Skip to content

References

Jump to bottom
Gabri Elles edited this page Aug 21, 2026 · 3 revisions

References

The bibliography behind DeepDelve's architecture decisions. Used to be README.md's "References" section. Full detail and still open leads live on the Literature Review page.

Why local models are hard for this task, briefly. DeepDelve's own model bake off, 10 candidates, 9 disqualified, gpt-oss:20b the only full pass, isn't an idiosyncratic gap. A published capacity floor study puts qwen2.5:14b at the "minimum viable production" threshold for tool invocation, with sub 8B models failing 40 to 85 percent of the time even on simpler tasks (Huang, Malwe, Wang, arXiv:2601.16280). Two other studies found small models specifically fail at structured tool call serialization in a way fine tuning can't fix, since it happens downstream of anything fine tuning touches (Li, Zhang, Lv, arXiv:2606.25605; Ray, arXiv:2605.26128). MAST's 14 mode failure taxonomy (Cemri et al., arXiv:2503.13657, NeurIPS 2025) maps closely onto DeepDelve's own bug catalog, evidence these are known, published patterns, not project specific quirks.

Papers

  • Huang, Malwe, Wang. When Agents Fail to Act. arXiv:2601.16280. Capacity floor evidence, 14B as the tool invocation threshold, part of why gpt-oss:20b over a smaller default.
  • Li, Zhang, Lv. Constraint Tax in Open Weight LLMs. arXiv:2606.25605. Confirms DeepDelve never mixes JSON schema output with tool availability, the exact combo that zeroes tool invocation rate.
  • Ray. The Constraint Tax. arXiv:2605.26128. Originating paper for the constraint tax concept, "reason free, constrain late."
  • Cemri, Pan, Yang, et al. Why Do Multi Agent LLM Systems Fail? (MAST). arXiv:2503.13657, NeurIPS 2025. 14 mode failure taxonomy matching DeepDelve's own bug catalog.
  • Cemri, Cojocaru, Pan, et al. Fantastic Adaptive Taxonomies and How to Use Them (ATLAS/AdaMAST). ICML 2026. ATLAS. Induces a domain specific taxonomy from a system's own traces rather than MAST's fixed vocabulary.
  • Liu, Lin, Hewitt, et al. Lost in the Middle. arXiv:2307.03172, TACL 2024. Candidate mechanism for DeepDelve's "content vanishes during synthesis" pattern.
  • Li, Zhang, Zhang, Chaturvedi. Coverage based Fairness in Multi document Summarization. arXiv:2412.08795, NAACL 2025. Validates the per task coverage design behind check_report_underuses_evidence.
  • Huang, Ma, Tahir, Zhang, Xiao, Xiao. Regression Accumulation in Multi Turn LLM Programming Conversations. arXiv:2607.01855. Informed the check priority shadowing fix (utils/grounding.py::cheap_grounding_problems).
  • Kamoi, Zhang, Zhang, Han, Zhang. When Can LLMs Actually Correct Their Own Mistakes? arXiv:2406.01297, TACL. Methodology critique, not the source of the 64.5% self correction figure (that's Tsui below, corrected 2026-08-17).
  • Tsui. Self Correction Bench. arXiv:2507.02778, COLM 2026. Source of the "Wait" prepend idea for _dispatch_writer_review_fix's retry instructions, 89.3% blind spot reduction.
  • Song Tae-Eun. Cross Context Review. arXiv:2603.12123. Validates PEER_REVIEWER_INSTRUCTIONS's artifact only review shape.
  • Xu, Zhu, Wang, et al. When Does Divide and Conquer Work for Long Context LLM? arXiv:2506.16411, ICLR 2026. Names "aggregator noise" as the mechanism behind DeepDelve's facet abandonment pattern.
  • Khanal, Tao, Zhou. Beyond pass@1. arXiv:2603.29231. Source of the pass@k/pass^k reliability framing in eval/evaluate.py.
  • Mustahsan, Lim, Anand, Jain, McCann. Stochasticity in Agentic Evaluations. arXiv:2512.06710, AAAI 2026. Why --runs 3 is a practical floor, not statistical convergence.
  • Jwalapuram, Lin, Li, et al. The Illusion of Multi Agent Advantage. arXiv:2606.13003. Confirms DeepDelve's fixed pipeline, not dynamic per query routing, is the architecture that actually wins.
  • Su, Wu. When Do Multi Agent Systems Outperform? arXiv:2602.08272. Maps DeepDelve's error concentration point (consolidation, not per facet research) onto PAC learning theory.
  • Jiang, Yang, Cui, et al. Deep Research in Physical Sciences (DelveAgent / PhySciBench). arXiv:2606.18648. Primary architecture source.
  • Huang, Chen, Zhang, et al. Deep Research Agents: A Systematic Examination and Roadmap. arXiv:2506.18096
  • Xu, Peng. A Comprehensive Survey of Deep Research. arXiv:2506.12594
  • Xi, Lin, Xiao, et al. A Survey of LLM based Deep Search Agents. arXiv:2508.05668
  • Del Rosario, Krawiecka, Schroeder de Witt. Architecting Resilient LLM Agents. arXiv:2509.08646. Source of the plan then execute pattern behind Builder's review/fix loop.
  • Min, Krishna, Lyu, et al. FActScore. arXiv:2305.14251, EMNLP 2023. Decompose then verify pattern behind the grounding layer's line scoped checks.
  • Goswami, Kurra. HALT RAG. arXiv:2509.07475. Source of "layer NLI on top of lexical checks, don't replace them."
  • Rasheed, Banerjee, Mukherjee, Hazra. From Fluent to Verifiable. arXiv:2602.13855. Source of the claim node framing behind decompose_claim_segments.
  • Thorne, Vlachos, Christodoulopoulos, Mittal. FEVER. NAACL 2018, fever.ai. Claim/evidence framing behind find_cross_source_contradictions.
  • Santos, Peng, Macdonald, Ounis. Explicit Search Result Diversification through Sub Queries. ECIR 2010. Source of _diversity_rerank's greedy marginal coverage pattern.
  • Anthropic. How we built our multi agent research system. anthropic.com/engineering/multi-agent-research-system. Independent confirmation multi agent research beats single agent by 90.2%.

Reference implementations

DeepDelve Wiki Home · Main repo · README

Home

History

Model Research

Reviews & Audits

Reference

Main repo

Clone this wiki locally