Trinity: Self-Evolving Vision-Language Models with a Self-Verifier

Youngwan Lee; Yong-Ju Lee; Sung Ju Hwang

cs.AIOct 3, 2026

Trinity: Self-Evolving Vision-Language Models with a Self-Verifier

Authors: Youngwan Lee, Yong-Ju Lee, Sung Ju Hwang

Organizations: Electronics and Telecommunications Research Institute (ETRI), South Korea · Korea Advanced Institute of Science and Technology (KAIST), South Korea · DeepAuto.ai, South Korea

Abstract

Self-evolving vision-language models (VLMs), a form of self-improvement in which a model generates its own training data from unlabeled images, are a promising route toward agents that expand their reasoning capability in an unsupervised manner, without relying on ever-larger annotation budgets. Existing methods pair a Questioner that proposes problems with a Solver that answers them, but reward both roles mainly by agreement among sampled answers. Agreement is a weak proxy for truth: it cannot tell whether a question is grounded in the image, whether the proposed reference answer is right, or whether a confident majority is wrong in the same way. We present Trinity, in which one VLM plays three roles, Questioner, Solver, and Verifier, and the Verifier is a self-verifier: an exponential moving average (EMA) of the policy itself, requiring neither labels nor an external judge. The Verifier screens every generated question for image grounding and answer correctness before it becomes supervision, scores Solver reasoning against the image, and adjudicates disputes between the reference answer and a strong Solver consensus, correcting the reference and penalizing the Questioner when the consensus is right. Trained on images alone, Trinity improves Qwen3-VL-8B on mathematical visual reasoning and on science benchmarks with biology content, for example, +8.6 points on the biology split of SciVQR and +12.8 on MathVerse, and its reward dynamics behave as a healthy self-play curriculum should. These results suggest that a self-evolving multimodal agent can strengthen its scientific reasoning from unlabeled scientific images alone, with the model itself serving as the verifier.

Figures & tables

Figure 1: Trinity training loop. The policy $\pi_{\theta}$ as Questioner proposes $G$ candidate QAs $\{(q,a)\}$ ; the EMA Verifier $\pi_{\phi}$ screens them and selects $(q^{*},a^{*})$ ; the policy as Solver produces $K$ completions $\{(\tau,\hat{a})\}$ , which the Verifier scores. $R_{Q}$ and $R_{S}$ drive one joint update $\nabla_{\theta}\mathcal{L}_{\rm trinity}$ , after which $\phi\leftarrow\alpha\phi+(1-\alpha)\theta$ .

Model	MathVision	MathVerse	MathVista	MMMU	EMVista	SciVQR-Bio	MME-SCI
Qwen3-VL-8B-Instruct	52.3	58.2	76.7	57.0	44.8	56.7	19.0
MM-Zero [ 10 ]	39.6	45.1	67.2	58.3	45.0	62.5	17.6
DPE [ 9 ]	53.9	57.2	76.2	69.1	44.1	60.8	20.4
Trinity (Vero-600K)	57.8	71.0	79.6	69.6	47.0	65.3	20.3
Trinity (OmniScience)	58.2	70.6	79.6	70.7	47.0	62.5	21.9

Table 1: Performance comparison with Self-evolving VLMs using Qwen3-VL-8B-Instruct. Trinity is trained on Vero-600K or OmniScience images.

Figure 2: Reward dynamics for Qwen3-VL-8B on OmniScience [ 18 ] (steps 0–299).

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Model	MathVision	MathVerse	MathVista	EMVista	SciVQR-Bio	MME-SCI
Qwen2.5-VL-7B-Instruct [ 2 ]	26.6	50.0	70.7	43.9	38.3	11.1
VisPlay [ 7 ]	31.2	39.1	–	–	–	–
iReasoner [ 17 ]	25.3	45.9	69.7	–	–	–
V-Zero [ 20 ]	27.0	42.6	69.2	–	–	–
EvoLMM [ 19 ]	24.8	44.9	70.5	46.3	38.3	12.0
Trinity (Vero-600K)	27.7	51.5	71.9	46.9	41.1	11.7

Appendix

Table 2: Qwen2.5-VL-7B-Instruct results (accuracy, %) against published self-evolving methods on the same backbone. Mathematical numbers for VisPlay [ 7 ] , iReasoner [ 17 ] , and V-Zero [ 20 ] are taken from the respective papers; all other values are measured by us. Science benchmarks use a single run with the same prompt for all three models and the judge scoring of Appendix A.4 . VisPlay, iReasoner, and V-Zero have not released checkpoints, so we could not evaluate them on the science benchmarks (‘–’).

