Pith. sign in

REVIEW 5 cited by

IKOD: Mitigating Visual Attention Degradation in Large Vision-Language Models

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2508.03469 v1 pith:65B75BXJ submitted 2025-08-05 cs.CV

IKOD: Mitigating Visual Attention Degradation in Large Vision-Language Models

classification cs.CV
keywords attentionhallucinationslvlmsikodmodelscollaborativedecodingmitigating
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
read the original abstract

Recent advancements in Large Vision-Language Models (LVLMs) have demonstrated significant progress across multiple domains. However, these models still face the inherent challenge of integrating vision and language for collaborative inference, which often leads to "hallucinations", outputs that are not grounded in the corresponding images. Many efforts have been made to address these issues, but each comes with its own limitations, such as high computational cost or expensive dataset annotation. Recent research shows that LVLMs exhibit a long-term bias where hallucinations increase as the sequence length grows, yet the underlying cause remains poorly understood. Building on extensive research into attention mechanisms in LVLMs, we analyze the relationship between this long-term bias and visual attention. In our research, we identify a consistent phenomenon in current LVLMs: the model's attention to visual input diminishes as the generated sequence grows, which we hypothesize to be a key factor contributing to observed increasing hallucinations. Based on these insights, we propose Image attention-guided Key-value merging cOllaborative Decoding (IKOD), a collaborative decoding strategy generating more image-focused sequences. This method derives logits from shorter sequences with higher image attention through key-value merging and combines them with those from the original decoding, effectively mitigating attention degradation and suppressing hallucinations while not incurring too much inference cost. Extensive experiments on both hallucination and comprehensive benchmarks demonstrate IKOD's superior effectiveness in mitigating hallucinations and improving comprehensive capacities for LVLMs. Importantly, IKOD requires no additional training or external tools, making it a lightweight and efficient framework applicable to various models.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 5 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. C-PTQ: Fisher-weighted Channel-wise Sensitivity for Post-training Quantization of MLLMs

    cs.CV 2026-07 conditional novelty 6.0

    C-PTQ weights quantization error by per-channel Fisher information of the task loss, improving low-bit accuracy of multimodal LLMs by small margins over existing channel-wise scaling methods.

  2. Visual Access Boundaries in Vision-Language Model Reasoning

    cs.AI 2026-07 conditional novelty 6.0

    CoT in VLMs extends language-side computation over early image-derived states rather than prolonging direct image-token access, with gains limited by perceptual readout reliability.

  3. Visual Access Boundaries in Vision-Language Model Reasoning

    cs.AI 2026-07 conditional novelty 6.0

    Chain-of-thought in VLMs does not require prolonged direct image-token attention; the visual-access boundary stays roughly fixed while generation grows, and CoT gains are gated by perceptual readout.

  4. Attention Degradation, Function Token Anchoring, and the Limits of Attention-Based Intervention in Large Language Models

    cs.AI 2026-07 conditional novelty 6.0

    In GPT-2, OPT-1.3B, and LLaMA-3.2 models, boosting attention to function-token positions by 16-24% does not improve contextual retrieval, so short-range attention degradation is descriptive rather than causal.

  5. StochasT: Learning with Stochastic Turn Depth for Visual Instruction Tuning

    cs.CV 2026-07 unverdicted novelty 6.0

    StochasT uses stochastic clustering of language tasks into varying turn depths for the same image to improve LVLMs on both single-turn and multi-turn scenarios without discarding data.