Project navigation
Results Demos Method Data Diversity Behavior

An Open RL Recipe for Visual Reasoning

Recipe: We introduce a single-stage, fully open RL recipe that trains broadly capable vision-language reasoners across six task categories, using 600K curated samples from 59 datasets with task-routed rewards.
Data: We construct Vero-600K, a diverse training set spanning Chart & OCR, STEM, Spatial & Action, Knowledge & Recognition, Grounding, Counting & Search, and Captioning & Instruction Following.
Evaluation: We assemble VeroEval, a suite of 30 challenging benchmarks, and show that Vero achieves state-of-the-art performance among open-weight 8B VLMs, improving four base models by 3.6–5.3 points on average.
Analysis: Systematic ablations reveal that different task categories elicit distinct reasoning patterns that transfer poorly in isolation, suggesting broad data coverage is the primary driver of strong RL scaling.
Paper logo Paper GitHub logo Code Dataset logo Dataset Model logo Models
Vero visual reasoning illustration
Gabriel Sarch*, Linrong Cai*, Qunzhong Wang, Haoyang Wu
Danqi Chen, Zhuang Liu†
*Project Leads   †Corresponding Author
Princeton shield Princeton University
Vero teaser showing performance across six task categories

State of the Art Performance Across Task Categories

We evaluate Vero on 30 benchmarks spanning six task categories. The same open recipe improves four different base models and reaches state-of-the-art overall performance among 8B open-weight VLMs.

Starting from Qwen3-VL-8B-Instruct, Vero-Q3I-8B raises the overall average from 60.7 to 66.0, with category gains of +8.5 on Chart and OCR, +6.4 on STEM, +3.7 on Spatial and Action, +1.0 on Knowledge and Recognition, +5.3 on Grounding, Counting and Search, and +5.6 on Captioning and Instruction Following. Applied on top of Qwen3-VL-8B-Thinking, Vero-Q3T-8B reaches 65.9 overall versus 62.3 for the base model, with its largest gains in Grounding, Counting and Search (+7.2) and Chart and OCR (+4.2). The same recipe also improves Qwen2.5-VL-7B-Instruct from 52.9 to 57.9 and MiMo-VL-7B-SFT to 63.3, exceeding MiMo-VL-7B-RL at 62.4.

† indicates evaluated by us. All other results are taken from official reports.

Vero Demos

Example conversations between a user and Vero across all six task categories. Each demo shows the model's reasoning trace and final answer.

1/3

Demo image
V
Vero-Q3T-8B

Method

Vero trains on 600K curated RL samples drawn from 59 datasets organized into six categories: Chart and OCR, STEM, Spatial and Action, Knowledge and Recognition, Grounding, Counting and Search, and Captioning and Instruction Following. The categories correspond to substantially different use cases, visual inputs, reasoning patterns, and answer formats.

The training mixture spans six task categories and 59 retained datasets after dataset-level and sample-level filtering.

Vero uses a single-stage RL recipe directly on top of instruction-tuned or RL'd base models. Vero uses GSPO-style optimization with task-routed verifiers, so numeric questions, multiple choice questions, grounding boxes, clicks, ordering problems, and open-ended instruction-following outputs.

Data Diversity and Transfer

We show that single-task RL does not generalize reliably across visual capabilities. Training on one category often improves that category while degrading others, especially Grounding and Captioning and Instruction Following. This is consistent with classic multi-task RL results showing that heterogeneous tasks can interfere and that task contributions must be balanced during training (Teh et al., 2017; Hessel et al., 2019). By contrast, the mixed model produces positive gains across categories and avoids the catastrophic spillover seen in single-task-category RL.

Behavioral Analysis

Different task categories do not simply induce more or less reasoning — they induce qualitatively different reasoning styles. STEM tasks trigger reflective, backtracking-heavy traces; grounding tasks favor direct perceptual search; chart tasks produce systematic regional synthesis. These distinct patterns help explain why single-task training transfers poorly: the model adapts not just its answers, but its reasoning policy.

Reasoning Length by Task Category

Beyond qualitative differences in reasoning style, task categories also elicit markedly different reasoning lengths. Spatial & Action produces the longest responses at 1,983 ± 51 words, followed by Chart & OCR (1,593 ± 32) and STEM (1,576 ± 40). Captioning & Instruction Following is much shorter (414 ± 13), while Grounding, Counting & Search (125 ± 13) and Knowledge & Recognition (76 ± 3) are shortest. The gap between the longest and shortest categories exceeds 26×, suggesting that long chain-of-thought behavior is concentrated in tasks requiring multi-step spatial state tracking or structured analytical decomposition.

RL on different task categories leads to varying reasoning lengths. Average reasoning length (in words) on the validation set, measured after training Qwen3-VL-8B-Instruct for 1,000 steps on each task category (100k samples) and evaluating on the same category. Error bars denote the standard error of the mean.
Task categories cultivate distinct skill repertoires. A logistic regression probe trained on 1,500 extracted skills per task category reaches high overall accuracy at recovering the source task category from the extracted skill lists, suggesting partially distinct skill repertoires rather than a generic increase in chain-of-thought length.

Interactive UMAP

The stacked-bar summary highlights the same task category separability at the category level. The interactive UMAP below shows the same story at the individual-skill level, where clusters can be inspected directly by task category, label, and description.

Scroll to zoom. Drag to pan. Hover a point for the behavior label and description.