Prefill-Time Intervention for Mitigating Hallucination in Large Vision-Language Models

2026-04-28

AI Summary

This work addresses the susceptibility of large vision-language models to hallucination during generation, which undermines output reliability. The authors propose a modality-aware guidance paradigm applied exclusively during the prefill stage, decoupling control over key-value (KV) cache: keys are steered to align with visually grounded objects while values are filtered to suppress background noise, thereby correcting hallucination-prone representations before error propagation occurs. This approach introduces, for the first time, an intervention mechanism that independently manipulates visual and textual modalities during prefill, operating orthogonally to existing decoding-stage methods. Experiments demonstrate that the method consistently reduces hallucination rates across diverse models, decoding strategies, and benchmarks, exhibiting strong generalization and plug-and-play compatibility.

Abstract

Large Vision-Language Models (LVLMs) have achieved remarkable progress in visual-textual understanding, yet their reliability is critically undermined by hallucinations, i.e., the generation of factually incorrect or inconsistent responses. While recent studies using steering vectors demonstrated promise in reducing hallucinations, a notable challenge remains: they inadvertently amplify the severity of residual hallucinations. We attribute this to their exclusive focus on the decoding stage, where errors accumulate autoregressively and progressively worsen subsequent hallucinatory outputs. To address this, we propose Prefill-Time Intervention (PTI), a novel steering paradigm that intervenes only once during the prefill stage, enhancing the initial Key-Value (KV) cache before error accumulation occurs. Specifically, PTI is modality-aware, deriving distinct directions for visual and textual representations. This intervention is decoupled to steer keys toward visually-grounded objects and values to filter background noise, correcting hallucination-prone representations at their source. Extensive experiments demonstrate PTI's significant performance in mitigating hallucinations and its generalizability across diverse decoding strategies, LVLMs, and benchmarks. Moreover, PTI is orthogonal to existing decoding-stage methods, enabling plug-and-play integration and further boosting performance. Code is available at: https://github.com/huaiyi66/PTI.

Problem

• hallucination
• Large Vision-Language Models
• reliability
• factually incorrect responses
• visual-textual understanding

Innovation

• Prefill-Time Intervention
• Hallucination Mitigation
• Vision-Language Models
• KV Cache Steering
• Modality-Aware Intervention