What they found
The authors study vision-language models (VLMs, which read images and text) that choose between a fast direct answer and a slow step-by-step trace. Fast answers and slow traces usually come from different sources, so a standard offline training goal weights one mode wrongly. Their method, DUAL-REFERENCE BOLT (DR-BOLT), separates the mixture of fast and slow behavior they want from the distribution the data was sampled from, and corrects the gap during reward-weighted training. In matched Qwen3-VL comparisons with identical candidate banks, DR-BOLT improves score-length trade-offs over the evaluated offline and online baselines on ScienceQA and three held-out transfer benchmarks. It reaches a 24.1% cross-benchmark slow ratio and a 2.90-fold end-to-end training speedup over reproduced GRPO.
Why it matters
Long reasoning on every input wastes compute and can cause overthinking. A model that reasons only when a question is hard could cut inference cost without giving up accuracy. The abstract frames the results as support for the studied two-reference setting only, and says broader robustness remains open.
How it changes the race
If the result holds, it makes adaptive reasoning cheaper to train and run, which helps smaller teams. It looks like a step forward on efficiency rather than a change in raw capability.
The abstract, by the authors
Reasoning traces improve vision-language models on complex multimodal tasks, but forcing every input to follow a long reasoning process leads to overthinking and unnecessary inference cost. We study adaptive fast/slow thinking for VLMs, where one model answers directly for simple inputs and reasons step by step when needed. Fast answers and slow traces often come from distinct behavior policies, making a single-reference offline objective misweight one mode. We propose DUAL-REFERENCE BOLT (DR-BOLT), which distinguishes the desired fast/slow mixture reference from the sampling distribution and corrects their mismatch in reward-weighted training. Fast/slow partition competition explains routing in the optimal target; an efficient source-wise correction approximates the exact objective when response formats are well separated. Matched Qwen3-VL comparisons use identical candidate banks to isolate the objective from slow-teacher quality. DR-BOLT improves score-length trade-offs over the evaluated offline and online baselines on ScienceQA and three held-out transfer benchmarks, with a 24.1% cross-benchmark slow ratio and a 2.90-fold end-to-end training speedup over reproduced GRPO. These results support adaptive reasoning in the studied two-reference setting, while broader robustness remains open. Our code is available at https://github.com/HongbinLin3589/DR-BOLT.
Quoted from arXiv:2611.09154. The research and its words belong to its authors.