Diagnosing Modality Transfer and Blind-Training Effects in Multimodal RLVR for Physics Reasoning

Paper | Project Page | Workshop | Challenge | Training Set | Benchmark

News

• [2026.05.13] We release the official code for SeePhys Pro.
• [2026.05.13] The arXiv paper and project page are online.
• [2026.05.13] We release the PhysRL training set and SeePhys Pro benchmark on Hugging Face.
• [2026.05.13] SeePhys Pro is included as Challenge 3 at the 3rd AI for Math Workshop at ICML 2026.

Overview

SeePhys Pro is a fine-grained modality-transfer benchmark for multimodal physics reasoning. Each seed problem is converted into four semantically aligned forms that progressively move task-critical information from language into images:

• L1: text-only problem statement.
• L2: physical structure is moved into the image.
• L3: variables and labels must be grounded visually.
• L4: the full problem is rendered as visual input.

This repository provides the code used for benchmark inference, judging, and the RLVR training diagnostics in the paper.

Highlights

• Progressive modality transfer. We introduce SeePhys Pro, a physics reasoning benchmark built on the principle of same physics, different representation. Its four aligned levels decompose model performance into structural transfer, visual variable grounding, full-rendering effects, and representation consistency.
• Frontier-model diagnostic. We evaluate strong closed- and open-weight MLLMs and find that even frontier systems remain fragile when task-critical information moves from text into diagrams, especially when variables and labels must be read from images.
• PhysRL training corpora. We release source-matched, benchmark-disjoint PhysRL training data for reproducing physics RLVR experiments and comparing normal RL with controlled variants across 4B and 7B models.
• Blind-training observation. We use masked-image RL as a negative-control diagnostic and observe that blind training can still improve unmasked validation accuracy. Our analysis suggests that residual language, templates, answer priors, and dataset regularities can drive part of the gain, so accuracy gains alone should not be taken as evidence of stronger visual grounding.

Core Conclusions

• Best current model. Gemini-3.1-Pro has the highest L1-L4 average accuracy among the listed systems, followed by Claude-4.7-Opus and GPT-5.4.
• Primary evaluation bottleneck. Across the aggregate row, accuracy drops from 49.2% on L1 to 35.8% on L4, with the largest mean gap at L2 -> L3 where variables and labels must be grounded from images.
• Training-time warning. PhysRL improves validation accuracy, but normal RL and blind RL do not reliably close modality-transfer gaps; reporting only final-answer accuracy can therefore overstate visually grounded learning.

Main Leaderboard

Rank	Model	Avg.	L1	L2	L3	L4	Cons4	Delta-S	Delta-V	Delta-R	Delta-T
Human	Human Performance	57.0	54.0	58.5	59.5	56.0	49.0	-4.5	-1.0	3.5	-2.0
1 🥇	Gemini-3.1-Pro	69.0	71.0	72.0	66.5	66.5	47.0	-1.0	5.5	0.0	4.5
2 🥈	Claude-4.7-Opus	61.0	74.0	67.0	56.5	46.5	33.5	7.0	10.5	10.0	27.5
3 🥉	GPT-5.4	60.1	67.4	64.1	55.8	53.0	32.6	3.3	8.3	2.8	14.4
4	Qwen-3.6-flash	54.8	61.4	59.3	49.9	48.4	29.9	2.1	9.4	1.5	13.0
5	Qwen3.5-27B	33.4	45.0	34.8	28.0	25.6	9.9	10.3	6.8	2.4	19.4
6	GPT-5	30.4	41.8	32.9	23.8	23.2	8.9	8.9	9.1	0.5	18.5
7	Gemma-4-31B-it	29.6	38.9	33.5	23.9	22.0	8.9	5.4	9.6	1.9	16.9
-	Aggregate Average	42.5	49.2	46.1	38.7	35.8	21.4	3.0	7.4	2.9	13.4

Table 1. Main SeePhys Pro leaderboard. Values follow the current project page. The highlighted Human Performance row is shown first as a reference and is not included in model ranks. Avg. is the mean of L1-L4 accuracy; ranked model rows are sorted by Avg., and the aggregate row is kept at the bottom for reference. L1-L4 and Cons4 are percentages; positive transfer gaps indicate degradation under a more visual representation. Delta-S, Delta-V, and Delta-R measure L1 -> L2 structural transfer, L2 -> L3 variable grounding, and L3 -> L4 rendering gaps; Delta-T is the overall L1 -> L4 gap.

Takeaway. Gemini-3.1-Pro leads by L1-L4 average accuracy, while Claude-4.7-Opus has the strongest L1 score but a much larger overall transfer gap. The aggregate row still drops from 49.2% on L1 to 35.8% on L4, with Delta-V as the largest mean component.

Benchmark Design

Figure 1. Four modality-transfer levels in SeePhys Pro.

Takeaway. The benchmark keeps the underlying physics fixed while moving structure, variables, and finally the full statement into vision. This makes it possible to diagnose where a model loses invariance instead of reporting only a single final score.

Data Pipeline

Figure 2. Data pipeline for constructing SeePhys Pro.

Takeaway. Problems are curated from heterogeneous physics sources, cleaned, deduplicated, transformed, and manually sketched into aligned levels. The result is a controlled modality-transfer benchmark rather than a loose collection of unrelated visual questions.

Training Diagnostic

Figure 3. Normal and blind RL diagnostics on SeePhys Pro.

Takeaway. Normal RL and blind RL both improve absolute accuracy, but the total transfer gap and the variable-grounding gap remain large. Accuracy gains therefore do not automatically mean that the model has learned to use the task-critical visual evidence.

