Trace: A Taxonomy-Guided Environment for Multidomain Visual Reasoning

• Visual RLVR remains constrained by the challenge of combining visually grounded questions, exactly verifiable answers, and variation beyond narrow templates; prior approaches trade breadth for coherent control or exact supervision.Large mixtures inherit heterogeneous task structures [29], while procedural systems often center on one grammar or closely related objectives [1] [8] [38] [45] [51].Source: §1 · p. 1
• TRACE separates scene grammars, executable task programs, reward contracts, and bounded query variation to provide stable task units across visual reasoning tasks.A scene grammar defines renderable objects, relations, and layout; a task program specifies computation and answer type, while the reward contract defines exact verification.Source: §1 · pp. 1–2
• TRACE comprises 1,000 tasks across 277 scene grammars and 11 visual domains, with each semantic state determining the image, prompt, typed answer, and verifier state.This shared state enables exact supervision from task computation and controlled semantic and visual variation; Figure 1 shows representative instances across the domains.Source: §1 · p. 2, Figure 1 · p. 2
• 3.51 percentage points at 3B and 4.06 points at 7B: RLVR on 64,000 TRACE instances improves the 24-benchmark macro-average, with positive mean changes on 21 of 24 benchmarks at 3B and all 24 at 7B.Both Qwen2.5-VL models were evaluated on external visual-reasoning benchmarks, testing transfer beyond held-out instances from TRACE.Source: §1 · p. 2

Related work positioning

• Structured datasets use explicit representations to control visual-reasoning coverage, and TRACE extends this lineage by making executable tasks stable units for generation, sampling, verification, and RL [16] [18] [5] [57] [53].CLEVR and GQA use scene graphs and functional programs; PuzzleVQA targets abstract patterns, Task Me Anything uses an extensible taxonomy, and WorldBench curates diverse real-image questions.Source: §2 · pp. 2–3
• Reasoning Gym and Enigmata show how inspectable generators and exact verifiers support broad training distributions, but primarily use textual or symbolic inputs [34] [3].Reasoning Gym spans more than one hundred environments, while Enigmata has 36 puzzle tasks in seven categories; TRACE applies this model to rendered visual states.Source: §2 · p. 3
• Multimodal RL mixtures route rewards across heterogeneous tasks; Vero scales to 600K examples from 59 datasets, while OpenThoughts emphasizes reproducible data recipes [15] [22] [25] [27] [31] [37] [52] [29] [12].The cited work spans classification, visual mathematics, counting, structural perception, spatial reasoning, and referring expressions; Vero’s mixture ablations indicate the importance of broad category coverage for transfer.Source: §2 · p. 3
• Example-selection research treats difficulty, multimodal value, source mixtures, influence, curricula, and verified synthesis as design variables [20] [44] [21] [50] [35].SoTA with Less selects challenging examples, RAP uses multimodal/text-only discrepancies and attention-based confidence, and MoDoMoDo predicts downstream outcomes from source distributions; Vision-G1 and WeThink use source selection or curricula, while DeepVision-103K expands visual-mathematics data.Source: §2 · p. 3
• Visual RL systems create verifiable objectives via rule-to-image pipelines, executable games, compositional questions, modified data, 25 Sphinx task families, and TRON’s 520 online generator–verifier environments across five ability groups [8] [38] [11] [43] [47] [1] [51].Source: §2 · pp. 3–4
• TRACE separates 277 scene grammars from 1,000 stable task programs across 11 visual domains, allowing objectives to share scenes and visual realizations to change without redefining tasks.Game-RL separates games from their tasks within one visual domain, while TRON treats each generator–verifier program as an online environment.Source: §2 · p. 4

3.1 Task identity and equivalence

• A task is defined by its scene grammar, executable program, answer schema, and reward contract; changes to visual or computational structure create distinct tasks.Programs specify candidate construction, operand roles, intermediate computations, the final operator, and answer binding; literal operand renaming alone does not change task identity.Source: §3.1 · p. 4

3.2 Task, query, and generation boundaries

• Bounded program arguments remain queries, while semantic changes may alter answers and render-only changes preserve answer-relevant relations and reward.Treating each operand choice as a task would overcount one computation, while merging different intermediate operations would obscure reasoning differences.Source: §3.2 · pp. 4–5

