Search papers, labs, and topics across Lattice.
This paper introduces Trace, a novel taxonomy-guided environment designed to enhance multidomain visual reasoning in vision-language models through reinforcement learning with verifiable rewards (RLVR). By factorizing task construction into a scene grammar and an executable task program, Trace enables a systematic approach to generating 1,000 tasks across 277 scene grammars and 11 visual domains, ensuring broad and reproducible training data. The application of RLVR on 64,000 instances from Trace results in significant improvements in performance on 24 external benchmarks, demonstrating the effectiveness of this structured training approach in transferring knowledge beyond the generated tasks.
Trace achieves a 4.06 percentage point boost in visual reasoning accuracy for large language models, showcasing the power of structured task generation in RLVR.
Reinforcement learning with verifiable rewards (RLVR) has substantially improved language-model reasoning, yet its extension to vision-language models remains constrained by the lack of training data that are simultaneously broad, exactly verifiable, and reproducible. We introduce Trace, a taxonomy-guided environment for multidomain visual reasoning. Trace factorizes task construction into a scene grammar and an executable task program, separating visual realization from answer computation. A shared semantic state determines the rendered image, prompt, typed answer, verifier state, and replayable instance trace. The resulting environment comprises 1,000 tasks over 277 scene grammars and 11 visual domains, with controlled semantic and visual variation. RLVR on 64,000 Trace instances improves the macro-average across 24 external benchmarks by 3.51 percentage points for Qwen2.5-VL-3B and 4.06 points for Qwen2.5-VL-7B, providing evidence that broad procedural training can transfer beyond the generated task distributions. Project page: https://maveryn.github.io/trace/.
