Vision-G1: Multi-Domain Visual Reasoning

• Vision-G1 is a 7B-scale visual reasoning model that broadens RL-based reasoning beyond math and logic tasks.
• It integrates a curated RL-ready corpus from 46 data sources across 8 domains using multi-round GRPO and influence-based filtering.
• The approach uses influence functions and difficulty filtering to select moderately challenging examples, driving balanced performance across heterogeneous tasks.

Vision-G1 is a 7B-scale reasoning vision-LLM trained to extend reinforcement-learning-based visual reasoning beyond the mathematical and logical tasks that dominate earlier pipelines. Initialized from Qwen2.5-VL-7B-Instruct, it combines a curated RL-ready corpus drawn from 46 data sources across 8 domains with multi-round GRPO training, influence-function-based sample selection, and difficulty-based curriculum filtering. The stated objective is general vision-language reasoning across heterogeneous domains such as infographic, spatial, cross-image, GUI, medical, commonsense, and general science reasoning, rather than competence confined to narrowly verifiable math-style tasks (Zha et al., 18 Aug 2025).

1. Problem setting and model identity

The central problem addressed by Vision-G1 is the domain narrowness of existing reasoning VLM pipelines. Prior RL-based systems are described as being optimized mainly for tasks with readily verifiable answers, especially math and logic, which creates models that can perform strongly on those tasks yet remain brittle on broader visual reasoning workloads. The paper attributes this brittleness to two linked constraints: the scarcity of RL-ready, verifiable reward data outside math/logical settings, and the uncertainty of whether heterogeneous domain datasets can be mixed without harming training (Zha et al., 18 Aug 2025).

Vision-G1 is presented as the resulting model from this reformulation of the training problem. It is a 7B reasoning VLM initialized from Qwen2.5-VL-7B-Instruct and trained with Group Relative Policy Optimization (GRPO) using exact or otherwise verifiable answer correctness as reward. Its intended scope includes tasks requiring visual perception, spatial understanding, commonsense, scientific knowledge, cross-image comparison, GUI or web interpretation, medical reasoning, and infographic or document reasoning. In this framing, the paper’s main claim is not merely that more domains are added, but that data curation + filtering + curriculum RL can produce broader generalization than math-centric RL alone.

A common source of confusion is nomenclature. Vision-G1 denotes the multi-domain visual reasoning model of Zhao et al. (Zha et al., 18 Aug 2025); it is distinct from the graph-reasoning G1 model (Guo et al., 24 May 2025), the game-oriented VLM G1 system (Chen et al., 19 May 2025), and the Unitree G1 humanoid robot used as an embodied testbed in edge-VLM deployment research (Ahmad et al., 21 Jan 2026).

2. RL-ready dataset construction and domain coverage

The training corpus is built from 46 visual reasoning datasets organized into 8 domains and also described as spanning 13 fine-grained dimensions. The paper’s emphasis is that heterogeneity is intentional, but that heterogeneity is made tractable by converting all sources into a unified RL-ready format consisting of prompt/question, images, and ground truth (Zha et al., 18 Aug 2025).

Domain	Example datasets	Representative task types
Infographic reasoning	FigureQA, DVQA, PlotQA, ChartQA, TabMWP, MapQA, ChartBench, UniChart, DocVQA, MultiUI	Charts, plots, tables, maps, documents, web
Mathematical reasoning	Geometry3K, GeoQA+, UniGeo, GeoQA, MMR1, GEOS, CLEVR-Math	Geometry, arithmetic
Cross-image reasoning	IconQA, NLVR2, ImageCode	Multi-image comparison
Spatial reasoning	VQA-AS, Super-CLEVR	3D spatial relations
Science / commonsense / medical	AI2D, TQA, ScienceQA, MMK12, VQA2.0, VizWiz, TextVQA, A-OKVQA, OK-VQA, PMC-VQA, VQA-RAD, SLAKE, Path-VQA	Science, commonsense, clinical and pathology image reasoning

To make these sources usable for RL with rule-based rewards, the preprocessing stage retains only instances whose answers can be judged automatically, including numeric values, multiple-choice options, yes/no answers, and other single-word ground truths. Open-ended, non-verifiable examples are discarded by rule-based checks and regular expressions. Prompt standardization appends a reasoning-oriented instruction inspired by ThinkLite and requires the final answer to appear in $\boxed{\cdot}$, which simplifies answer extraction and reward computation.

