SelfThinker: Self-Evolving Reasoning via Reliability-Verified Intrinsic Visual Thoughts

Zhangquan Chen; Ruqi Huang

cs.CVOct 29, 2026

SelfThinker: Self-Evolving Reasoning via Reliability-Verified Intrinsic Visual Thoughts

Authors: Zhangquan Chen, Ruqi Huang

Organizations: Tsinghua Shenzhen International Graduate School, Tsinghua University, Shenzhen, 518055, Guangdong, China.

Abstract

Unified multimodal models can both generate and understand images within a single set of weights, offering an opportunity to use image synthesis as training supervision for visual comprehension. However, existing visual chain-of-thought methods rely on human-annotated helper images during training and require slow image generation at test time. We present SelfThinker, the first self-evolving framework that turns a unified model's self-generated visual thoughts into stronger understanding without human image annotations. Given a visual problem, SelfThinker samples candidate helper images from its generation pathway and uses its frozen understanding pathway to score each candidate via conditional pointwise mutual information. High-utility visual thoughts then refine the generator via flow matching and serve as privileged teacher context for on-policy self-distillation. At inference, SelfThinker reasons directly from the original image in text without generating intermediate images. Although trained only on small synthetic visual-search and navigation datasets, SelfThinker transfers broadly to natural-image and spatial benchmarks, raising the seven-benchmark average of ThinkMorph-7B from 52.7 to 64.0 (+11.3 points). Controlled image-perturbation probes further show that SelfThinker surpasses ThinkMorph-7B by up to 40.4 points and its pretrained BAGEL-7B backbone by up to 23.2 points across fine-grained and spatial perception tasks. Our code is available at https://github.com/zhangquanchen/SelfThinker.

Figures & tables

Figure 1: Overview of SelfThinker . The upper-left Self-Evolution panel summarizes the closed loop of imagining candidate visual thoughts, verifying reliable ones, and evolving into a stronger model. Imagine : a shared UMM with Generation and Understanding pathways produces candidate visual thoughts and decoded answers. Verify : the frozen Understanding pathway scores each candidate by its log-likelihood utility, calibrates the acceptance threshold against thoughts from other problems, and selects the top verified thought under the ground-truth answer or majority vote. Evolve : the verified thought is distilled into the Student Understanding pathway via on-policy token-level JSD distillation from the EMA Teacher ( SelfThinker ), supported by flow-matching generator refinement ( SelfThinker -Gen). No human image annotation is used

Method	Visual thoughts	Self-generated imagery	No image annotations	Thought verification	Text-only inference
Visual Sketchpad ( Hu et al., 2024 )	$\boldsymbol{\checkmark}$	$\boldsymbol{\times}$	$\boldsymbol{\times}$	$\boldsymbol{\times}$	$\boldsymbol{\times}$
MVoT ( Li et al., 2025 ) , ThinkMorph ( Gu et al., 2025 )	$\boldsymbol{\checkmark}$	$\boldsymbol{\checkmark}$	$\boldsymbol{\times}$	$\boldsymbol{\times}$	$\boldsymbol{\times}$
Visual-OPSD ( Li et al., 2026a ) , Vision-OPD ( Yuan et al., 2026 )	$\boldsymbol{\checkmark}$	$\boldsymbol{\times}$	$\boldsymbol{\times}$	$\boldsymbol{\times}$	$\boldsymbol{\checkmark}$
u-OPSD ( Li et al., 2026b ) , Canon ( Gkountouras et al., 2026 )	$\boldsymbol{\times}$	$\boldsymbol{\times}$	$\boldsymbol{\checkmark}$	$\boldsymbol{\times}$	$\boldsymbol{\checkmark}$
SelfThinker (ours)	$\boldsymbol{\checkmark}$	$\boldsymbol{\checkmark}$	$\boldsymbol{\checkmark}$	$\boldsymbol{\checkmark}$	$\boldsymbol{\checkmark}$

Table 1: Comparison of SelfThinker with related visual-thought and self-distillation methods across five key properties: reasoning with intermediate visual thoughts, learning from the model’s own self-generated visual thoughts rather than external tools or curated datasets, operating without human image annotations (bounding boxes, masks, or path overlays), explicitly verifying candidate visual thoughts, and supporting text-only inference at test time

