Skip to content

Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning

🕒 Published (v1): 2026-07-22 06:17 UTC · Source: Arxiv · link

Why this paper was selected

RLVR training environment for VLMs with taxonomy-guided verifiable visual rewards

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

TRACE is a procedural training environment for vision-language models that decomposes visual task construction into reusable scene grammars and executable task programs, enabling exact verifiable rewards across 11 visual domains. Training Qwen2.5-VL models on 64K TRACE instances with RLVR improves macro-average accuracy across 24 external benchmarks by 3.51 pp (3B) and 4.06 pp (7B), demonstrating broad transfer from procedurally generated visual tasks.

Problem

RLVR has proven effective for language-model reasoning but its extension to VLMs is bottlenecked by training data that lacks all three of: broad domain coverage, exact verifiability, and reproducibility. Existing procedural visual environments focus on a single scene grammar or narrow task family, while large dataset mixtures inherit heterogeneous task granularity and cannot define stable sampling units for cross-domain analysis.

Method

TRACE factorizes task construction into a three-level hierarchy: domain → scene grammar → task. A task is defined as \(T = (S, P, Y, C)\), where \(S\) is the scene grammar, \(P\) is the executable task program, \(Y\) is the answer schema, and \(C\) is the reward contract. A shared semantic state \(x\) is generated by \(x = G_s(z, \theta_s)\); the task program then computes both the typed answer and verifier state \((y, v) = P_t(x; q)\); the renderer and prompt function independently realize the image and question from the same \(x\); and the reward contract is bound to produce an instance scorer. This separation means visual realization (palette, layout, typography) can vary without altering the task computation or answer. Bounded query variation (e.g., largest vs. smallest IQR) changes a prompt argument without spawning a new task, preventing inventory inflation. RLVR training uses GRPO with reward \(R = 0.95R_a + 0.05R_f\), where \(R_a\) is exact-match answer correctness and \(R_f\) is format compliance. Both Qwen2.5-VL-3B and 7B are trained on 64,000 instances (64 per task) for 500 GRPO updates on 8×H100 GPUs.

Key Contributions

  • A program-centered taxonomy formalizing task identity via equivalence relations on \((S, P, Y, C)\), yielding 1,000 tasks over 277 scene grammars across 11 visual domains.
  • A deterministic generator–renderer–verifier pipeline where images, prompts, typed answers, and verifier states all derive from a single shared semantic state, enabling exact supervision and replayable instance traces.
  • Controlled two-level instance variation: semantic parameters change the answer-relevant state; render-only parameters change visual presentation while preserving the typed answer.
  • A two-scale RLVR study (3B and 7B) demonstrating transfer to 24 external benchmarks not present in training.

Results

  • Qwen2.5-VL-3B: +3.51 pp macro-average across 24 external benchmarks; positive mean change on 21/24 benchmarks.
  • Qwen2.5-VL-7B: +4.06 pp macro-average across 24 external benchmarks; positive mean change on all 24 benchmarks.
  • Training set: 64K instances (64 per task × 1,000 tasks); held-out TRACE validation: 2,000 instances (2 per task).
  • External evaluation covers 32,805 examples per model across six groups: charts/tables, visual mathematics, science/general reasoning, spatial reasoning, perception/counting, puzzles/logic.
  • Baselines are the corresponding Qwen2.5-VL-3B-Instruct and Qwen2.5-VL-7B-Instruct base checkpoints without TRACE fine-tuning.

Limitations

  • The 11 domains and 277 scene grammars are authored manually; scalability to new domains requires engineering new renderers and task programs.
  • Automatic validation checks structural consistency but cannot guarantee semantic clarity or uniform visual legibility for every rendered instance; human inspection is used as a complement but may not be exhaustive.
  • Training uses a single fixed set of 64K instances (one shuffled pass); curriculum, difficulty filtering, or dynamic sampling strategies are not explored.
  • Transfer gains are measured via macro-average, which treats all 24 benchmarks equally regardless of difficulty or relevance; per-benchmark variance is not fully reported in the excerpt.
  • The paper does not compare against state-of-the-art multimodal RL mixture systems (e.g., Vero at 600K examples) under controlled conditions, so the efficiency/scale tradeoff is not quantified.

Relevance to Vision-Language Models

TRACE directly addresses the core bottleneck in applying RLVR to VLMs: the absence of training data that is simultaneously broad, exactly verifiable, and reproducible. The program-centered taxonomy provides a principled framework for constructing VLM training curricula that decouple visual rendering from reasoning computation, a design principle applicable to future benchmark and training-data engineering efforts. The demonstrated cross-domain transfer from purely procedural synthetic tasks to 24 real-world benchmarks strengthens the case that VLMs can acquire generalizable visual reasoning skills without task-specific human annotation. For researchers tracking VLMs, TRACE establishes a reproducible RLVR training protocol and a controlled diagnostic environment for isolating which visual reasoning competencies benefit most from procedural training.