Program-Verified Self-Evolution for Vision-Language Models

Ahmed Heakl; Sungik Choi; Moontae Lee; Salman Khan

cs.CVSep 27, 2026

Program-Verified Self-Evolution for Vision-Language Models

Authors: Ahmed Heakl, Sungik Choi, Moontae Lee, Salman Khan

Organizations: LG AI Research · MBZUAI · University of Illinois at Chicago · Australian National University

Abstract

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

Figures & tables

Figure 1: VQS improves every benchmark. Scores are for Qwen3-VL-2B. VISE loses on LogicVista and MMStar.

Figure 2: Majority-vote labels vs. computed answers. (a) In prior self-play, the solver’s majority answer becomes the label. Here three of five answers say 2 mugs, so the wrong answer is rewarded and the correct answer “1” is discarded. (b) VQS computes the answer instead. (1) A schema-constrained parser turns the image into a structured parse. A fixed template reads the parse (filter mugs, then count) to write a question and compute its answer “1” and the parser then paraphrases the question. (2) A visual checker verifies each claim behind the answer, a blind gate drops questions answerable without the image, and a difficulty band keeps those the solver gets right in some but not all of 8 rollouts. (3) GRPO rewards each rollout by exact match against the computed answer, so training reinforces the correct answer “1” and suppresses the wrong one “2” .

Figure 3: Label-free parser training. (1) The parser samples $K$ structured parses of an unlabeled image under a schema-constrained decoder. (2) A visual checker verifies each parse one claim at a time. In the example, $z_{2}$ wrongly calls the mug blue and places it to the right of the book, so it passes 6 of 8 claims. (3) Each parse is scored by its fraction of verified claims. We keep the best parse $z_{1}$ (8/8) only because it beats the mean of all four parses (0.75) and contains at least $m{=}4$ claims. (4) The kept parses train the parser with SFT. The trained parser then feeds the templates that write questions and compute their answers for the solver.

Figure 4: Examples of structured scene parsing. The parser converts natural images, charts, diagrams, and infographics into domain-specific JSON records that preserve spatial relations, chart values and ticks, graph structure, and verbatim strings for deterministic QA construction.

Qwen3-VL-2B-Instruct
Method	GQA	OK-VQA	InfoVQA	SQA	MMMU	MMB	ESB	LogicV	MMStar	SEED	Avg.
Base	58.25	40.76	69.02	79.42	38.92	74.48	68.54	35.04	55.62	72.16	59.22
VisPlay	58.65 +0.40	41.15 +0.39	69.96 +0.94	80.74 +1.32	39.27 +0.35	74.52 +0.04	68.56 +0.02	34.93 -0.11	55.37 -0.25	71.62 -0.54	59.48 +0.26
Vision-Zero	58.98 +0.73	41.39 +0.63	70.93 +1.91	81.96 +2.54	39.58 +0.66	75.07 +0.59	69.72 +1.18	35.28 +0.24	55.58 -0.04	71.53 -0.63	60.00 +0.78
EvoLMM	59.01 +0.76	38.03 -2.73	70.69 +1.67	83.01 +3.59	39.08 +0.16	74.62 +0.14	69.32 +0.78	34.99 -0.05	55.50 -0.12	71.17 -0.99	59.54 +0.32
iReasoner	59.13 +0.88	38.13 -2.63	70.82 +1.80	83.12 +3.70	39.11 +0.19	74.75 +0.27	69.67 +1.13	35.09 +0.05	55.59 -0.03	71.25 -0.91	59.67 +0.45

Table 1: Main results across three scales. Every model trains on unlabeled images only, with no human question, answer or gold label at any stage. Subscripts give the change against that backbone’s own base, in green when positive and red when negative. VQS gives the largest average gain at all three scales , +3.18 at 2B, +2.32 at 4B and +2.43 at 8B, against +1.01, +1.04 and +0.66 for the strongest published method. InfoVQA stands for infographicsVQA, SQA for ScienceQA, MMB for MMBench, ESB for EmbSpatial, and LogicV for LogicVista.

Figure 5: What scale and coverage change. (a) Share of benchmark answers RL changes, measured per sample against each model’s base. (b) Smoothed training reward. (c) 2B gain over base against the number of template families.

