Enhancing the Outcome Reward-based RL Training of MLLMs with Self-Consistency Sampling
Paper • 2511.10648 • Published • 1
Datasets:
Dataset Viewer
The dataset viewer is not available for this dataset.
The JWT signature verification failed. Check the signing key and the algorithm.
Error code: JWTInvalidSignature
Exception: InvalidSignatureError
Message: Signature verification failed
Traceback: Traceback (most recent call last):
File "/src/libs/libapi/src/libapi/jwt_token.py", line 286, in validate_jwt
decoded = jwt.decode(
jwt=token,
...<2 lines>...
options=options,
)
File "/usr/local/lib/python3.14/site-packages/jwt/api_jwt.py", line 368, in decode
decoded = self.decode_complete(
jwt,
...<8 lines>...
leeway=leeway,
)
File "/usr/local/lib/python3.14/site-packages/jwt/api_jwt.py", line 265, in decode_complete
decoded = self._jws.decode_complete(
jwt,
...<3 lines>...
detached_payload=detached_payload,
)
File "/usr/local/lib/python3.14/site-packages/jwt/api_jws.py", line 270, in decode_complete
self._verify_signature(
~~~~~~~~~~~~~~~~~~~~~~^
signing_input,
^^^^^^^^^^^^^^
...<4 lines>...
options=merged_options,
^^^^^^^^^^^^^^^^^^^^^^^
)
^
File "/usr/local/lib/python3.14/site-packages/jwt/api_jws.py", line 417, in _verify_signature
raise InvalidSignatureError("Signature verification failed")
jwt.exceptions.InvalidSignatureError: Signature verification failedNeed help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.
[NeurIPS 2025] Enhancing the Outcome Reward-based RL Training of MLLMs with Self-Consistency Sampling
A simple, general sampling method for RLVR with multi-choice dataset to solve unfaithful reasoning phenomenon!
SCS Resouces
🔔News
- 🔥[2025-11-9] Release the eval codes! 🚀
- 🔥[2025-10-13] Release the dataset the codes! 🚀
- 🔥[2025-9-17] Our SCS paper is accepted by NeurIPS 2025! 🚀
To-do
- Release the eval codes
📖 Introduction
Self‑Consistency Sampling (SCS) improves outcome‑reward reinforcement learning for multimodal large language models (MLLMs). In multiple‑choice reasoning tasks, models often get the correct answer through faulty reasoning and receive unmerited rewards. SCS mitigates this by introducing visual perturbations and repeated resampling of reasoning trajectories, rewarding only consistent reasoning paths. Integrated into methods like RLOO, GRPO, and REINFORCE++, SCS boosts accuracy by up to 7.7% on six multimodal benchmarks with minimal extra cost, and generalizes across models including Qwen2.5‑VL and InternVL3.
Training
Please refer to code repo for more details.
Evaluation
Please refer to code repo for more details.
Contact
- Jiahao Wang: wjhwdscience@stu.xjtu.edu.cn
- Weiye Xu: ustcxwy0271@mail.ustc.edu.cn
Citation
BibTeX:
@article{wang2025enhancing,
title={Enhancing the Outcome Reward-based RL Training of MLLMs with Self-Consistency Sampling},
author={Wang, Jiahao and Xu, Weiye and Yang, Aijun and Zhou, Wengang and Lu, Lewei and Li, Houqiang and Wang, Xiaohua and Zhu, Jinguo},
journal={arXiv preprint arXiv:2511.10648},
year={2025}
}
- Downloads last month
- 70