Skip to content

Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning

Trace is a procedural environment for broad, exactly verifiable visual reasoning. It separates scene grammar, executable task program, and bounded query variation so that visual realization and reasoning structure can vary independently while preserving exact supervision. Every instance includes a rendered image, prompt, typed answer, verifier state, and replayable execution trace.

1,000 reasoning tasks
277 scene grammars
11 visual domains
66,000 generated examples
Trace examples from charts, games, geometry, graphs, icons, illustrations, pages, physics, puzzles, symbolic reasoning, and 3D scenes, followed by the Trace mark.
One example from each Trace domain: charts, games, geometry, graphs, icons, illustrations, pages, physics, puzzles, symbolic reasoning, and 3D scenes.

How Trace works

Trace organizes visual reasoning as domain → scene grammar → task program. A deterministic seed instantiates semantic scene state, and the selected task program executes over that state to derive a unique typed answer and verifier state. The image and prompt are then rendered from the same underlying state.

A four-stage Trace instance pipeline showing semantic scene state, task execution, validation, and the final RLVR record.
Semantic state drives task execution, validity checks, exact scoring, and the replayable training record.

This shared-state design keeps generation, supervision, and verification aligned. Each finalized record contains the rendered problem, exact scoring contract, image-space annotation, and an execution-trace reference for inspection and replay.

TRACE validation

The released checkpoints improve accuracy on 2,000 previously unseen instances generated from the same 1,000 task programs. Each model is evaluated once with decoding seed 42.

Qwen2.5-VL-3B
Base24.45 Trace41.05
Improvement: +16.60
Qwen2.5-VL-7B
Base34.25 Trace51.55
Improvement: +17.30

These results measure new realizations within the TRACE task distributions. The evaluation below measures transfer to external benchmarks.

External benchmark transfer

Qwen2.5-VL models trained on 64,000 Trace instances improve the macro-average across 24 external benchmarks at both evaluated model scales. Values are mean ± sample standard deviation across decoding seeds 42, 43, and 44; paired deltas compare matched seeds.

Qwen2.5-VL-3B
Base39.34 ± 0.63 Trace42.85 ± 0.39
Paired improvement: +3.51 ± 0.25
Qwen2.5-VL-7B
Base47.93 ± 0.30 Trace51.99 ± 0.17
Paired improvement: +4.06 ± 0.41

Base versus TRACE by benchmark

The table reports all 24 external benchmarks for the matched 3B and 7B comparisons. Parenthesized values are mean seed-paired TRACE-minus-Base changes in percentage points.