3.3 Canonical task programs and extensibility

• Operators are canonicalized across scene grammars to expose shared reasoning, but tasks remain distinct when their grammars, metrics, or visual result bindings differ.TRACE samples tasks uniformly by default and draws query values after task selection; extensions can add a grammar or add a program and reward contract to an existing grammar.Source: §3.3 · pp. 5–6

4.1 Executable instance generation

• TRACE executes each task over a semantic state, then binds its typed answer and verifier state to an instance scorer while rendering the image and prompt and recording a replayable trace.The trace records the semantic state, query, task execution, rendering decisions, and verifier state; independent seeded streams allow reconstruction without coupling prompt or rendering variation to task computation.Source: §4 · p. 6, §4.1 · p. 6

4.2 Environment composition

• The environment contains 1,000 tasks across 277 scene grammars and 11 visual domains, reusing visual constructions across multiple reasoning objectives.Scene grammars support a median of three tasks (mean 3.61, range 1–26); 790 tasks use open-form integer, numeric, or string answers.Source: §4.2 · p. 7
• Of 1,000 tasks, 317 offer multiple queries, yielding 1,475 variants without changing task sampling probability; TRACE also analyzes tasks across 13 multi-label operation families.The operation families are an analytical view of task programs, and every task belongs to at least one family.Source: §4.2 · p. 7

4.3 Controlled visual realization

• Semantic controls vary task-relevant state, while scene-compatible realization controls vary presentation only when queried relations and the typed answer remain unchanged.For information-rich scenes, curated context outside answer-bearing regions must pass fit, collision, and contrast checks; Figure 6 illustrates answer-preserving variation in theme, typeface, layout, context, and raster treatment.Source: §4.3 · p. 7, Figure 6 · p. 8

4.4 Instance validation

• Instances are retained only when execution yields a unique typed answer, prompt and image match the semantic state, and deterministic replay reconstructs the same instance.Automatic structural and rendering checks are complemented by sample review for semantic clarity, prompt–image consistency, and visual legibility.Source: §4.4 · p. 8

5.1 Training data and reward

• Both model scales train on the same 64,000 instances, with 2,000 held-out validation examples and all results reported from the final checkpoint after 500 updates.The training set has 64 generated examples per task; validation has two previously unseen instances per task.Source: §5.1 · p. 9
• The response must end in a JSON object with the requested key; exact type-aware answer matching determines reward, while a smaller format term encourages parseable output.The model may produce an unrestricted reasoning trace before the required JSON object.Source: §5.1 · p. 9

5.2 Optimization

• GRPO updates sample 128 prompts with eight responses each; 500 updates make one shuffled pass over training prompts and yield 512,000 sampled responses.Rewards are normalized within each response group and optimized with a clipped token-level policy objective.Source: §5.2 · p. 9
• Both scales use full-parameter BF16 training at learning rate 10^-6 and share data, reward, sampling schedule, and optimization hyperparameters.Training uses eight H100 80GB GPUs and takes 12.2 hours for 3B and 13.9 hours for 7B.Source: §5.2 · p. 9

5.3 External evaluation

• Evaluation covers 24 external benchmarks across six groups, with full designated splits totaling 32,805 examples per model and decoding seed.Category scores average four benchmarks equally, and the overall score averages all 24 equally so larger evaluation sets do not dominate.Source: §5.3 · p. 9
• Each fixed checkpoint is decoded with three seeds and evaluated using the same prompts, preprocessing bounds, parsers, scorers, and VLMEvalKit version.Seeds affect response generation only; model weights remain unchanged, and reported values are means with sample standard deviations.Source: §5.3 · p. 9
• At 7B, synthetic-data RLVR checkpoints are evaluated alongside the models, while Vero is reported separately as a real-image reference; unmatched training and compute make these outcome comparisons, not controlled method comparisons.The compared checkpoints are from Game-RL, Sphinx, and PC-GRPO; Vero uses a 600K-example mixture from 59 datasets, whereas TRACE runs use 64K procedural instances.Source: §5.3 · p. 9

6.1 TRACE validation

• Accuracy on unseen instances from the same task programs increases from 24.45 to 41.05 at 3B (+16.60 points) and from 34.25 to 51.55 at 7B (+17.30 points).These results measure generalization to new semantic and visual realizations within TRACE task distributions.Source: §6.1 · p. 10
• Formula evaluation and aggregation have the largest operation-family gains at both scales, while results vary by domain, computational structure, answer interface, and query count.Performance improves across all four answer interfaces and both single- and multi-query tasks; numeric-answer tasks increase most despite comprising 51 tasks, while option-letter tasks increase least.Source: §6.1 · p. 10

