SciReasoner

Multimodal physics problem solving — caption → reason → critic.

🥇 1st Place — ICML 2025 AI for Math Workshop, Track 2: Physics Reasoning with Diagrams and Expressions

---

What it does

A three-stage pipeline that solves multimodal physics problems by describing the figure, deriving an answer, and then critiquing and refining that answer:

Caption (image → text) → Reason (solve) → Critic (review & correct)

Default model: Gemini-3.1-Pro on any OpenAI-compatible endpoint.

Use it

git clone https://github.com/OpenDCAI/SciReasoner.git
cd SciReasoner
pip install -e .

export OPENAI_API_KEY=<your-key>
export OPENAI_BASE_URL=<endpoint>          # optional, OpenAI-compatible proxy

As a CLI:

scireasoner solve --problem "A 2 kg block slides from rest down a 30° incline of length 5 m, μ=√3/10, g=10. Find the speed at the bottom."
scireasoner solve --problem "Find I(t)." --image circuit.png

As a Python library:

from scireasoner import solve
res = solve(problem="...", image="figure.png")
print(res.answer, res.reasoning)

Inside Claude Code (one-click install):

cd plugins/claude-code/scireasoner && bash install.sh

Inside Codex (one-click install):

cd plugins/codex/scireasoner && bash install.sh

Both plugins expose three MCP tools — scireasoner_solve, scireasoner_caption, scireasoner_reason — and auto-trigger a solve-physics-problem skill when the user asks for help with a physics problem.

Reproduce SeePhys Pro 2026

pip install -e ".[batch]"
hf download Kun-Xiang/SeePhysPro --repo-type dataset --local-dir ./data/SeePhysPro

python seephys_pro_codabench/scripts/run_v2.py \
    --run v2_pub830 --split testmini --levels level1 level2 level3 level4 level5 \
    --caption-model gemini-3.1-pro-preview --reason-model gemini-3.1-pro-preview \
    --critic-model gemini-3.1-pro-preview --use-critic --k-samples 1 --workers 50

python seephys_pro_codabench/scripts/audit_fix.py --run output/v2_pub830
# Upload submission_audited.zip to https://www.codabench.org/competitions/16010/

Per-stage caches under output/<run>/cache/ make crash-resume automatic — re-run to continue.

Results

ICML 2025 SeePhys Challenge — 🥇 1st Place.

SeePhys Pro 2026 (Codabench 16010, public testmini) — current best:

#	Submission	Overall	L1	L2	L3	L4	L5
1	gemini baseline	0.7651	0.770	0.810	0.755	0.700	0.933
4	+ L4 verbatim caption	0.7747	0.765	0.810	0.760	0.740	0.933
8	+ L1 few-shot reason	0.7771	0.770	0.800	0.765	0.750	0.933

Full iteration log: seephys_pro_codabench/output/submissions/README.md.

Layout

scireasoner/              Python package (CLI + MCP server, thin shell)
plugins/claude-code/      Claude Code one-click install
plugins/codex/            Codex one-click install
seephys_pro_codabench/    Active 2026 competition workspace (don't touch)
caption.py / answer.py    Original 2025 single-script implementation

The scireasoner/ package imports stages directly from seephys_pro_codabench/scripts/run_v2.py without copying — so any improvement we push during the live competition flows to all downstream users on the next git pull.

Citation

@article{liang2025multimodal,
  title  = {Multimodal Reasoning for Science: Technical Report and 1st Place
            Solution to the ICML 2025 SeePhys Challenge},
  author = {Liang, Hao and Wu, Ruitao and Zeng, Bohan and Niu, Junbo
            and Zhang, Wentao and Dong, Bin},
  journal= {arXiv preprint arXiv:2509.06079},
  year   = {2025}
}

License

GPL-3.0. See LICENSE. Thanks to the ICML 2025 AI for Math Workshop organizers and Codabench / AWS for hosting.