Research map / Vision-language and multimodal / Multimodal LLMs
LASER: A Corrective Lens for LVLMs via Visual Attention Preservation and Sink Suppression
Bowen Yuan, Zijian Wang, Yadan Luo, Shijie Wang, Zi Helen Huang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add GRPO training (ablation) | Qwen2.5-VL-7B · average over 8 benchmarks | average accuracy: 48.4 -> 51.1 | Improved |
| add visual grounding reward R_vis (ablation) | GRPO on Qwen2.5-VL-7B · average over 8 benchmarks | average accuracy: 51.1 -> 52.2 | Improved |
| add sink suppression reward R_supp (ablation) | GRPO + R_vis on Qwen2.5-VL-7B · General reasoning | accuracy: 55.1 -> 58.6 | Improved |
| add sink suppression reward R_supp (ablation) | GRPO + R_vis on Qwen2.5-VL-7B · average over 8 benchmarks | average accuracy: 52.2 -> 53.7 | Improved |
| add LASER rewards on SFT cold-start model | Revisual-R1 · average over benchmarks | average accuracy: 52.2 -> 58.0 | Improved |
| add LASER rewards on top of GRPO for Revisual-R1 (ablation) | GRPO on Revisual-R1 · average over benchmarks | average accuracy: 56.5 -> 58.0 | Improved |
| one-time suppression of visual value states at 1/8 decoding stage (full masking) (ablation) | Qwen2.5-VL-7B inference · MMStar | accuracy: 62.8 -> 48.9 | Hurt |
| one-time full masking of visual states at 3/4 decoding stage (ablation) | Qwen2.5-VL-7B inference · MMStar | accuracy: 2.0% decrease | Hurt |
| moderate suppression (kappa 0.5) at early stage versus later stage (ablation) | Qwen2.5-VL-7B inference · MMStar | accuracy: 3.7% drop vs 1.3% | Hurt |
| redistribute attention from sink to non-sink tokens instead of uniform scaling (ablation) | Qwen2.5-VL-7B inference · MMStar | accuracy | Improved |
| early-weighted R_vis instead of uniform R_vis (ablation) | LASER · 6-benchmark average | average accuracy: 52.2 vs 52.3 | No clear effect |
| vary lambda in R_vis (ablation) | LASER · average accuracy over benchmarks | average accuracy: 52.4 / 53.3 / 54.6 / 53.9 / 53.2 for lambda 1 / 2.5 / 5 / 7.5 / 10 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- OpenVLThinker-7B — OpenVLThinker: Complex Vision-Language Reasoning via Iterative SFT-RL Cycles (NeurIPS 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Remember Me: Bridging the Long-Range Gap in LVLMs with Three-Step Inference-Only Decay Resilience Strategies (AAAI 2026 · similarity 0.90)
- Dismantling Pathological Shortcuts: A Causal Framework for Faithful LVLM Decoding (ICML 2026 · similarity 0.90)
- To Sink or Not to Sink: Visual Information Pathways in Large Vision-Language Models (ICLR 2026 · similarity 0.90)
- Mitigating Mask Prior Drift and Positional Attention Collapse in Large Diffusion Vision-Language Models (ICML 2026 · similarity 0.89)
- Finding the Correct Visual Evidence Without Forgetting: Mitigating Hallucination in LVLMs via Inter-Layer Visual Attention Discrepancy (ICML 2026 · similarity 0.89)
- Dual-Latent Memory Routing for Vision-Language Reasoning (ICML 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
