TimeRLM: Recursive Language Models Are Precise Temporal Reasoners

Code and benchmark for TimeRLM, a Recursive Language Model formulation for time-series, and AnomalyXL, a synthetic long-context anomaly-localization benchmark released together with its generator.

🤗 TimeRLM / AnomalyXL datasets on Hugging Face

TimeRLM treats a long recording as an environment variable rather than something to ingest: the signal lives in a file in a sandboxed Python REPL, and the model queries it over multiple turns with numpy/scipy — and, in the image variant, with a look() tool that renders plots back into context — before committing a structured, evidence-grounded answer. This lets it operate on the full, un-decimated series instead of a single compressed or subsampled view.

The orchestrator reasons in natural language and emits Python into a sandboxed REPL, where the signal is an environment variable. numpy and scipy are common to all variants; including matplotlib exposes a look() tool that inserts plots into the model's context. A FINAL() action commits a grounded answer. (vector version)

What's here

Directory	What it is
anomalyXL/	The AnomalyXL generator. Five precise task families (localize, classify-with-evidence, measure-magnitude, localize-all-channels, lead–lag), each with a coarse MCQ projection over the same signal. Also the ECG/sleep generalization-set builders and the classical no-LLM baseline.
environments/timeseries_qa/	The evaluation environment (verifiers, id timeseries-qa). One env, four inference-time configurations: single-pass text, single-pass image, TimeRLM text, TimeRLM image. Also the ARFBench and generalization-set loaders.
environments/timeseries_qa_v1/	The RL training environment (id timeseries-qa-v1): the generator-backed taskset, the reward (the benchmark's own scorer, plus the −0.25 turn-budget penalty), and the harness that drives the agent.
rlm-harness/	The agent loop: a persistent IPython sandbox with auto-compaction and a turn budget. Vendored from Prime Intellect's rlm — see rlm-harness/NOTICE.
configs/rl/	The GRPO post-training config behind the paper's RL row.
sft/tslm/	The TSLM baselines: standalone SFT + eval for ChatTS, OpenTSLM-Flamingo and ITFormer, all on a frozen Chronos-2 encoder feeding a LoRA-tuned Qwen3.5-4B.
scripts/	Bootstrap, data splitting, checkpoint eval, and the generalization-set scorer.

Two RLM code paths exist by design and are both load-bearing: evaluation runs through environments/timeseries_qa (which owns the look() vision tool), while RL training runs through environments/timeseries_qa_v1 + rlm-harness.

Quickstart

bash scripts/setup.sh                  # prime-rl + the three packages
cd anomalyXL && uv sync
uv run python scripts/generate_anomalyxl_sweep.py --out-root datasets --precise-only --seed 42
uv run pytest tests/ -q                # 157 tests, no GPU needed

The full walkthrough — building the benchmark, building the real ECG/sleep generalization set, running all four inference-time configurations, and reproducing the GRPO run — is in docs/REPRODUCE.md.

The benchmark

Five task families — localize, classify, measure magnitude, localize across all channels, and lead–lag — each posed in two paired formulations over the same signal: a precise open-ended version scored on class and IoU, and a coarse multiple-choice projection. (vector version)

AnomalyXL is synthesized programmatically, so every gold label is exact and strict scoring is well-defined. Difficulty is configurable along context length (16k–131k samples), channel count (1–16), and anomaly kind, and 20% of samples are first-class negatives. Anomaly width is an absolute sample count (8–256) regardless of series length — which is what makes long context hard: at L = 131,072 an anomaly can occupy as little as 0.006% of the series.

Every item is posed twice over the identical signal: a precise, open-ended version scored by a continuous metric (temporal IoU, kind×IoU, relative error, set-matched F1, or a direction-and-lag composite), and a MCQ multiple-choice projection. Holding the signal fixed isolates the effect of evidence-grounded scoring from that of the multiple-choice format, with no different-signal confounder.

The released shards — the AnomalyXL monorepo (Localize precise + MCQ coarse partitions) and the real ECG/sleep anomalyxl-generalization set — are collected on 🤗 Hugging Face.

What's not included

To be explicit about the boundary of this artifact:

• The prime-rl patches used for the paper's RL run. scripts/setup.sh clones upstream prime-rl; the exact commit and patch set are not pinned here (see docs/REPRODUCE.md).
• The PhysioNet preprocessing that produces the derived layout the generalization-set builders consume. Use the released Hub shards instead.
• Model weights. The fine-tuned LoRA adapters are not committed; sft/toto/eval.py and the TSLM configs train them from scratch.

Every baseline row in the paper is reproducible from this repo: the three TSLMs (sft/tslm/), Toto (sft/toto/), and the classical no-LLM detector (anomalyXL/scripts/classical_baseline.py). All four are scored by the same scorer as TimeRLM — enforced by environments/timeseries_qa/tests/test_scorer_parity.py, which fails if the subprojects' vendored copies ever drift apart.

Acknowledgements

• rlm (PrimeIntellect-ai/rlm) — the recursive agent harness, vendored here as rlm-harness/ (see its NOTICE), together with prime-rl and verifiers.
• PhysioNet — the Long Term AF Database and the 2018 Sleep-PSG challenge set, from which the generalization windows are spliced.

License

MIT — see LICENSE. rlm-harness/ is a vendored copy of Prime Intellect's rlm (MIT, see rlm-harness/NOTICE); rlm_env_vendored.py is frozen from verifiers (MIT); ARFBench and the PhysioNet recordings retain their own licenses.