The paper also introduces explicit category labeling. Each instance is assigned a category label by a VLM classifier, specifically Qwen2.5-VL-32B-Instruct, so that domain distribution can be tracked and balanced sampling can be enforced. This indicates that dataset construction is not treated as passive aggregation; it is treated as a controlled precondition for RL.

3. Influence-based selection and difficulty filtering

A core technical contribution of Vision-G1 is the claim that mixed-domain RL benefits from selecting samples not only by domain coverage but also by estimated training utility. The first mechanism is an influence function implemented as gradient similarity. The paper defines pairwise influence between two instances $z$ and $z'$ as

$$I(z,z') \propto \operatorname{Sim}\bigl(\nabla l(z,\theta), \nabla l(z',\theta)\bigr),$$

where $l(\cdot,\cdot)$ is the cross-entropy loss and $\operatorname{Sim}$ is cosine similarity after projection (Zha et al., 18 Aug 2025).

Instance-level influence is then aggregated over both same-domain and cross-domain neighborhoods:

$$I(z) = \frac{1}{|\mathcal{D}_{\operatorname{dom}(z)}|} \sum_{z' \in \mathcal{D}_{\operatorname{dom}(z)}} I(z,z') + \frac{1}{|\mathcal{D}\setminus \mathcal{D}_{\operatorname{dom}(z)}|} \sum_{z' \in \mathcal{D}\setminus \mathcal{D}_{\operatorname{dom}(z)}} I(z,z').$$

Operationally, the method fine-tunes a small LoRA module on high-quality rollouts from a subset, computes gradients for remaining examples, applies random projection for efficiency, uses cosine similarity as the influence estimator, and filters out instances with low or negative influence. The stated intuition is that such examples are likely to be unhelpful, noisy, incompatible with other tasks, or detrimental to generalization.

The second mechanism is difficulty-based filtering. The paper argues that extremely easy examples contribute little, while extremely hard examples may induce misleading rollouts because the reward verifies only the final answer rather than the intermediate reasoning chain. For each instance, the previous-round checkpoint performs $k$ rollouts and computes average accuracy; only examples satisfying

$0.2 \le \text{avg accuracy} \le 0.8$

are retained, with

$k = 16.$

This design keeps moderately difficult samples and implements a curriculum in which the trainable subset evolves with model competence. A plausible implication is that Vision-G1 treats dataset quality as a dynamic quantity conditioned on the current policy, rather than as a fixed property of the raw corpus.

4. Multi-round GRPO training procedure

The optimization stage uses GRPO, described as a PPO-style algorithm that uses multiple samples per prompt and does not require a separate value function. Vision-G1 uses rule-based exact-answer checking, with reward defined as $z$0 for a correct response and $z$1 for an incorrect response, followed by group normalization of the advantage:

$z$2

The reference model for KL regularization is Qwen2.5-VL-7B-Instruct (Zha et al., 18 Aug 2025).

The training loop is explicitly iterative: start from Qwen2.5-VL-7B-Instruct, curate data with influence filtering, measure difficulty using previous-round model rollouts, keep moderate examples, train with GRPO, and repeat for the next round. The main setup uses two rounds, and the paper reports a training trajectory showing stability and improvement after each round.

The reported experimental configuration is unusually specific. After filtering, the final training set contains 40k questions. Training uses batch size 128, 32 rollouts per question, 8 × NVIDIA H200, and about 18 hours of training time. Evaluation uses greedy decoding, pass@1 / accuracy, and answer matching with math-verify + normalized exact string matching. These details matter because the contribution is not only algorithmic; it is also an explicit recipe for post-training a general visual reasoner with verifiable RL.

5. Benchmarks and reported empirical performance

Vision-G1 is evaluated on 17 benchmarks spanning general visual reasoning, math-related visual reasoning, and domain-specific tasks. On the general visual reasoning suite, the reported scores are 76.1 on MathVista, 53.4 on MMMU-Val, 41.2 on MMMU-Pro, 66.0 on MMStar, 50.2 on LogicVista, and 90.8 on ChartQA. The paper further reports gains over Qwen2.5-VL-7B of +8.7 on MathVista, +2.9 on MMMU-Pro, +3.2 on MMStar, +7.6 on LogicVista, and +2.5 on ChartQA (Zha et al., 18 Aug 2025).