Cross-Benchmark Control

Figure 4. Cross-benchmark normal and blind RL gain summary.

Takeaway. Blind-training gains also appear on external physics and math benchmarks, and in several settings recover a large fraction of normal-RL gains. This points to residual language, templates, answer priors, and dataset-level regularities as plausible non-visual learning routes.

Mechanism Controls

Figure 5. Mechanism controls for blind-training gains.

Takeaway. Text deletion, targeted deletion, mask-rate ablations, and format-saturation controls weaken simple explanations such as all-black-image artifacts or format-only learning. The blind signal is better understood as a distributed shortcut effect from residual textual and distributional cues.

Quick Start

Benchmark Evaluation Environment

Python 3.10-3.12 is recommended.

conda create -n seephyspro-eval python=3.10 -y
conda activate seephyspro-eval

cd lmms-eval
pip install -U pip
pip install -e .

Download Benchmark Data

The SeePhys Pro lmms-eval tasks expect Level 1-4 parquet files under lmms-eval/data/seephys_pro/:

cd lmms-eval
mkdir -p data/seephys_pro data/seephys_pro_raw

huggingface-cli download Kun-Xiang/SeePhysPro \
  --repo-type dataset \
  --local-dir data/seephys_pro_raw

for level in level1 level2 level3 level4; do
  cp "data/seephys_pro_raw/${level}/train-00000-of-00001.parquet" \
     "data/seephys_pro/${level}-00000-of-00001.parquet"
done

Run Benchmark Evaluation

Local open-model evaluation:

cd lmms-eval
python -m lmms_eval eval --config configs/seephys_pro_local.yaml

Closed-model or OpenAI-compatible evaluation:

cd lmms-eval
export OPENAI_API_KEY="<your_api_key>"
python -m lmms_eval eval --config configs/seephys_pro_api.yaml

Common task entry points:

seephys_pro_all
seephys_pro_level1
seephys_pro_level2
seephys_pro_level3
seephys_pro_level4
seephys_pro_testmini

RLVR Training

Use a Linux CUDA environment for training. Install the PyTorch build matching your CUDA driver before installing EasyR1 dependencies, because flash-attn requires torch at build time.

conda create -n seephyspro-rl python=3.10 -y
conda activate seephyspro-rl

cd EasyR1
pip install -U pip

# Replace cu124 if your cluster uses another CUDA version.
pip install torch torchvision --index-url https://download.pytorch.org/whl/cu124

pip install -r requirements.txt --no-build-isolation
pip install -e . --no-build-isolation

Run the retained reproduction launchers:

cd EasyR1

# PhysRL physics training
bash examples/repro/physrl_4b_normal.sh
bash examples/repro/physrl_4b_blind.sh
bash examples/repro/physrl_7b_normal.sh
bash examples/repro/physrl_7b_blind.sh

# Visual-math control
bash examples/repro/math_vn_4b_normal.sh
bash examples/repro/math_vn_4b_blind.sh
bash examples/repro/math_vn_7b_normal.sh
bash examples/repro/math_vn_7b_blind.sh

Blind RL is used as a negative-control diagnostic. It masks only training images:

data.train_image_mask_mode=all_blank

Validation images remain unmasked. Dataset placeholders can be overridden at runtime:

PHYSRL_TRAIN_FILES="<your_physrl_train_split>" \
PHYSICS_VAL_FILES="<name1=validation_split1,name2=validation_split2>" \
MODEL_PATH="Qwen/Qwen3-VL-4B-Instruct" \
GPUS=4 \
NNODES=1 \
bash examples/repro/physrl_4b_normal.sh

The main reproduction configs use 4 rollouts per prompt, rollout temperature 1.0, top-p 1.0, 4096 prompt tokens, 4096 response tokens, AdamW learning rate 1e-6, weight decay 0.01, one PPO epoch, BF16 FSDP, vLLM rollout, and validation every 5 iterations.

Notes

• API credentials should be provided through environment variables; do not hard-code keys in configs or scripts.
• Some training configs use sanitized dataset placeholders. Override them with the PhysRL and validation splits used in your environment.
• For SeePhys Pro, report A1-A4, transfer gaps, and consistency. Final accuracy alone is not sufficient for diagnosing visual grounding.
• This release excludes private keys, .env files, local user paths, .git directories, temporary scripts, cache files, and OS metadata.

Citation

@article{xiang2026seephyspro,
  title={SeePhys Pro: Diagnosing Modality Transfer and Blind-Training Effects in Multimodal RLVR for Physics Reasoning},
  author={Xiang, Kun and Zhang, Terry Jingchen and Liu, Zirong and Zhou, Bokai and Tang, Yueling and Yu, Junjie and Lu, Jiacong and Huang, Shangrui and Li, Heng and Zhang, Likui and Liu, Kunkun and Zhang, Changzheng and Fang, Yangle and Guo, Boqiang and Zhen, Hui-Ling and Tu, Dandan and Huang, Yinya and Liang, Xiaodan},
  journal={arXiv preprint arXiv:2605.09266},
  year={2026}
}

@article{xiang2026seephys,
  title={Seephys: Does seeing help thinking?--benchmarking vision-based physics reasoning},
  author={Xiang, Kun and Li, Heng and Zhang, Terry Jingchen and Huang, Yinya and Liu, Zirong and Qu, Peixin and He, Jixi and Chen, Jiaqi and Yuan, Yu-Jie and Han, Jianhua and others},
  journal={Advances in Neural Information Processing Systems},
  volume={38},
  year={2026}
}

Acknowledgements

This release builds on lmms-eval, EasyR1, and veRL.