Benchmark 3B Base 3B TRACE (Δ) 7B Base 7B TRACE (Δ)
Charts & Tables
ChartQAPro 31.57 ± 0.66 31.43 ± 1.33 (-0.14) 45.81 ± 0.24 48.05 ± 0.40 (+2.24)
CharXivReason 28.90 ± 1.11 34.67 ± 1.47 (+5.77) 39.73 ± 1.17 47.13 ± 0.35 (+7.40)
TableVQABench 69.27 ± 0.91 71.99 ± 0.28 (+2.72) 75.20 ± 0.90 78.31 ± 0.17 (+3.11)
EvoChart 48.51 ± 0.45 46.83 ± 1.36 (-1.68) 57.07 ± 0.72 64.91 ± 0.05 (+7.84)
Visual Math
MathVision 19.25 ± 0.65 25.35 ± 0.86 (+6.10) 24.81 ± 0.75 27.42 ± 0.62 (+2.61)
MathVista 58.13 ± 3.74 64.43 ± 2.12 (+6.30) 68.67 ± 0.40 73.37 ± 0.45 (+4.70)
MathVerse 33.59 ± 1.95 40.02 ± 1.52 (+6.43) 43.44 ± 0.92 47.76 ± 0.51 (+4.31)
WeMath 17.87 ± 1.24 28.82 ± 0.40 (+10.95) 35.20 ± 2.92 46.16 ± 1.32 (+10.96)
Science & General
PhyX mini MC 32.80 ± 9.96 37.47 ± 5.58 (+4.67) 40.97 ± 3.57 48.70 ± 0.82 (+7.73)
MMMU-ProVis 26.59 ± 0.32 31.16 ± 1.20 (+4.57) 35.70 ± 0.39 39.36 ± 0.71 (+3.66)
RealWorldQA 60.35 ± 0.42 62.14 ± 1.18 (+1.79) 65.45 ± 0.87 68.50 ± 0.68 (+3.05)
MMStar 52.27 ± 0.52 55.24 ± 1.02 (+2.98) 61.89 ± 0.87 65.64 ± 0.34 (+3.76)
Spatial Reasoning
EmbSpatial 59.07 ± 1.04 60.88 ± 1.13 (+1.81) 69.42 ± 0.90 70.95 ± 0.69 (+1.53)
SpatialVizBench COT 30.08 ± 1.22 31.84 ± 1.52 (+1.75) 35.14 ± 0.10 35.68 ± 0.34 (+0.54)
CV-Bench 3D 58.67 ± 8.32 66.97 ± 3.68 (+8.31) 76.25 ± 1.96 81.00 ± 0.43 (+4.75)
ERQA 35.33 ± 0.80 36.42 ± 0.52 (+1.08) 39.00 ± 1.56 41.17 ± 1.28 (+2.17)
Perception & Counting
BLINK 44.52 ± 2.10 47.13 ± 0.70 (+2.61) 53.31 ± 1.23 56.57 ± 0.73 (+3.26)
CountBenchQA 65.43 ± 1.59 68.65 ± 0.83 (+3.22) 82.14 ± 1.48 84.80 ± 1.44 (+2.67)
CountQA 14.88 ± 1.80 15.64 ± 0.57 (+0.76) 19.59 ± 1.17 22.51 ± 0.91 (+2.92)
TreeBench 39.26 ± 1.96 38.60 ± 1.17 (-0.66) 38.02 ± 1.37 40.25 ± 2.11 (+2.22)
Puzzles & Logic
PuzzleVQA 32.55 ± 1.21 39.13 ± 1.13 (+6.58) 44.57 ± 0.64 50.02 ± 0.50 (+5.45)
VisualPuzzles 26.20 ± 2.74 28.51 ± 0.70 (+2.31) 31.08 ± 0.15 34.02 ± 0.77 (+2.94)
LogicVista 36.47 ± 1.25 40.12 ± 1.71 (+3.65) 42.65 ± 2.24 47.35 ± 0.68 (+4.70)
MME-Reasoning 22.62 ± 1.81 25.06 ± 1.53 (+2.44) 25.17 ± 1.18 28.03 ± 0.89 (+2.86)

TRACE improves the mean score on 21 of 24 benchmarks at 3B and all 24 at 7B. The research page includes the full eight-model comparison and links to the evaluation artifacts.

Released models

Released checkpoint Base model
TRACE Qwen2.5-VL-3B Qwen2.5-VL-3B-Instruct
TRACE Qwen2.5-VL-7B Qwen2.5-VL-7B-Instruct

See the 3B model card or 7B model card for pinned Transformers inference.

Reproduce

Follow the paper-results reproduction guide for training, evaluation, progress reporting, and validation.

Explore the documentation

For complete examples and maintenance interfaces, see the runnable Python examples and repository maintenance and release scripts.

Citation and acknowledgements

Use CITATION.cff for repository citation metadata and paper.pdf for the manuscript. The released checkpoints build on Qwen2.5-VL and EasyR1. External evaluator and dataset integration builds on VLMEvalKit, and bundled resource attributions are listed in THIRD_PARTY_NOTICES.md.

Repository branches. main contains the stable task-generation package and contracts; dev contains contributor review tools and development workflows; and rlvr contains model training, TRACE validation, and the trace_eval_v1 external-evaluation workflow.