cs.CVApr 6, 2026
SaveVero: An Open RL Recipe for General Visual Reasoning
Organizations: Princeton University
Abstract
What does it take to build a visual reasoner that works across charts, science, spatial understanding, and open-ended tasks? The strongest vision-language models (VLMs) suggest that broad visual reasoning is within reach, yet their closed data and reinforcement learning (RL) pipelines make their gains difficult to study, reproduce, or extend. We introduce Vero, a family of fully open VLMs that match or exceed existing open-weight models across diverse visual reasoning tasks. We scale RL data and rewards across six broad task categories, constructing Vero-600K, a 600K-sample dataset from 59 datasets, and designing task-routed rewards that handle heterogeneous answers. Across VeroEval, our 30-benchmark suite, Vero-600K outperforms existing RL datasets under controlled comparisons. Applied to five starting models, Vero variants gain 2.9-5.4 points on average over their initial models. Notably, Vero-Qwen3I-8B, trained on the Instruct model, surpasses Qwen3-VL-8B-Thinking by 3.8 points on average without additional distillation. Systematic ablations reveal that different task categories elicit distinct reasoning patterns and that broad gains depend on learning them jointly rather than in isolation. All data, code, and models are publicly available.
Figures & tables
Figure 1: Vero achieves state-of-the-art performance across six task categories using a fully open RL recipe. Top left: Training curves versus RL FLOPs for Vero-600K compared with existing open RL datasets, all finetuned from Qwen2.5-VL-7B-Instruct; dashed lines indicate training beyond one epoch. Reported is mean sem (3 seeds). Top center: Summary of the broad task categories targeted in Vero-600K and VeroEval . Top right: Overall performance of Vero (Avg@5) against each initial RL model across the 30 VeroEval benchmarks. Bottom row: Same as top right, but displaying per-category scores for Vero trained on Qwen3-VL-8B-Instruct and Qwen3.5-9B.
Figure 2: Composition of Vero-600K . The inner ring shows six task categories, each allocated 100K samples (600K total), and the outer ring shows their 59 constituent datasets. The categories represent real-world use cases and cover distinct visual reasoning capabilities (Sections 5 and 7 ). Categories are uniformly sampled to balance learning across tasks.
Figure 3: Vero-600K data curation pipeline. Starting from over 250 candidate datasets, we assign each to one of six task categories and apply multi-stage selection and filtering: heuristic screening (size, resolution, answer format), manual quality control, LLM-based question filtering for ambiguity and verifiability, and answer filtering for stable reward computation. The retained data are combined into a uniformly weighted mixture across task categories and used for on-policy RL training with task-routed rewards.
Figure 4: Examples from each task category illustrate the breadth of Vero-600K . We show representative samples of our training data from the six categories, highlighting the diversity of visual inputs, question formats, and answer types covered by our training set.
Figure 5: Accuracy verifiers for our multi-task reward. Each card illustrates a verifier used in Vero , with an example visual question, ground-truth answer, and model prediction. Verifiers include binary rewards (string match, multiple choice (not displayed), list string match (not displayed), numeric, ordering, point-in-box) and graded rewards (IoU/F1 for grounding, field match for web actions, rule-based checks for instruction following, and LLM-as-judge for open-ended responses). This task-routed design enables accurate reward computation across diverse answer formats.
Figure 6: Per-category RL training curves. Evaluation score vs. RL FLOPs for RL runs using Vero-600K and three prior open RL datasets, all starting from Qwen2.5-VL-7B-Instruct under identical RL settings. Curves report mean and SEM across seeds where shown. Dashed segments indicate training beyond one epoch. Vero-600K leads on five of six categories throughout training; on STEM it remains within 2 points of the best prior dataset.
Figure 7: Diverse task mixing eliminates negative cross-task transfer. Each row shows a model trained on a single task category (or mixture of all categories). Values are absolute score changes relative to the base model. Single-task training yields selective transfer, while mixing achieves consistent gains.
Figure 10: Performance improves with greater training data exposure. We plot benchmark performance as each model sees progressively more samples from the fixed 600K RL training mixture over the course of a single training pass.
Figure 11: Single-task training elicits a range of behavioral profiles, many of which are distinct. Average behavior presence rates for Vero on Qwen3-VL-8B-Instruct, aggregated over six validation task categories after balanced subsampling. 34 high-level cognitive behaviors, sorted by overall prevalence. Single-task training elicits distinct high-level strategies, and these differences carry through to the skill level.
Figure 12: Task category skills are largely distinct. Each bar shows the predicted classification distribution (as proportions) from a logistic regression probe trained on skill embeddings, with the values indicating per-category accuracy. The probe achieves 0.77 overall accuracy, with STEM (0.84) and Chart & OCR (0.82) yielding the most distinctive skills. Knowledge & Recognition is the least separable (0.59), with notable confusion toward Grounding, Counting & Search, reflecting shared visual grounding operations across these categories.