Answer source	Correct (%)	Kept (%)
Templates + checker	94.4%	76.0
Majority vote	76.4%	92.0
Model judge	82.2%	90.8

Table 2: Human evaluation. Correct is accuracy among retained questions.

Parser target	Parse Prec.	Answer Acc.	Solver Acc.
Untrained	80.8	86.8	58.0
Random	82.1	91.0	61.5
Whole parse	85.0	94.0	62.0
Per claim	87.6	96.1	62.4

Table 3: Parser target selection. Parse precision, correct answer accuracy, and solver accuracy (%).

Figure 6: Parser Structured-F1 Score.

Figure 7: Parser target selection: quality and coverage. Each panel varies one setting around

(K,\delta,m)=(4,0.10,4)

. Parse precision is the percentage of correct visual claims after training. Images retained is the percentage of source images yielding a selected training target.

Figure 8: Gains across cycles. Average gain over base per benchmark category for Qwen3-VL-2B.

Pipeline	Solver Acc.	Parser Prec.
Untrained parser	58.0	80.8
Full VQS	62.4	87.6
✗ constrained decoding	61.9	85.3
✗ fact-checker	60.9	87.6
✗ blind gate	61.8	87.6
✗ difficulty band	62.1	87.6

Table 4: Component ablations at 2B.

Rank	Family	Gain
$\blacktriangle$ 1	Chart differences	7.4
$\blacktriangle$ 2	Diagram connectivity	5.8
$\blacktriangledown$ 1	Object existence	-0.2
$\blacktriangledown$ 2	Attribute lookup	0.4

Table 5: Per-family gains. The two largest and two smallest.

Comparison	Setting	Solver acc.
Family difficulty	Easy	60.8
	Middle	62.4
	Hard	61.1
Generator	Fixed templates	62.4
	Generated programs	61.4
Curriculum	Shuffled	61.3

Table 6: Question design at 2B. Each block changes one choice at an equal training budget. Solver accuracy averages the 10 benchmarks in Table 1 . Bold marks the VQS default.

Appendix figures & tables5 assets

Supplementary material from the paper’s appendix.

Appendix

Adapter
Setting	Value
LoRA rank, alpha	64, 128
Target modules	all linear layers, .visual. excluded
Trainable visual parameters	0.00M
Gradient checkpointing	on
Optimiser

Appendix

Table 7: Solver training settings. Identical across 2B, 4B, and 8B except for the per-device micro-batch.

Category	Benchmark	What it tests
Perception	GQA	objects, attributes, and relations
	InfoVQA	text and layout in infographics
	EmbSpatial (ESB)	spatial relations in embodied scenes
Knowledge & Reasoning	OK-VQA	outside knowledge
	ScienceQA (SQA)	school science
	MMMU	college-level subjects

Appendix

Table 8: Benchmark categories. Each benchmark belongs to one category.

Checker	Relation to the parser	Accepts (%)	Trap YES (%)	$\bm{\kappa}$ vs. ours
Qwen3-VL-2B (ours)	same weights	94.4	1.2	–
Qwen3-VL-8B	same family, 4 $\times$ larger	93.3	1.6	0.837
InternVL3.5-8B	different family, 8B	93.1	4.0	0.810
Qwen2.5-VL-7B	previous Qwen generation	89.5	0.4	0.807
SmolVLM2-2.2B	different family, same size	86.1	5.7	0.660

Appendix

Table 9: Checker independence. All five checkers grade the same 1,417 claims. Accepts is the share of the 1,170 parser claims a checker calls true. Trap YES is the share of the 247 corrupted claims it wrongly accepts.

\kappa

is Cohen’s kappa with our checker over all claims.

Gemma3-12B-It
Method	GQA	OK-VQA	InfoVQA	SQA	MMMU	MMB	ESB	LogicV	MMStar	SEED	Avg.
Base	61.20	62.40	52.10	94.30	46.50	76.40	24.50	41.80	58.90	74.20	59.23
VQS (ours)	62.02 +0.82	62.71 +0.31	55.29 +3.19	95.23 +0.93	50.84 +4.34	79.06 +2.66	27.70 +3.20	42.87 +1.07	59.80 +0.90	74.47 +0.27	61.00 +1.77
InternVL3-8B-Instruct
Base	65.40	64.10	68.80	96.50	51.20	81.30	28.10	48.60	63.70	76.90	64.46
VQS (ours)	66.01 +0.61	64.36 +0.26	71.81 +3.01	96.71 +0.21	54.87 +3.67	83.17 +1.87	31.99 +3.89	49.42 +0.82	64.32 +0.62	77.08 +0.18	65.97 +1.51

Appendix

Table 10: Results on other model families. Every model trains on unlabeled images only. Subscripts give the change against that backbone’s base. Abbreviations follow Table 1 . VQS improves all ten benchmarks for every family , raising the average by +1.77 on Gemma3-12B-It, +1.51 on InternVL3-8B-Instruct and +1.84 on Llama-3.2-11B-Vision-Instruct.