Panel of Figure 3 — Figure 3: Case A — a stated constraint that changes the answer. The largest ppr in the figure belongs to group 5 ( $1.33$ ), which the question’s $\tau_{F}=12$ ms condition excludes; the answer is group 4 ( $1.04$ ). A model that ignores the constraint and reports the maximum is wrong, and the Solver’s transcript shows the exclusion being made explicitly. Panels b–g of the source figure are omitted here for space. Source figure: Nature Communications , 10.1038/s41467-022-35395-y , reproduced under CC BY 4.0. OmniScience subject annotation [ 18 ] : Biology .

Panel of Figure 4 — Figure 4: Case B — cross-panel mapping under domain vocabulary. Panels d–f carry no titles of their own, so the mutant identity of each must be carried over from the trace panels a–c above them, and the question’s terms (PIP2 regulation, open probability, voltage dependence) are what make the relevant curves findable at all. Source figure: Nature Communications , 10.1038/s41467-024-52469-1 , reproduced under CC BY 4.0. OmniScience subject annotation [ 18 ] : Biology .

Panel of Figure 5 — Figure 5: Case C — normalization reverses the visible ordering. Kelly is the taller bar in both charts, so comparing bar heights gives the wrong cell line; the two charts also differ tenfold in axis range (ITCH $0$ – $1.5$ , TP73 $0$ – $15$ ), which is only apparent from the tick labels. Forming the ratio inverts the ordering ( $0.10$ for Kelly against $0.49$ for BE2). Remaining weakness: with two candidate cell lines, guessing succeeds half the time, so this episode is better read as an illustration than as evidence of reasoning on its own. Source figure: Scientific Reports , 10.1038/s41598-020-57854-6 , reproduced under CC BY 4.0. OmniScience subject annotation [ 18 ] : Medicine .

Panel of Figure 6 — Figure 6: Case D — a shared visual scale over different numerical ranges. The four boxes occupy similar heights on the page, but only MYH3 is plotted on an axis spanning $10^{-3}$ to $1.1\times 10^{6}$ while the other three span roughly $0.03$ to $8$ ; the five-order-of-magnitude difference is legible only from the tick labels. The canonical differentiation marker among these genes is MyoG, whose median sits more than five orders of magnitude below that of MYH3 once the axes are read. Source figure: Scientific Reports , 10.1038/s41598-018-30123-3 , reproduced under CC BY 4.0. OmniScience subject annotation [ 18 ] : Pharmacy .

Vision-Language Models

Vision-language models connect visual inputs with text for retrieval, question answering, generation, and grounded multimodal reasoning.

See all 734 papers

Program-Verified Self-Evolution for Vision-Language Models
Sep 27, 2026Ahmed Heakl, Sungik Choi, Moontae Lee +1Visual Question Answering Vision-Language Models
LG AI Research · MBZUAI · University of Illinois at Chicago +1
RISE: Reliable Improvement in Self-Evolving Vision-Language Models
May 20, 2026Chaoran Xu, Yingmao Miao, Pengfei Zhang +3Vision-Language Models VLM Adaptation
AMAP, Alibaba Group
When to Rethink: Learning Multi-Perspective Self-Verification for Vision-Language Models
Oct 4, 2026Ziquan Zhu, Hanruo Zhu, Si-Yuan Lu +10Self-Correction in VLMs VLM Reasoning
University of Leicester · Nanjing University of Posts and Telecommunications · University of North Carolina at Chapel Hill +5

Sep 27, 2026J/K move · Enter open · S save

Program-Verified Self-Evolution for Vision-Language Models

Ahmed Heakl, Sungik Choi, Moontae Lee, Salman Khan

LG AI Research · MBZUAI · University of Illinois at Chicago+1

Visual Question Answering Vision-Language Models

Open paper AskPDF arXiv

Self-evolving vision-language models train on questions they generate from unlabeled images. Since these questions have no gold answers, prior methods label them by majority vote over sampled answers or by a model judge. In a human evaluation, we find that 24% of majority-vote labels and 18% of model-judge labels produced during self-evolution are wrong. To address this problem, we present Verifiable QA Generation for Self-Evolving Models (VQS), which changes how the model judges answers. Instead of voting on an answer, the model parses each image into a structured record, such as a scene graph, a chart table, or a diagram graph. Fixed programs then write a question from the record and compute its answer. The model still acts as a visual checker, but it only confirms the individual facts the program reads, one short claim at a time. These claim-level checks select the parser's training targets, so the parser also improves without labels. Human raters find 94% of VQS answers correct, against 76% for majority voting. Across ten benchmarks, VQS improves Qwen3-VL by up to 3.18 points at the 2B, 4B, and 8B scales and outperforms the strongest self-evolving baseline at each. Gains keep growing over three training rounds, reaching 3.84 points at 2B. Code is released at https://github.com/ahmedheakl/VQS

Abstract

Vision-Language Models

Explore similar work

Program-Verified Self-Evolution for Vision-Language Models

RISE: Reliable Improvement in Self-Evolving Vision-Language Models

When to Rethink: Learning Multi-Perspective Self-Verification for Vision-Language Models