| --- |
| license: other |
| tags: [reasoning, chain-of-thought, early-stopping, uncertainty, qwen3] |
| --- |
| |
| # ICLR 2-action controller — rollout chains |
|
|
| Chain-of-thought rollouts with **counterfactual stopping probes**, for a single-chain |
| `{continue, terminate}` controller on Qwen3-1.7B (native thinking). |
|
|
| Each JSONL line is one chain: |
|
|
| | field | meaning | |
| |---|---| |
| | `qkey`, `rep`, `gold` | question id, chain index, gold answer | |
| | `outcome` | `natural` \| `length` (budget cut) | |
| | `answer`, `correct` | the chain's own answer and whether it is right | |
| | `blocks`, `thinking_tokens` | length of the reasoning | |
| | `completion_ids` | the full token ids — the CoT itself | |
| | `block_ends`, `query_points` | block boundaries and the probed prefixes | |
| | `probes[]` | `{block, prefix_len, forced, forced_tokens}` — the answer the model gives if forced to stop **there** | |
|
|
| Probes are placed at every 4th block. Boxed pools force `</think>\n\n\boxed{` (≤ 32 |
| tokens); code pools force a ```` ```python ```` program (≤ 768 tokens) graded by unit |
| tests. That makes the value of stopping at any point observable, which is what the |
| controller is trained and evaluated on. |
|
|
| ## Contents |
|
|
| | file | questions × chains | domain | |
| |---|---|---| |
| | `chains/math_union_fit.jsonl` | 4 241 × 10 | hard math (DAPO, Numina-hard, DeepScaleR) — training | |
| | `chains/math_union_select.jsonl` | 1 065 × 10 | same — model selection | |
| | `chains/math_t1_report.jsonl`, `chains/math_t2_report.jsonl` | 841 / 853 × 10 | sealed math test splits | |
| | `chains/math_ood_report.jsonl` | 951 × 8 | math OOD (MATH-500, AIME 24/25, OlympiadBench, GaoKao) | |
| | `chains/code_pool_{fit,select,report}.jsonl` | 891 / 223 / 386 × 8–16 | SYNTHETIC-2-RL code pool | |
| | `chains/code_ood_pooled.jsonl` | 958 × 8 | HumanEval+, MBPP+, LiveCodeBench ≥ 2024-08 | |
| | `chains/gpqa_diamond_report.jsonl` | 198 × 8 | GPQA-Diamond as numbered-option MCQ | |
| | `chains/mixed_{fit,select}.jsonl` | concatenations of the math and code splits | |
| | `manifests/`, `fire_tables/` | question manifests and cross-fitted π₀ tables | |
| | `ladders/ladders_json.tar.gz` | full decision-rule sweeps behind every reported table | |
| | `ASSET_INVENTORY.json` | path → bytes → lines → **sha256** for every file here | |
|
|
| Companion checkpoints: [`JingweiNi/iclr-2action-controller-heads`](https://huggingface.co/JingweiNi/iclr-2action-controller-heads). |
|
|
| ## Provenance and reproduction |
|
|
| Code, frozen question manifests, all 70 result tables and the full reproducibility |
| record live on the branch `archive/iclr-2action-controller-2026-08-30` of the project |
| repo; see `docs/reproducibility/iclr-2action-controller-assets.md` there. Generation is |
| seeded by `(qkey, rep, seed)`, so any shard regenerates byte-identically from |
| `(manifest, commit, seed)`. |
|
|
| GPQA-Diamond questions are derived from the gated `Idavidrein/gpqa` (revision |
| `633f5ee8…`); this repo redistributes only the model's own rollouts over them, not the |
| source CSV. |
|
|
