Learning When to Think: Dual-Reference Offline Optimization for Adaptive VLM Reasoning
Summary
- Researchers introduce Dual-Reference Offline Optimization for Adaptive VLM Reasoning (DR-BOLT) to balance fast and slow thinking in vision-language models.
- Forcing every input into long reasoning traces causes overthinking and high inference costs, prompting an adaptive approach.
- DR-BOLT distinguishes the desired fast/slow mixture reference from the sampling distribution, correcting distribution mismatches in reward-weighted training.
- Experiments show improved score-length trade-offs on ScienceQA and transfer benchmarks, achieving a 2.90-fold end-to-end training speedup over reproduced GRPO.
Why it matters
Adaptive reasoning techniques like DR-BOLT help reduce unnecessary inference costs in vision-language models while maintaining task performance, offering faster and more efficient multimodal deployments.