Method	VSP	ChartQA	V*	MMVP	SAT	BLINK	CV-B.	Average	$\Delta$
ThinkMorph-7B ( Gu et al., 2025 )	1.2	52.7	53.9	79.0	55.3	53.6	73.4	52.7	–
SelfThinker (Nav)	2.0	78.1	63.4	80.3	58.0	60.3	78.0	60.0	$+7.3$
SelfThinker (Mix)	5.5	74.7	70.2	79.3	54.7	58.0	83.5	60.8	$+8.1$
SelfThinker (Vote)	13.0	81.9	67.0	80.0	60.0	59.3	82.4	63.4	$+10.7$
SelfThinker	15.8	80.1	70.2	82.3	60.0	58.2	81.7	64.0	$+11.3$
$\Delta$ vs. ThinkMorph-7B	$+14.6$	$+27.4$	$+16.3$	$+3.3$	$+4.7$	$+4.6$	$+8.3$	$+11.3$	–

Table 2: Main results on seven multimodal benchmarks under a unified judge (Qwen3-30B-A3B-Instruct-2507) and text-only decoding. SelfThinker variants distill verified visual thoughts from synthesized navigation ( Nav ), combined ( Mix ), or visual-search pools (under label-free rollout Vote or ground-truth answer direction) into single-image text reasoning. Bold marks the best score in each column and

\Delta

gives the gain over ThinkMorph-7B. CV-B.: CV-Bench

Figure 2: Per-benchmark and average scores of ThinkMorph-7B, SelfThinker (Nav), SelfThinker (Mix), and SelfThinker under the unified judge and text-only decoding. Numbers above the bars indicate the gains of SelfThinker over ThinkMorph-7B

Figure 3: Analysis of the distilled student. a Percentage of outputs that state a final answer before the image slot under the benchmark protocol, for ThinkMorph-7B and SelfThinker . b Image-condition probe at 512 thinking tokens for pretrained BAGEL-7B, ThinkMorph-7B, and SelfThinker across V*, MMVP, CV-Bench-2D, and CV-Bench-3D. Full bar height gives accuracy with the real problem image, and the pale base below the horizontal tick gives accuracy on its pixel scramble, so the upper saturated segment represents the visual gain

Task and direction	Yield	Top-1 AUC	Win rate (%)	Margin (nats)
Navigation, answer direction	0.674	0.821	57.0	2.10
Visual search, answer direction	0.578	0.660	54.1	0.20
Visual search, vote direction (label-free)	0.578	0.647	54.5	0.21
accurate-vote subset (93% of items)	0.583	0.654	54.9	0.23

Table 3: Reliability verification on the diagnostic splits (500 items each). Yield is the fraction of items with a verified thought, Top-1 AUC compares the top intrinsic thought with the mean of mismatched utilities, Win rate is the percentage of intrinsic thoughts that outscore their paired mismatched thought, and Margin is the mean utility difference in nats

Figure 4: What intrinsic verification measures. a Ground-truth thoughts of ThinkMorph’s training tasks and of fresh navigation maps, verified against mismatched ground-truth thoughts. b Utility

u

(nats) on fresh navigation for verified intrinsic thoughts

v^{\star}

, programmatic GT path overlays, all self-generated candidates, and mismatched thoughts from other maps. Dots show means and vertical bars show interquartile ranges

Figure 5: Verified visual thoughts on diagnostic items. Each row shows the problem, the programmatic reference thought (dashed green frame, not used by verification), and the six intrinsic visual thoughts sorted by utility $u$ (nats). The dark blue frame marks the verified top thought, and light blue frames mark other thoughts exceeding the item’s 95th-percentile mismatched threshold printed above each row. Row 1: the verified thought draws the shortest path. Row 2: the verified thoughts re-render the scene consistently. Rows 3 and 4 (visual search): marking the queried target scores highest, whereas boxes on distractor objects score lowest

Method	Configuration	Average	$\Delta$ vs. ThinkMorph-7B
ThinkMorph-7B	Base interleaved model	52.7	–
SelfThinker (Nav)	On-policy self-distillation with verified navigation thoughts	60.0	$+7.3$
SelfThinker (Mix)	On-policy self-distillation with combined navigation and visual-search thoughts	60.8	$+8.1$
Prompt-OPSD	Visual-search distillation with transition prompt only (no second image)	60.7	$+8.0$
Noise-OPSD	Visual-search distillation with uniform RGB noise in place of the thought	61.8	$+9.1$
SelfThinker (Vote)	Visual-search distillation with thoughts verified via majority voting (label-free)	63.4	$+10.7$

Table 4: Ablation of the self-evolving understanding update across training task pools, teacher contexts, and supervision directions. Average gives the seven-benchmark mean and

\Delta

gives the improvement over ThinkMorph-7B

Task	Method	Margin (nats)	Win rate (%)	Top-1 AUC	Yield
Navigation	ThinkMorph-7B	2.10	57.0	0.821	0.674
Navigation	Rand-RFT	2.06	57.3	0.825	0.659
Navigation	SelfThinker -Gen	2.96	60.0	0.847	0.694
Visual search	ThinkMorph-7B	0.20	54.1	0.660	0.578
Visual search	SelfThinker -Gen	0.27	54.4	0.675	0.632

Table 5: Ablation of the supporting generation update on the diagnostic splits (500 items per task). Each generator draws six visual thoughts per item, and the frozen ThinkMorph-7B understanding pathway scores them. SelfThinker -Gen improves every verification metric over both ThinkMorph-7B and Rand-RFT

Appendix figures & tables14 assets

Supplementary material from the paper’s appendix.

