Attention Reallocation: Towards Zero-cost and Controllable Hallucination Mitigation of MLLMs
This addresses hallucinations in MLLMs for users needing efficient and controllable mitigation, though it is incremental as it builds on existing training-free methods.
The paper tackles the problem of hallucinations in Multi-Modal Large Language Models (MLLMs) by proposing attention reallocation (AttnReal), a method that mitigates hallucinations with nearly zero extra inference cost and allows for a trade-off between response faithfulness and overall performance, validated across six MLLMs and three decoding strategies.
Multi-Modal Large Language Models (MLLMs) stand out in various tasks but still struggle with hallucinations. While recent training-free mitigation methods mostly introduce additional inference overhead via retrospection strategy and contrastive decoding, we propose attention reallocation (AttnReal) to mitigate hallucinations with nearly zero extra cost. Our approach is motivated by the key observations that, MLLM's unreasonable attention distribution causes features to be dominated by historical output tokens, which further contributes to hallucinated responses because of the distribution gap between different token types. Based on the observations, AttnReal recycles excessive attention from output tokens and reallocates it to visual tokens, which reduces MLLM's reliance on language priors and ensures the decoding process depends more on the visual inputs. More interestingly, we find that, by controlling the intensity of AttnReal, we can achieve a wide-range trade-off between the response faithfulness and overall performance. Comprehensive results from different benchmarks validate the effectiveness of AttnReal across six open-source MLLMs and three decoding strategies.