Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning

Analysis

Why This Paper Matters

This paper addresses a critical bottleneck in extending reinforcement learning with verifiable rewards (RLVR) to vision-language models (VLMs): the lack of training data that is simultaneously broad, exactly verifiable, and reproducible. Existing datasets for visual reasoning are often static, limited in domain coverage, or lack the procedural structure needed for RLVR. Trace introduces a taxonomy-guided environment that systematically generates diverse, verifiable tasks across 11 visual domains, enabling RLVR training that demonstrably transfers to external benchmarks. This is significant because it provides a principled way to scale up training data for VLMs without manual annotation, potentially unlocking the same reasoning improvements that RLVR has brought to language models.

Technical Contributions

• Taxonomy-guided task construction: Trace factorizes tasks into a scene grammar (defining visual elements and their relationships) and an executable task program (computing the answer), separating visual realization from answer computation. This allows controlled variation in both visual and semantic dimensions.
• Shared semantic state: A unified state determines the rendered image, prompt, typed answer, verifier state, and replayable instance trace, ensuring exact verifiability and reproducibility.
• Large-scale environment: 1,000 tasks over 277 scene grammars and 11 visual domains, with 64,000 instances used for RLVR training.
• Transferable training: RLVR on Trace instances improves macro-average across 24 external benchmarks by 3.51 pp (Qwen2.5-VL-3B) and 4.06 pp (Qwen2.5-VL-7B), demonstrating that procedural training generalizes beyond the generated distribution.

Results

• RLVR on 64,000 Trace instances yields a 3.51 percentage point improvement in macro-average accuracy across 24 external benchmarks for Qwen2.5-VL-3B.
• For Qwen2.5-VL-7B, the improvement is 4.06 percentage points.
• These gains indicate that the procedural training data from Trace transfers effectively to diverse real-world visual reasoning tasks, validating the approach.

Significance

Trace provides a scalable, principled framework for generating verifiable training data for vision-language reasoning, addressing a key data bottleneck in multimodal AI. By enabling RLVR training that transfers to external benchmarks, it opens the door to systematic improvement of VLMs without reliance on static, manually annotated datasets. This could accelerate progress in applications requiring robust visual reasoning, such as autonomous driving, medical imaging, and robotics. The taxonomy-guided approach also offers a template for building similar environments in other domains, potentially advancing the broader field of procedural data generation for AI.