Not All Tokens and Heads Are Equally Important: Dual-Level Attention Intervention for Hallucination Mitigation

Large vision-language models (LVLMs) have shown remarkable capabilities across a wide range of multimodal tasks. However, they remain prone to visual hallucination (VH), often producing confident but incorrect descriptions of visual content. We present VisFlow, an efficient and training-free framework designed to mitigate VH by directly manipulating attention patterns during inference. Through systematic analysis, we identify three key pathological attention behaviors in LVLMs: (1) weak visual grounding, where attention to visual tokens is insufficient or misallocated, over-focusing on uninformative regions; (2) language prior dominance, where excessive attention to prior response tokens reinforces autoregressive patterns and impairs multimodal alignment; (3) prompt redundancy, where many attention heads fixate on system prompt tokens, disrupting the integration of image, instruction, and response content. To address these issues, we introduce two inference-time interventions: token-level attention intervention (TAI), which enhances focus on salient visual content, and head-level attention intervention (HAI), which suppresses over-attention to prompt and nearby text tokens. VisFlow operates without additional training or model modifications. Extensive experiments across models and benchmarks show that VisFlow effectively reduces hallucinations and improves visual factuality, with negligible computational cost.

Learning What Reinforcement Learning Can't: Interleaved Online Fine-Tuning for Hardest Questions

Recent advances in large language model (LLM) reasoning have shown that sophisticated behaviors such as planning and self-reflection can emerge through reinforcement learning (RL). However, despite these successes, RL in its current form remains insufficient to induce capabilities that exceed the limitations of the base model, as it is primarily optimized based on existing knowledge of the model rather than facilitating the acquisition of new information. To address this limitation, we employ supervised fine-tuning (SFT) to learn what RL cannot, which enables the incorporation of new knowledge and reasoning patterns by leveraging high-quality demonstration data. We analyze the training dynamics of RL and SFT for LLM reasoning and find that RL excels at maintaining and improving performance on questions within the model's original capabilities, while SFT is more effective at enabling progress on questions beyond the current scope of the model. Motivated by the complementary strengths of RL and SFT, we introduce a novel training approach, \textbf{ReLIFT} (\textbf{Re}inforcement \textbf{L}earning \textbf{I}nterleaved with Online \textbf{F}ine-\textbf{T}uning). In ReLIFT, the model is primarily trained using RL, but when it encounters challenging questions, high-quality solutions are collected for fine-tuning, and the training process alternates between RL and fine-tuning to enhance the model's reasoning abilities. ReLIFT achieves an average improvement of over +5.2 points across five competition-level benchmarks and one out-of-distribution benchmark compared to other zero-RL models. Furthermore, we demonstrate that ReLIFT outperforms both RL and SFT while using only 13\% of the detailed demonstration data, highlighting its scalability. These results provide compelling evidence that ReLIFT overcomes the fundamental limitations of RL and underscores the significant potential.