--- pretty_name: RE-PoLar license: cc-by-nc-4.0 language: - en tags: - llm - layer-skipping - layer-repetition - programs-of-layers - mcts - routing - reproduction - math-reasoning task_categories: - text-generation size_categories: - 10MA router reads the prompt and picks a per-layer skip/keep/repeat program before the frozen transformer runs.

## Explore the search trees The dataset viewer above shows the searches as tables: pick `programs_` for the programs and verdicts of each question (split = search seed), `answers_` for the generated answers, `polar_split_programs_` for the PoLar split. The search trees themselves, node by node, are on the [project page's interactive explorer](https://datexis.github.io/RE-PoLar/). ## Layout ``` mcts_results//seed/dart-math-diff-/merged_mcts_samples.json.gz programs + verdicts per question generated_answers//seed/diff.jsonl.gz one generated answer per evaluation router_checkpoints//seed/diff.pt trained routers (weights only) router_results//seed/diff.json router test results, per question fixed_program//seed/diff.json fixed-program baseline, per question menu_crossexec//diff.json.gz top-100 programs x 2,000 questions menu_crossexec_heldout//diff.json.gz top-100 by train+val, run on the 500 test questions polar_split/... same artifacts under //seed/ cross_checks/{their_search_our_split,our_search_polar_split}/... DATA.md full schema of every file ``` ``: `qwen15_moe_a27b`, `qwen25_3b`, `qwen25_7b`, `qwen3_8b`, `qwen3_32b`, `llama32_3b`; ``: search seed 42, 1, 7; ``: difficulty level 1-5. [`DATA.md`](DATA.md) documents every field, the search configurations and which runs have no answer texts. ## Download and use The code expects the files under `data/` of a clone of the repository: ```bash git clone https://github.com/DATEXIS/RE-PoLar && cd RE-PoLar pip install -U huggingface_hub hf download SWAMBIT/re-polar --repo-type dataset --local-dir data # everything, ~6 GB hf download SWAMBIT/re-polar --repo-type dataset --local-dir data --include "router_results/*" "fixed_program/*" ``` One file with the `datasets` library, or directly: ```python from datasets import load_dataset programs = load_dataset("SWAMBIT/re-polar", "programs_qwen3_8b", split="seed42") print(programs[0]["sample_info"]["query_id"], len(programs[0]["final_valid_transitions"])) import gzip, json router = json.load(open("data/router_results/qwen3_8b/seed42/diff1.json")) print(router["overall"]) # Base tau=0, router p@1 and p@5 on the 500 test questions with gzip.open("data/generated_answers/qwen3_8b/seed42/diff1.jsonl.gz", "rt") as f: print(json.loads(next(f))) ``` ## Questions are not included No file holds question text or reference answers, only `query_id`s. The MATH half of the DART-Math pool traces back to the delisted `hendrycks/competition_math`, so the questions are rebuilt locally from the public pool with the repository's code and joined on `query_id` (`re_polar.datasets.dart_math`, `re_polar.datasets.polar_split`, `re_polar.datasets.attach_mcts_questions`; see the repository README). The models sometimes restate their question in a generated answer; such restatements are replaced by `[question]` (DATA.md, "Generated text"). ## Citation ```bibtex @article{westerhoff2026repolar, title={Programs-of-Layers in LLMs through the Lens of Cortical Areas}, author={Westerhoff, Justus and Olbrich, Stephan and Oraby, Hatem and Senn, Walter and Larkum, Matthew Evan and Gers, Felix Alexander}, journal={arXiv preprint arXiv:2609.31360}, year={2026} } ``` ## License **CC BY-NC 4.0.** The generated answers and verdicts are outputs of the models above and also fall under their licenses: Qwen1.5-MoE-A2.7B-Chat (Tongyi Qianwen License), Qwen2.5-3B-Instruct (Qwen Research License), Qwen2.5-7B-Instruct, Qwen3-8B and Qwen3-32B (Apache 2.0), LLaMA-3.2-3B-Instruct (Llama 3.2 Community License; Built with Llama). The routers embed questions with Qwen3-Embedding-0.6B (Apache 2.0). The code is MIT-licensed, see the repository.