6.2 External benchmark transfer

• 3B benchmark-macro average rises from 39.34 ± 0.63 to 42.85 ± 0.39 (+3.51 points), and 7B rises from 47.93 ± 0.30 to 51.99 ± 0.17 (+4.06 points).The averages are unweighted across the 24 benchmarks.Source: Table 2 · p. 11, §6.2 · p. 11
• All six category averages increase at both scales, mean performance improves on 21 of 24 benchmarks at 3B and all 24 at 7B, and bootstrap intervals exclude zero on 18 and 21 benchmarks, respectively.EvoChart at 3B is the only benchmark with an interval entirely below zero.Source: §6.2 · p. 11
• Visual mathematics gains 7.44 points at 3B and 5.65 at 7B, while WeMath has the largest individual improvement: 10.95 and 10.96 points, respectively.The gains extend beyond held-out procedural images to external benchmarks with different visual formats, task structures, and evaluation metrics.Source: §6.2 · p. 11
• The TRACE-trained 7B checkpoint exceeds Game-RL-7B by 3.94 macro-average points, Sphinx-7B by 2.64, and PC-GRPO-7B by 3.18.It has the highest reported mean on 21 of 24 benchmarks and highest category average in all six groups; these checkpoint comparisons do not isolate method effects because data, optimization, and compute are unmatched.Source: Table 2 · p. 11, §6.2 · p. 12
• At 3B, ChartQAPro, TreeBench, and EvoChart shift from −0.14, −0.66, and −1.68 points to gains of 2.24, 2.22, and 7.84 at 7B.The scale-level transfer pattern is positively associated overall, but the benefit of a fixed procedural training mixture depends on model capacity and benchmark demands.Source: §6.2 · p. 12

6.3 Training dynamics

• Mean training reward rises from 0.240 to 0.432 at 3B and from 0.353 to 0.542 at 7B, while completion length grows and sampled-token entropy falls at both scales.Completion length changes from 198 to 308 tokens at 3B and 211 to 237 at 7B; entropy changes from 0.829 to 0.137 and 0.361 to 0.088, respectively, with multi-hundred-token completions retained.Source: §6.3 · p. 12

7 Limitations and Future Work

• The human-designed taxonomy’s 1,000 tasks do not exhaust visual reasoning, and its operation families do not define comparable difficulty across domains.Future work could model complexity using program structure, semantic-state size, visual density, and empirical solve rates.Source: §7 · p. 12
• Transfer from the complete TRACE mixture does not identify which domains, operation families, sampling choices, or rendering controls drive gains.The study uses uniform task sampling for one fixed training pass; task identities and generation parameters could enable mixture ablations, difficulty-aware curricula, adaptive sampling, and optimization based on competence or learning progress.Source: §7 · p. 12
• One training run per model scale leaves independent training variation unmeasured, while unmatched RLVR checkpoints support only descriptive comparisons.The external suite cannot establish transfer to every natural-image distribution, some benchmarks use model-based graders, and procedural consistency does not ensure equal perceptual difficulty; richer rendering sources and replay-based failure analysis are proposed directions.Source: §7 · p. 13

8 Conclusion

• RLVR on 64,000 TRACE instances improves the 24-benchmark macro-average by 3.51 points at 3B and 4.06 points at 7B, with positive mean changes on 21 and all 24 benchmarks, respectively.These results provide evidence that broad, structured procedural experience can support transferable visual reasoning beyond regenerated instances from the same task distributions.Source: §8 · p. 13
• TRACE makes scene grammars, task programs, bounded queries, and reward contracts explicit and separable within an executable environment.This provides stable units for generation, task-balanced sampling, verification, and analysis.Source: §8 · p. 13
• Every TRACE instance is exactly verifiable and deterministically replayable across semantic and visual variation.Source: §8 · p. 13

Broader Impact

• Trace may make multimodal-reasoning research more reproducible and reduce reliance on sensitive or copyrighted imagery, but consequential uses and training costs warrant documented safeguards.Releases should document intended uses, licensing, resource requirements, and downstream-use constraints.Source: Broader Impact · p. 13

