| title: TRM Complex Reasoning Reproduction | |
| colorFrom: blue | |
| colorTo: green | |
| sdk: gradio | |
| sdk_version: "6.20.0" | |
| app_file: app.py | |
| pinned: false | |
| tags: | |
| - icml2026-repro | |
| - paper-IMFgiWw4jd | |
| # TRM Complex Reasoning Reproduction | |
| This submission checks released artifacts for "Characterizing, Evaluating, and | |
| Optimizing Complex Reasoning" (`IMFgiWw4jd`). | |
| It targets three claims: | |
| - ME2 characterizes reasoning traces along macro/micro and | |
| efficiency/effectiveness dimensions. | |
| - Reasoning traces are represented as DAGs with progression, branching, and | |
| merging structures. | |
| - TRM is trained from the TRM-Preference dataset with a preference-loss reward | |
| model path. | |
| The evidence is CPU-only. It pins the arXiv source hash, GitHub revision, HF | |
| dataset revision, and HF model revision; it avoids downloading the 2 GB | |
| training JSON and 30 GB model weights. | |
| Run: | |
| ```bash | |
| uv run python generate_evidence.py | |
| uv run pytest -q | |
| ``` | |