12 papers
Subliminal Learning as Trait-Direction Drift: A Mechanism and Targeted Control under SFT Distillation
Zhixuan Liu, Zhichen Dong, Yuyu Fan +2
Beyond intended capabilities, model distillation can transfer hidden traits from a teacher. A teacher biased by a system prompt can generate semantically clean training data, such…
Scaling Model-Generated Distillation Data Can Make Latent Teacher Traits More Recoverable
Zhichen Dong, Zhixuan Liu, Yuyu Fan +3
Scaling model-generated data is usually viewed as improving distillation: more examples should increase coverage, reduce noise, and produce stronger students. We show a second effe…
Decoupled Contrastive Decoding via Expert-Aligned Drafting
Zhixuan Liu, Zhichen Dong, Yuanfu Wang +1
Contrastive Decoding (CD) improves generation quality, but its amateur-model pass makes decoding expensive. Accelerating CD with speculative decoding raises a proposal-alignment qu…
Native Reasoning Models: Training Language Models to Reason on Unverifiable Data
Yuanfu Wang, Zhixuan Liu, Xiangtian Li +2
The prevailing paradigm for training large reasoning models--combining Supervised Fine-Tuning (SFT) with Reinforcement Learning with Verifiable Rewards (RLVR)--is fundamentally con…
RePO: Replay-Enhanced Policy Optimization
Siheng Li, Zhanhui Zhou, Wai Lam +2
Reinforcement learning (RL) is vital for optimizing large language models (LLMs). Recent Group Relative Policy Optimization (GRPO) estimates advantages using multiple on-policy out…
SafeCoT: Improving VLM Safety with Minimal Reasoning
Jiachen Ma, Zhanhui Zhou, Chao Yang +1
Ensuring safe and appropriate responses from vision-language models (VLMs) remains a critical challenge, particularly in high-risk or ambiguous scenarios. We introduce SafeCoT, a l…
Mitigating Object Hallucination via Robust Local Perception Search
Zixian Gao, Chao Yang, Zhanhui Zhou +2
Recent advancements in Multimodal Large Language Models (MLLMs) have enabled them to effectively integrate vision and language, addressing a variety of downstream tasks. However, d…
VLMs Can Aggregate Scattered Training Patches
Zhanhui Zhou, Lingjie Chen, Chao Yang +1
One way to mitigate risks in vision-language models (VLMs) is to remove dangerous samples in their training data. However, such data moderation can be easily bypassed when harmful…
Critique-GRPO: Advancing LLM Reasoning with Natural Language and Numerical Feedback
Xiaoying Zhang, Yipeng Zhang, Hao Sun +4
Recent advances in reinforcement learning (RL) using numerical rewards have significantly enhanced the complex reasoning capabilities of large language models (LLMs). However, we i…
Adversarial Preference Learning for Robust LLM Alignment
Yuanfu Wang, Pengyu Wang, Chenyang Xi +13
Modern language models often rely on Reinforcement Learning from Human Feedback (RLHF) to encourage safe behaviors. However, they remain vulnerable to adversarial attacks due to th…
Emergent Response Planning in LLMs
Zhichen Dong, Zhanhui Zhou, Zhixuan Liu +2
In this work, we argue that large language models (LLMs), though trained to predict only the next token, exhibit emergent planning behaviors: $\textbf{their hidden representations…
Towards AI- Law: A Roadmap to Trustworthy AGI
Chao Yang, Chaochao Lu, Yingchun Wang +1
Ensuring Artificial General Intelligence (AGI) reliably avoids harmful behaviors is a critical challenge, especially for systems with high autonomy or in safety-critical domains. D…