A.1 Domain construction profiles

• Chart, game, geometry, graph, and icon scenes vary layouts and appearance while keeping labels, measurements, or other queried elements tied to semantic state.Their generators use procedural chart marks, game states, solved geometric constructions, explicit graph structures, and curated icon assets.Source: §A.1 · p. 18
• Illustration, page, physics, puzzle, symbolic, and 3D scenes use domain-specific structures and vary presentation while preserving links to the queried state or relation.Examples include structured documents, validity-constrained puzzles, executable notation, and 3D scenes with explicit camera, depth, and occlusion constraints.Source: §A.1 · pp. 18–19

A.2 Operation-family definitions

• Task operations include visible lookup, filtering, counting, comparison, ranking, and aggregation over sets, candidates, or collections.Aggregation includes statistics such as sums, means, medians, cumulative totals, shares, mass, or related measures.Source: §A.2 · p. 19
• Other task operations cover logical composition, spatial relations, topology, matching, transformations, state updates, and formula evaluation.They include operations such as connectivity and reachability, geometric or symbolic changes, discrete transitions, and domain-specific calculations.Source: §A.2 · p. 19

A.3 Invariant enforcement

• Semantic constraints and answer uniqueness precede rendering; answer-bearing geometry is projected before validation, while render-only controls cannot change operands, answers, or rewards.Independent seeded streams and instance traces support replay of semantic state, prompt, rendering choices, and verifier state.Source: §A.3 · pp. 19–20, Figure 10 · p. 20

B Experimental Details

• Table 5 reports a configuration shared by the 3B and 7B training runs.Source: B Experimental Details · p. 20
• External inference uses temperature 0.6, top-p 1.0, a 4,096-token response limit, and decoding seeds 42, 43, and 44.Source: B Experimental Details · p. 20
• The decoding seeds affect response generation only; model weights remain fixed across decoding runs.Source: B Experimental Details · p. 20

B.1 TRACE validation slices

• Table 6 reports held-out TRACE accuracy by answer interface and query structure, with two previously unseen instances per task and one decoding seed for all slice scores.Source: §B.1 · p. 20
• The Table 6 rows are overlapping analytical views of the same task inventory, not separate evaluation sets.Source: §B.1 · p. 20

B.2 External evaluation suite

• Each external benchmark uses its standard scorer when available; open-answer benchmarks retain fixed model-based grading procedures.The same grader, prompt, and parser are used for every checkpoint on those open-answer benchmarks.Source: §B.2 · p. 20
• Table 7 lists the complete external benchmark suite, references, and evaluated sample counts.Source: §B.2 · p. 20

B.3 External benchmark uncertainty

• Table 8 reports paired confidence intervals for each base-to-TRACE improvement, jointly resampling items and all three decoding seeds, with WeMath problem families and TableVQABench split items as scoring units.For grouped metrics, resampling follows each benchmark’s official scoring unit.Source: §B.3 · p. 20

Charts

• Each page pairs rendered instances with questions and typed answers; chart examples include numeric intervals, reference containment, and centroid-option selection.Rule-heavy questions are condensed only when needed for legibility.Source: C Representative Task Atlas · p. 21, Figure 11 · p. 24

Games and geometry

• Game and geometry examples cover checkers capture-chain length, pyramid cross sections, circle-pair tangents, and protractor angles.Source: Figure 12 · p. 25, Figure 13 · p. 26

Graphs

• Graph examples ask for an undirected component or a binary tree’s local relative node.Source: Figure 14 · p. 27

Icons and illustrations

• Icon and illustration examples include matching wallpaper patterns, finding a named path’s successor, and completing a size progression.Source: Figure 15 · p. 28, Figure 16 · p. 29

Pages

• Page tasks include comparing infographic field totals, identifying a web action target, and finding a section-icon total difference.Source: Figure 17 · p. 30

Physics

• The physics example asks for an electrostatic field direction choice.Source: Figure 18 · p. 31

Puzzles

• Puzzle examples cover a cube-net marked-edge neighbor, a sheet-fold projection, and a word-search location.Source: Figure 19 · p. 32

Symbolic and 3D tasks

• Symbolic examples ask for a one-step Life cell state, bond order, and joint even-value probability; the atlas also includes representative 3D tasks.Source: Figure 20 · p. 33, Figure 21 · p. 34