On math-related visual reasoning, Vision-G1 reports 31.3 on MathVision, 51.9 on MathVerse, 23.7 on OlympiadBench, 45.1 on WeMath, 58.5 on DynaMath, and 42.1 average. The paper notes that Vision-R1 can perform very well on some math-centric benchmarks, but states that Vision-G1 is more balanced across domains and even performs better than OpenAI o1 on MathVista.

On domain-specific reasoning, the reported scores are 44.0 / 65.5 on ChartXiv, 47.7 on ChartQAPro, 72.1 on VQA-RAD, 66.7 on PathVQA, 78.3 on SLAKE, and 61.5 on MuirBench. The paper characterizes this as notable because several RL-trained models reportedly degrade on domain-specific benchmarks when those domains are underrepresented in training.

Benchmark group	Selected benchmarks	Reported scores
General visual reasoning	MathVista, MMMU-Val, MMMU-Pro, MMStar, LogicVista, ChartQA	76.1, 53.4, 41.2, 66.0, 50.2, 90.8
Math-related reasoning	MathVision, MathVerse, OlympiadBench, WeMath, DynaMath	31.3, 51.9, 23.7, 45.1, 58.5
Domain-specific reasoning	ChartXiv, ChartQAPro, VQA-RAD, PathVQA, SLAKE, MuirBench	44.0 / 65.5, 47.7, 72.1, 66.7, 78.3, 61.5

The comparison set includes similar-sized models such as Qwen2.5-VL-7B, Ovis2-8B, MiniCPM-V2.6, LLaVA-Next-34B, Llava-OV-7B, MM-Eureka-7B, Vision-R1-7B, ThinkLite-VL-7B, and VL-Rethinker-7B; larger open models such as Qwen2.5-VL-72B, InternVL2.5-78B, InternVL3-78B, and VL-Rethinker-32B / 72B; and proprietary models including GPT-4o, Claude-3.5, Gemini-1.5 Flash, and Gemini-1.5 Pro. The paper’s summary claim is that Vision-G1 achieves state-of-the-art performance across various visual reasoning benchmarks, outperforming similar-sized VLMs and even some proprietary systems.

6. Cross-domain transfer, ablations, and limitations

The ablation results are used to support the paper’s central interpretation that breadth alone is insufficient; curation and curriculum are necessary. Removing multi-round RL hurts performance across most benchmarks. Removing influence-function filtering and difficulty filtering also causes significant drops. Training only on a smaller high-quality subset such as ThinkLite + ViRL39k improves some math scores but hurts domain-specific generalization. The intended conclusion is explicit: general visual reasoning requires diverse, carefully curated multi-domain training data (Zha et al., 18 Aug 2025).

The paper also reports a domain-transfer heatmap. The described patterns are asymmetrical and instructive: math data transfers well to infographic and cross-image tasks, infographic data can benefit math, medical-only training underperforms even on medical benchmarks, and mixing all domains with the proposed selection strategy yields the best overall performance. This suggests that cross-domain compatibility exists, but is not automatic; it must be mediated by sample selection and curriculum design.

Several practical limitations or caveats are either stated directly or implied. The model relies on verifiable-answer filtering, so open-ended supervision remains outside the main RL loop. The training corpus is broad, but the final effective set is only 40k questions after filtering, indicating that curation removes a substantial amount of raw data. The abstract states that the model, code and dataset are publicly available, but the checklist notes that the fully reproducible package was not yet released at submission time and would be released after acceptance. That discrepancy is best read as a release-timing caveat rather than a methodological one.

In the broader landscape, Vision-G1 occupies a specific position. Unlike the graph-reasoning G1 work, which uses synthetic graph-theoretic tasks and RLVR to improve LLM graph reasoning (Guo et al., 24 May 2025), and unlike the game-oriented G1 work, which studies the “knowing-doing” gap through VLM-Gym and interactive game RL (Chen et al., 19 May 2025), Vision-G1 is a multi-domain post-training recipe for static but heterogeneous visual reasoning. It is therefore best understood as an argument that RL-based VLM reasoning should not be treated as a math-only problem, and that verifiable supervision can be broadened substantially when dataset construction, influence estimation, and difficulty-aware curricula are designed jointly.