Family	Lv	Template $\rightarrow$ example question	Answer key
Photos
existence	L2	Are there any $\langle$ objects $\rangle$ in the image? $\rightarrow$ Are there any windows in the image?	no
relation_existence	L2	Are there any $\langle$ objects $\rangle$ $\langle$ relation $\rangle$ the $\langle$ anchor $\rangle$ ? $\rightarrow$ Are there any chairs to the right of the painting?	yes
attribute_filtered_existence	L3	Is there a $\langle$ attribute $\rangle$ $\langle$ object $\rangle$ in the picture? $\rightarrow$ Is there white grass in the picture?	no
state_filtered_existence	L3	Do you see a $\langle$ state $\rangle$ $\langle$ object $\rangle$ in the picture? $\rightarrow$ Do you see an open oven door in the picture?	yes
gated_object_recognition	L3	Which kind of $\langle$ category $\rangle$ is $\langle$ attribute $\rangle$ ? $\rightarrow$ Which kind of furniture is green?	bench

Appendix

Table 11: Template families, with one question each. A slot in

\langle

angle brackets

\rangle

is filled from the parse, and a part in [ square brackets ] appears only when the chart has several series. Each example is a question the family wrote for a real image, shown after paraphrasing where the paraphrase passed the guard, with the key its program computed. We checked every key against its image.

Explore similar work

Self-Evolving Visual Questioner
Jun 11, 2026Yijun Liang, Hengguang Zhou, Ming Li +3Faithful Question Generation Question
University of Maryland, College Park · University of California, Los Angeles · Peking University +2
RISE: Reliable Improvement in Self-Evolving Vision-Language Models
May 20, 2026Chaoran Xu, Yingmao Miao, Pengfei Zhang +3Recent Vision-Language Models Multimodal Reasoning
AMAP, Alibaba Group
AutoVQA-G: Self-Improving Agentic Framework for Automated Visual Question Answering and Grounding Annotation
Apr 19, 2026Rongsheng Hu, Runwei Guan, Yicheng Di +2Knowledge-Based Visual Question Answering Textvqa
School of Artificial Intelligence and Computer Science, Jiangnan University, Wuxi, China · Thrust of Artificial Intelligence, HKUST(GZ), Guangzhou, China

Jun 11, 2026, cs.CVJ/K move · Enter open · S save

Self-Evolving Visual Questioner

Yijun Liang, Hengguang Zhou, Ming Li, Lichen Li+2

University of Maryland, College Park · University of California, Los Angeles · Peking University+2

Faithful Question Generation Question

Open paper AskPDF arXiv

Vision-language models (VLMs) are typically trained as passive answerers, while their ability to actively ask diverse, non-trivial, visual-centric and grounded questions remains underexplored. Existing visual questioners' performance is bottlenecked by the availability of high-quality training data or the cost of curating them. We show that a VLM can continuously improve itself as a visual questioner without any external supervision. We propose a self-evolving framework that uses a VLM itself as both a proposer and a filter to produce harder, more informative, and visual-centric questions, while maintaining their exploration diversity to avoid training collapse. These questions are then used to train the VLM in both questioner and answerer modes. To evaluate the questioner, we introduce an agentic protocol that assesses questions along perception, reasoning, and diversity dimensions. Experiments across various backbone VLMs show that our method substantially enhances the quality and substantially expands the difficulty boundary of autonomous question generation. Under the same budget, our self-supervision is more effective than training on the static source data. Moreover, the self-evolving questioner remains a competitive or even better answerer.

Abstract

Explore similar work

Self-Evolving Visual Questioner

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

AutoVQA-G: Self-Improving Agentic Framework for Automated Visual Question Answering and Grounding Annotation