Research map / Vision-language and multimodal / Multimodal LLMs
Learning When to Think: Dual-Reference Offline Optimization for Adaptive VLM Reasoning
Hongbin Lin, Sizhe Zou, Juangui Xu, Xinyue Xu, Zhuoli Ouyang, Stanislav Kolchin, Chengwei Qin, Zhongxiang Dai, Yao SHU
Read the paperCheck your idea against 137,217 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.
Studies adaptive fast/slow thinking for vision-language models, where one model answers directly for simple inputs and reasons step by step only when needed, using offline dual-reference optimization.
Problem
Vision-language models forced to produce long reasoning traces for every input overthink and incur unnecessary inference cost; a single model should decide when to think.
Approach
Frames offline adaptive thinking as more than a length-compression problem and proposes dual-reference offline optimization to train a VLM to switch between direct answering and step-by-step reasoning. Only the abstract is available, so finer details are not stated.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add behavior-reference correction kappa (ablation) | DR-BOLT · MathVision/DynaMath/MMMU-Pro | score and slow ratio: slow ratio 0.13% -> 24.1%; MathVision 0.260 -> 0.606 | Improved |
| use exact instead of mode-wise correction (ablation) | DR-BOLT · three transfer benchmarks | score: 0.606 -> 0.623 (MathVision) | No clear effect |
| use dual-reference mixture instead of single-reference slow (ablation) | BOLT · cross-benchmark | slow ratio: 97.4% -> 24.1% | Improved |
| use DR-BOLT instead of mixture-reference DPO (ablation) | DR-BOLT · MathVision | score: 0.568 -> 0.606 | Improved |
| use DR-BOLT instead of mode-conditioned reward-weighted SFT (ablation) | DR-BOLT · MathVision | score: 0.513 -> 0.606 | Improved |
| use DR-BOLT instead of Reward SFT (ablation) | DR-BOLT · MathVision | score: 0.421 -> 0.606 | Improved |
| use offline DR-BOLT instead of online GRPO | DR-BOLT · MathVision | score: 0.463 -> 0.606 | Improved |
| use offline DR-BOLT instead of online GRPO | DR-BOLT · training cost | time (h): 202.4 -> 69.7 | Improved |
| change sampling fast-prob q from 0.7 to 0.3 (ablation) | DR-BOLT · ScienceQA | score: 0.002 change | No clear effect |
| increase length penalty lambda from 0 to 0.10 (ablation) | DR-BOLT · ScienceQA | response length (words): 28.9 -> 15.7 | Improved |
Closest papers by meaning
- More Thought, Less Accuracy? On the Dual Nature of Reasoning in Vision-Language Models (ICLR 2026 · similarity 0.90)
- Semi-off-Policy Reinforcement Learning for Vision-Language Slow-Thinking Reasoning (NeurIPS 2025 · similarity 0.90)
- Empowering Small VLMs to Think with Dynamic Memorization and Exploration (ICLR 2026 · similarity 0.90)
- See Further, Think Deeper: Advancing VLM's Reasoning Ability with Low-level Visual Cues and Reflection (CVPR 2026 · similarity 0.90)
- Critic-V: VLM Critics Help Catch VLM Errors in Multimodal Reasoning (CVPR 2025 · similarity 0.89)
- From Seeing to Thinking: Decoupling Perception and Reasoning Improves Post-Training of Vision-Language Models (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.
