flips the verdict |
| 🌐 Environment | terminal predicate, reset, observation schema, replay | success transcript cannot be replayed |
Agent data is more than a cleaned success transcript. A trainable or auditable episode should expose the environment contract.
| Field | Why it matters |
|---|---|
| 🧭 Task and initial state | Defines what the agent was actually asked to solve. |
| 👀 Observation stream | Separates visible context from hidden evaluator state. |
| 🛠️ Action schema | Makes tool, browser, OS, code, or API calls inspectable. |
| ⏱️ Budget | Records step limit, time, token budget, and retries. |
| 🧪 Terminal predicate | States exactly how success or failure is decided. |
| 🔁 Replay metadata | Lets another team re-run the episode and verify the result. |
| 🧱 Scaffold metadata | Captures planner, memory, retrieval, tool wrapper, and guardrails. |
| 🧯 Failure trace | Keeps near-misses and verifier failures instead of deleting them. |
Scaling claims become much clearer when you treat the training data, verifier, and inference budget as part of the same ledger.
| Claim | Ask for | Watch out |
|---|---|---|
| RLVR improves reasoning | reward contract, verifier coverage, base-model pass rate | reward hacking or easy-example filtering |
| More data improves performance | unique examples, reuse count, source mixture | repeated prompts counted as fresh data |
| Test-time compute scales | pass@k, pass@(k,T), budget, search topology | hidden inference budget changes |
| Distillation transfers reasoning | teacher identity, trace policy, filtering | teacher leakage or style imitation |
| Frontier report shows recipe | data partitions, curricula, ablations | optimizer details without data details |
| Path | What it is for |
|---|---|
| docs/ | Conceptual lessons: mental model, taxonomy, construction cookbook, verifiers, agent trajectories, scaling, and failure modes. |
| papers/ | Field navigation map: category pages with read-first tables, full paper lists, audit checklists, related cards, and open gaps. |
| cards/ | Learning cards: paper/data/verifier/recipe/benchmark/failure summaries with links and audit questions. |
| data/papers.yaml | Structured source of truth for paper metadata, roles, contracts, summaries, links, and curation levels. |
| docs/index.html | Searchable web atlas generated from structured data. |
| reports/ | Public QA and coverage: link coverage, needs-search, release notes, quality audits, and live-link reports. |
| exports/ | CSV, JSON, and BibTeX exports for readers who want to reuse the atlas data. |
| scripts/ | Reproducible generators and validators for README, paper pages, cards, site data, exports, and QA. |
| ROADMAP.md | Public priorities for making the atlas more useful, citable, and contribution-friendly. |
| If your question is... | Use this path |
|---|---|
| "I am new. What should I read first?" | Start with docs/00, then the Starter Pack. |
| "I want to build a reasoning dataset." | Read docs/05, then inspect release cards and recipe cards. |
| "I want to know whether a benchmark is reusable." | Open the relevant benchmark card, then check its verifier, data split, contamination risk, and official links. |
| "I want to understand RLVR." | Follow programmatic math/code/proof papers, verifier cards, and scaling/RLVR category pages. |
| "I want to study agents." | Follow agent papers, then inspect action schema, terminal predicate, and replay fields. |
| "I want to contribute." | Pick an item from needs_search, verify official links, then add structured metadata and a card. |
The repository becomes more useful and citable when it improves depth, trust, and reuse rather than raw length. The public roadmap is in ROADMAP.md.
| Priority | What to improve next | Why it helps readers cite or reuse the atlas |
|---|---|---|
| P0 | Keep public hygiene clean: no private planning files, prompt/spec artifacts, or local OS metadata. | Readers should see a maintained research resource, not a build workspace. |
| P1 | Promote high-impact L1_link_verified entries into L4_carded or L5_audit_ready. |
Deep cards are what make the repo useful beyond a paper list. |
| P1 | Add official code, data, Hugging Face, and project links for already verified papers. | Builders can jump from survey reading to reusable artifacts. |
| P1 | Strengthen thin subfields before adding long-tail seeds. | Researchers can trust the taxonomy as a balanced field map. |
| P2 | Improve bilingual polish and paper-specific citation metadata. | The atlas becomes easier to share in reading groups, surveys, and course notes. |
The long categorized lists live in papers/. Each category page includes a category explanation, read-first table, full paper list, audit checklist, related cards, and open gaps.
- Foundations and Primers
- Instruction, Demonstration, and Rationale Data
- Preference and Reward Feedback Data
- Programmatically Verifiable Outcome Data
- Process and Trace Supervision Data
- Rollout, Search, and Test-Time Trace Data
- Environment and Agent Trajectory Data
- Judgment, Rubric, and Domain-Expert Data
- Data Construction and Open Release Recipes
- Training Usage and Optimization Objectives
- Scaling, RLVR, and Test-Time Compute
- Benchmarks and Evaluation Surfaces
- Frontier Reports and Data Disclosure Ledger
- Audit, Failure, Contamination, and Verifier Attacks
Cards turn citations into engineering decisions. They explain the data object, verifier/reward/environment, construction recipe, post-training use, audit questions, and risks.
- Card index
- Release cards
- Verifier cards
- Agent/environment cards
- Recipe cards
- Failure/audit cards
- Benchmark cards
Open the live atlas or docs/index.html locally. The site loads generated JSON on hosted pages and includes docs/assets/atlas-data.js as a local fallback for browsers that block direct JSON loading. It supports search plus filters for year, venue, source role, verification contract, supervision granularity, training use, curation level, status, and artifact availability.
| Level | Meaning |
|---|---|
L0_seeded |
Only a title or bibliography seed is known. |
L1_link_verified |
Official paper, arXiv, venue, or DOI link is pinned. |
L2_artifact_verified |
Code, data, project, or model artifact links are also checked. |
L3_summary_ready |
One-line summary and why-it-matters rationale are present. |
L4_carded |
A local card explains data object, verifier, use, and audit fields. |
L5_audit_ready |
The card includes concrete risks, missing fields, and audit questions. |
Please do not submit only a paper title. A useful contribution includes official links, source role, verification contract, supervision granularity, training use, a one-line summary, a why-it-matters rationale, and card/audit fields when available. Start with CONTRIBUTING.md.
If this atlas helps your related work, dataset construction, verifier design, or reading group, please cite the companion paper and link this repository. See CITATION.cff.
MIT. See LICENSE.