Appendix

Pool	Problem image	Question and answer	Reference thought
Navigation	Frozen-Lake grid, $4\times 4$ to $8\times 8$ , 12–28% holes, $192\times 192$ px	Move sequence along the unique shortest path (2–14 moves, mean 5.1)	Path overlay
Visual search	5–8 non-overlapping shapes (5 shapes, 8 colours), $384\times 384$ px	Colour of a target shape or shape of a target colour (13 classes)	Red bounding box

Appendix

Table 6: Synthesized problem pools built outside the training data of ThinkMorph-7B. Each pool contains 1,500 training items and 500 diagnostic items. Programmatic reference thoughts are used only as diagnostic baselines and never enter verification or model training

Setting	SelfThinker -Gen (generation update)	SelfThinker (understanding update)
Trainable pathway	Generation and understanding (shared decoder)	Understanding; generation pathway unused
Objective	Flow matching on thought latents (weight 1.0), cross-entropy on answer (0.25)	Token-level JSD, $\beta=0.5$ , temperature 1.0, top- $k$ 256, per-token clip 0.05
Teacher	–	EMA of the student, decay 0.995
Student sampling	–	Temperature 1.0, up to 128 new tokens, one image skip
Schedule	1,000 steps; cosine decay, warmup 100, learning rate $10^{-5}$ to $10^{-7}$	1,000 steps; cosine decay, warmup 200, learning rate $10^{-5}$ to $10^{-7}$
Optimizer	AdamW	AdamW $(0.9,0.95)$ , gradient clipping 1.0

Appendix

Table 7: Training hyperparameters for SelfThinker -Gen and SelfThinker . Both updates start from ThinkMorph-7B and run on one node with eight NVIDIA H20 GPUs

Method	VSP	ChartQA	V*	MMVP	SAT	BLINK	CV-B.	Average	$\Delta$
ThinkMorph-7B	1.2	52.7	53.9	79.0	55.3	53.6	73.4	52.7	–
Understanding evolution: navigation and mixed pools
SelfThinker (Nav)	2.0	78.1	63.4	80.3	58.0	60.3	78.0	60.0	$+7.3$
SelfThinker (Mix)	5.5	74.7	70.2	79.3	54.7	58.0	83.5	60.8	$+8.1$
Understanding evolution: visual-search pool and teacher controls
Prompt-OPSD	3.3	77.7	68.6	79.7	56.7	58.4	80.6	60.7	$+8.0$

Appendix

Table 8: Per-benchmark breakdown of task-pool variants, teacher-context controls, and SelfThinker under the unified judge and text-only decoding.

\Delta

gives the improvement in the seven-benchmark average relative to ThinkMorph-7B. CV-B.: CV-Bench

Figure 6: Comparison of SelfThinker variants and teacher-context controls. a Gain in the seven-benchmark average relative to ThinkMorph-7B across navigation, mixed-pool, and visual-search distillation runs. b Per-benchmark gains relative to ThinkMorph-7B for Prompt-OPSD, Noise-OPSD, and SelfThinker

Figure 7: Answer consensus versus imagery on visual search (500 diagnostic items). a Items split by answer agreement (vote share at least 0.8) and visual agreement (mean pairwise cosine similarity of the six thoughts, median split), with the vote accuracy of each cell. b Paired win rate against mismatched thoughts under the answer direction, the vote direction, and the accurate-vote subset. c Utility difference $u(\text{self})-u(\text{mismatched})$ for thoughts whose trace answered correctly versus wrongly

Figure 8: Improvements in the generation pathway under SelfThinker -Gen. a Utility margin (nats) of navigation thoughts drawn by ThinkMorph-7B, Rand-RFT, and SelfThinker -Gen on the 500-item diagnostic split. b Bounding-box generation rate and above-threshold pass rates on visual search before and after SelfThinker -Gen

