Research map / Vision-language and multimodal / Multimodal LLMs

Large Vision-Language Models Get Lost in Attention

ICML 2026 · Vision-language and multimodal › Multimodal LLMs · approach: VLM capability analysis · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes RID and MixIG, information-theoretic/geometric metrics for residual-stream updates, finds attention is subspace-preserving and FFN subspace-expanding in 15 LVLMs, and shows replacing visual attention scores with predefined priors (incl. Gaussian noise) preserves or improves benchmark accuracy.

Problem

Module-level interpretability of LVLM decoders lacks a unified information-theoretic and geometric framework, and it is unclear whether learned visual attention scores are functionally necessary.

Approach

Define RID (effective-rank spectrum change plus subspace innovation) and MixIG (change in token mixing entropy) to decompose residual updates into innovation vs reconfiguration, then test attention replacement with shared attention priors (visual-encoder attention, patch complexity, Gaussian noise).

Key results

• Attention updates show low RID (0.06) and high MixIG (0.61), while FFN updates show higher RID (0.21) and low MixIG (0.02), supporting a functional decoupling across 15 LVLMs.
• Replacing visual attention with a visual-encoder prior on Qwen-2.5-VL-7B raises POPE from 86.54 to 87.62 and MathVista from 63.3 to 63.1 (roughly neutral), with the Noise variant giving comparable results.
• On LLaVA-1.5-7B, the visual-encoder prior improves HallusionBench from 41.63 to 44.29 and VMCBench from 48.71 to 52.23.

Surprising findings

• Replacing learned decoder attention scores with Gaussian noise in selected layers yields comparable or better benchmark accuracy than vanilla models.
• Attention updates show negligible innovation, close to the RoPE calibration level.

Limitations

• Replacement is a proof-of-concept applied to selected layers and heads chosen via ablations on the test benchmarks.
• Mixed results on some benchmarks (e.g., MathVista changes are small and sometimes negative).
• Metric tolerance epsilon is calibrated by RoPE discrepancy, which is a heuristic.

Experimental setup

Datasets: POPE 3DSRBench RealWorldQA MMMU VMC-Bench MathVista HallusionBench

Benchmarks: POPE 3DSRBench RealWorldQA MMMU VMC-Bench MathVista HallusionBench

Metrics: accuracy RID MixIG

Baselines: Vanilla LVLM (unmodified attention) Visual-encoder attention prior Patch complexity prior

Builds on: tuned lens attention rollout manifold hypothesis

Closest papers by meaning

• A Comprehensive Information-Decomposition Analysis of Large Vision-Language Models (ICLR 2026 · similarity 0.93)
• Mitigating Mask Prior Drift and Positional Attention Collapse in Large Diffusion Vision-Language Models (ICML 2026 · similarity 0.92)
• A-VL: Adaptive Attention for Large Vision-Language Models (AAAI 2025 · similarity 0.91)
• Synthesize Diagnose and Optimize: Towards Fine-Grained Vision-Language Understanding (CVPR 2024 · similarity 0.91)
• What's in the Image? A Deep-Dive into the Vision of Vision Language Models (CVPR 2025 · similarity 0.91)
• Lifting the Veil on Visual Information Flow in MLLMs: Unlocking Pathways to Faster Inference (CVPR 2025 · similarity 0.90)