Research map / Vision-language and multimodal / Multimodal LLMs
Starve to Perceive: Taming Lazy Perception in VLMs with Constrained Visual Bandwidth
Yuhuan Wu, Haozhe Wang, Cong Wei, Chong Peng, Fangzhen Lin, Wenhu Chen
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 |
|---|---|---|---|
| train with constrained visual bandwidth (budget-aware SFT + RL) vs DeepEyes | Qwen2.5-VL-7B · Overall avg, unconstrained | overall score: 61.77 -> 64.59 | Improved |
| train with constrained visual bandwidth vs DeepEyes, evaluated under B=256 | Qwen2.5-VL-7B · Overall avg, constrained B=256 | overall score: 45.57 -> 52.35 | Improved |
| use budget-aware SFT instead of standard SFT on identical trajectories (ablation) | Qwen2.5-VL-7B SFT · HR-Bench 8K | accuracy: 59.88 -> 65.88 | Improved |
| use budget-aware SFT instead of standard SFT (ablation) | Qwen2.5-VL-7B SFT · V* Overall | accuracy: 75.92 -> 80.62 | Improved |
| use budget-aware SFT instead of standard SFT on V* relative position (ablation) | Qwen2.5-VL-7B SFT · V* rela_pos | accuracy: 73.68 -> 73.68 | No clear effect |
| skip budget-aware SFT warm start before budget RL (ZeroShot + BudgetRL) (ablation) | Qwen2.5-VL-7B RL · HR visual search benchmarks | accuracy: lowest across all benchmarks (no numbers in text) | Hurt |
| apply visual bandwidth constraint during RL instead of unconstrained RL (ablation) | BA-SFT initialized RL · All Focus Ratio | ratio: 0.40 -> 0.50 | Improved |
| apply bandwidth constraint during RL (ablation) | BA-SFT+RL · Unconstrained overall avg | overall score: 64.48 -> 64.59 | No clear effect |
| apply bandwidth constraint during RL (ablation) | BA-SFT+RL · Constrained B=256 overall avg | overall score: 51.54 -> 52.35 | Improved |
| add RL on top of budget-aware SFT (ablation) | BA-SFT · Unconstrained overall avg | overall score: 58.95 -> 64.59 | Improved |
| bandwidth-constrained RL vs vanilla RL across inference budgets 128-1024 (ablation) | Ours · Budget sweep avg | average accuracy: +1.38 (B=128), +0.81, +0.50, +0.60 (B=1024) | Improved |
| constrained inference budget (B=256) vs unconstrained | Ours · Inference latency | speedup: 2.7-5.1x speedup | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- DeepEyes — DeepEyes: Incentivizing "Thinking with Images" via Reinforcement Learning (ICLR 2026) · better on 7 of 7 shared results
- PixelReasoner — Pixel Reasoner: Incentivizing Pixel Space Reasoning via Curiosity-Driven Reinforcement Learning (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
- The Perceptual Bandwidth Bottleneck in Vision-Language Models: Active Visual Reasoning via Sequential Experimental Design (ICML 2026 · similarity 0.89)
- In-Context Learning Can Help Vision Language Models Overcome Training Prior (NeurIPS 2026 · similarity 0.89)
- SPARC: Separating Perception And Reasoning Circuits for Test-time Scaling of VLMs (ICML 2026 · similarity 0.89)
- From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (ICML 2026 · similarity 0.88)
- Learning to See Before Seeing: Demystifying LLM Visual Priors from Language Pre-training (ICLR 2026 · similarity 0.88)
- Prismatic VLMs: Investigating the Design Space of Visually-Conditioned Language Models (ICML 2024 · similarity 0.88)
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.
