Research map / Vision-language and multimodal / Multimodal LLMs

VideoAnchor: Reinforcing Subspace-Structured Visual Cues for Coherent Visual-Spatial Reasoning

ICLR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Training-free token reduction for MLLMs · 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 VideoAnchor, a training-free test-time module that uses sparse subspace clustering self-expression scores to scale query, key and value attention for visual tokens in MLLMs, improving visual-spatial reasoning on VSI-Bench, All-Angles-Bench and Video-MME.

Problem

MLLMs underperform on visual-spatial reasoning across frames or views because text tokens dominate attention and visual cues are not consistently preserved across frames.

Approach

Connects the self-expressiveness property of sparse subspace clustering to Transformer attention; computes per-token sharing expression scores (subspace cardinality times self-expression row-sum) and uses them as multiplicative gates on query-key logits and value vectors at inference, with no retraining.

Key results

• VideoAnchor raises InternVL2-8B (8 frames) on VSI-Bench from 34.6 to 37.8 (+3.2).
• On Video-MME, Qwen2.5VL-72B improves from 75.4 to 80.0 on the spatial-related subset average (+4.6).
• On All-Angles-Bench, LLaVA-Video-72B improves from 49.9 to 53.0 (+3.1).

Surprising findings

• Applying the Q/K scaling before softmax collapses performance (VSI-Bench 5.2 vs 34.6 baseline), while applying it after softmax gives +3.2.
• Video-R1 (GRPO-trained) beats VideoAnchor at 16 frames (34.6 vs 33.3) on VSI-Bench, though VideoAnchor is better at 32 frames.

Limitations

• Scaling coefficients (alpha_Q, alpha_K, alpha_V) are set manually per model, with different values per backbone.
• Gains are modest on some backbones (e.g., +0.3 to +1.2 on Video-MME general tasks) and Video-R1 outperforms it at 16 frames on VSI-Bench.
• Sensitive to SSC hyperparameters such as the number of subspaces (fixed at 24 for experiments).

Experimental setup

Datasets: VSI-Bench All-Angles-Bench Video-MME

Benchmarks: VSI-Bench All-Angles-Bench Video-MME

Metrics: accuracy average score

Baselines: InternVL2-8B Qwen2.5-VL-7B LLaVA-Video-72B Video-R1-7B k-Means attention scaling UniformBoost

Builds on: sparse subspace clustering ADMM spectral clustering Video-R1

Closest papers by meaning

• Why Is Spatial Reasoning Hard for VLMs? An Attention Mechanism Perspective on Focus Areas (ICML 2025 · similarity 0.91)
• Fill the GAP: A Granular Alignment Paradigm for Visual Reasoning in Multimodal Large Language Models (NeurIPS 2026 · similarity 0.90)
• SIFThinker: Spatially-Aware Image Focus for Visual Reasoning (AAAI 2026 · similarity 0.90)
• Visual Structures Help Visual Reasoning: Addressing the Binding Problem in LVLMs (NeurIPS 2025 · similarity 0.90)
• Multi-step Visual Reasoning with Visual Tokens Scaling and Verification (NeurIPS 2025 · similarity 0.89)
• Spatial-MLLM: Boosting MLLM Capabilities in Visual-based Spatial Intelligence (NeurIPS 2025 · similarity 0.89)