Trace builds a reproducible RLVR environment for 11 visual reasoning domains

Md Tanvirul Alam · hf · 2026-07-23

Trace is a taxonomy-guided environment for multidomain visual reasoning designed to make RLVR usable for vision-language models.

• It factorizes each task into a scene grammar and an executable task program, separating visual rendering from answer computation.
• A shared semantic state generates the image, prompt, typed answer, verifier state, and replayable instance trace, making the environment exactly verifiable and reproducible.
• The release includes 1,000 tasks, 277 scene grammars, and 11 visual domains with controlled semantic and visual variation.
• RLVR trained on 64,000 Trace instances improves the macro-average across 24 external benchmarks by 3.51 points for Qwen2.5-VL-3B and 4.06 points for Qwen2.5-VL-7B.

Project page: https://maveryn.github.io/trace/

Original post →

More from Research

• Oxford study says AI writing tools can subtly steer public opinion — SandraWachter5 · 2026-07-23
• Learnable novelty improves reward and speeds convergence across RL tasks — menhguin · 2026-07-23
• USC benchmark shows GPT-5.5 scores 10.6% on active visual observation tasks — UniversityofSouthernCalifornia · 2026-07-23
• RECAP lets probes verify activation explanations that reconstruction scores can fake — Hiskias Dingeto · 2026-07-23
• Paper proposes Linearizing Softmax Attention into Gated DeltaNet — bronzeagepapi · 2026-07-23
• Overcoming RGB Sim-to-Real Gap: Gaussian Splatting Enables Zero-Shot Robot Transfer — ZeYanjie · 2026-07-23

Research channel →