Figure 13: Single-task training elicits distinct behavioral skills. Word clouds of extracted low-level skills reveal that models develop highly specialized strategies adapted to their specific training domains. For instance, STEM training fosters formula-driven logic (e.g., “Apply Arc Length Formula”); Grounding models prioritize localized perception (e.g., “Determine Relative Position”).
Appendix figures & tables9 assets
Supplementary material from the paper’s appendix.
Appendix
| Dataset | Ret. | Answer type | Reward type(s) |
|---|---|---|---|
| Grounding, Counting & Search | |||
| AerialVG ( Liu et al., 2025a ) | 12,634 | bbox coordinates | grounding |
| GroundUI ( Zheng et al., 2025b ) | 12,064 | click coordinates | clicking |
| MultiHop ( Li et al., 2026 ) | 6,316 | integer count | counting |
| Objects365-QA ( Shao et al., 2019 ) | 12,632 | bbox coordinates | grounding |
| OOD-VQA ( Shi and Lee, 2024 ) | 5,028 | integer count | counting |
Table A1 : Retained training datasets used in our RL mixture. Retained sizes are taken from the composition source used to generate the training-data figure. Captioning and instruction-following retained sizes are rounded display values.
| Benchmark | Description |
|---|---|
| Chart & OCR | |
| ChartQA-Pro ( Masry et al., 2025 ) | diverse chart question answering |
| ChartQA ( Masry et al., 2022 ) | chart reasoning and question answering |
| InfoVQA ( Mathew et al., 2022 ) | infographic question answering |
| CharXiv ( Wang et al., 2024b ) | scientific chart understanding |
| ChartMuseum ( Tang et al., 2025 ) | chart visual reasoning |
Table A2 : Evaluation benchmarks in VeroEval , organized by task category. The right column gives a short description of what each benchmark evaluates.
| Reason | Description | Example |
|---|---|---|
| Single-ground-truth datasets | ||
| Multi-value answers ( ) | Ground truth contains multiple distinct values that cannot be reduced to a single verifiable target. | AC = 4, BD = 4 |
| Ambiguous text labels ( ) | Descriptive phrases requiring fuzzy or semantic matching beyond exact string comparison. | “Isosceles triangle” |
| Unsupported notation ( ) | Scientific notation or symbolic algebraic expressions outside our numeric parser’s scope. | |
| Empty / invalid GT | Ground truth is missing, empty, or malformed, making reward computation impossible. | “Empty case” |
| Unit mismatch | Unit is inconsistent with the question context or cannot be cleanly stripped to a numeric value. | Mass answered in cm |
Table A3: Common reasons for answer filtering. The top section covers single-ground-truth datasets (frequency estimates from a manual sample); the bottom section covers multi-annotator datasets.
| Hyperparameter | Value |
| Framework | VeRL |
| FSDP strategy | fsdp2 |
| Rollouts per prompt ( ) | 8 |
| Train batch size | 256 |
| PPO mini-batch size | 128 |
| Learning rate |
Table B1: RL training hyperparameters for all Vero models. Training uses VeRL with GSPO ( Zheng et al., 2025a ) , asymmetric clipping ( ), and no KL penalty to allow less-restricted policy updates.
| Base model | GPUs | Steps | Ctx. | Max px. | Dtype | Coords |
|---|---|---|---|---|---|---|
| MiMo-VL-7B-SFT | 8 H100 | 2,000 | 28,672 | bf16 | absolute | |
| Qwen2.5-VL-7B-Inst. | 8 H100 | 2,000 | 24,576 | fp16 | absolute | |
| Qwen3-VL-8B-Inst. | 8 H200 | 2,000 | 36,864 | fp16 | norm. 0–1k | |
| Qwen3-VL-8B-Think. | 8 H200 | 2,000 | 36,864 | fp16 | norm. 0–1k |
Table B2: Base-model training configurations for the Vero variants, including context length, maximum image resolution, precision, and coordinate format. Qwen-family models use fp16 for improved training stability ( Qi et al., 2025 ) .
Figure D4 : Behavioral skill analysis presence rates for models trained on individual task categories and one trained on all categories (mix all). Training on individual task categories impacts the emergence of fine-grained skills within each domain, with green indicating higher skill presence and red indicating lower. For instance, chart skills center on reading operations like "Axis Analysis", while spatial skills center on physical state reasoning like "Mental Simulation".
| Skill | Definition |
|---|---|
| Captioning & Instruction Following | |
| Adjective Selection For Visual Description | Choosing descriptive adjectives that reflect style and function to capture a scene. |
| Adopt Perspective | Use vivid language to convey a specific perspective without using pronouns. |
| Analyze Atmosphere | Establish emotional connection to a setting via environmental interaction. |
| Analyze Contrasts | Use visual contrast (color, structure) to highlight key information. |
| Analyze Visual Composition | Assess visual elements like focal points to determine mood or message. |
Table D2 : Summary of skill definitions. A comprehensive glossary providing operational definitions for the fine-grained skills evaluated across all six task categories.