Method	Own	None	Neutral	True ans.	Thought only	Own $-$ None
ThinkMorph-7B	45.9	39.7	38.0	40.7	35.4	$+6.1$
SelfThinker (Nav)	92.9	92.9	91.2	96.4	89.1	$+0.0$
SelfThinker (Vote)	95.4	95.3	95.3	99.7	93.0	$+0.1$
SelfThinker	99.0	98.9	98.9	99.5	96.5	$+0.1$

Appendix

Table 9: Context interventions on the 1,092 verified visual-search items (accuracy in %). Own: the item’s verified visual thought as second image. None: no second image (single-image deployment). Neutral: a neutral sentence appended. True ans.: the ground-truth answer appended. Thought only: the verified visual thought replaces the problem image. The last column reports the difference between Own and None, measuring single-image internalization

Figure 9: Context interventions on the 1,092 verified visual-search items across ThinkMorph-7B, SelfThinker (Nav), SelfThinker (Vote), and SelfThinker under five conditions (accuracy in %). Rows compare providing the item’s verified thought as second image, no second image (single-image deployment), appending a neutral sentence or the ground-truth answer, or replacing the problem image with the verified thought alone

ChartQA, horizontal bars (444 items, 222 with a different previous image)
Method	real	real †	wrong †	gray	scr.	$\times 2$	$\times 4$	$\times 8$	$\times 16$	$\sigma 1$	$\sigma 2$	$\sigma 4$	$\sigma 8$
ThinkMorph-7B	32.2	13.1	8.6	5.0	5.6	30.6	10.1	4.7	3.8	32.7	20.5	4.1	4.7
SelfThinker (Nav)	51.4	33.3	11.7	9.7	10.1	48.9	19.4	–	–	48.2	36.7	10.8	10.1
Prompt-OPSD	63.3	52.7	9.0	11.7	10.4	59.0	24.5	12.4	–	60.1	45.0	14.4	–
Noise-OPSD	66.2	58.1	14.4	12.6	10.8	62.4	26.6	12.6	–	67.1	46.8	18.2	14.4
SelfThinker	67.3	52.3	14.9	10.4	11.3	64.9	25.7	11.7	13.7	68.9	52.3	18.2	12.8

Appendix

Table 10: Image-condition probe on ChartQA and V*. Accuracy (%) under each condition of the problem image (256 thinking tokens). real † and wrong † : real image and another item’s image on the subset whose previous row shows a different image. scr.: pixels in random order.

\times k

: mosaic with

k

-pixel blocks.

\sigma k

: Gaussian blur with

\sigma=k

pixels. Dashes: not run

Figure 10: Conditions of the image-condition probe, illustrated on a ChartQA horizontal-bar item and a CV-Bench-2D item. The problem image is replaced by another item’s image, a gray canvas, its pixels in random order, a $2\times 2$ or $4\times 4$ tile permutation, a mosaic with square blocks of 4 or 16 pixels, or a Gaussian blur with $\sigma$ of 2 or 8 pixels at the native resolution shown under each benchmark name

Figure 11: Image-condition probe at 256 thinking tokens across ThinkMorph-7B, SelfThinker (Nav), Prompt-OPSD, Noise-OPSD, and SelfThinker . a Accuracy with the real problem image. b – d Visual sensitivity (accuracy gain in percentage points of the real problem image over pixel scramble, another item’s distinct image, and Gaussian blur with $\sigma=8$ pixels)

MMVP (300 items)
Method	real	real †	wrong †	gray	scr.	$\times 2$	$\times 4$	$\times 8$	$\times 16$	$\sigma 1$	$\sigma 2$	$\sigma 4$	$\sigma 8$
ThinkMorph-7B	44.7	44.7	47.7	49.0	45.3	45.3	47.3	49.7	47.0	47.3	51.7	46.3	50.0
SelfThinker (Nav)	69.3	69.3	40.0	48.3	45.0	69.7	61.7	53.0	49.7	67.7	68.7	63.3	57.3
Prompt-OPSD	67.0	67.0	36.7	49.7	49.0	66.7	61.0	54.0	–	68.0	65.0	62.3	56.7
Noise-OPSD	73.3	73.3	29.0	49.3	48.7	70.3	61.3	56.7	–	72.7	69.3	62.7	54.7
SelfThinker	78.0	78.0	26.7	48.7	46.3	74.0	62.7	53.3	52.7	79.7	72.3	63.0	62.7

Appendix

Table 11: Image-condition probe on MMVP and CV-Bench. Accuracy (%) under each condition of the problem image (256 thinking tokens). real † and wrong † : real image and another item’s image on the subset whose previous row shows a different image. scr.: pixels in random order.

\times k

: mosaic with

k

-pixel blocks.

\sigma k

: Gaussian blur with

\sigma=k

pixels. Dashes: not run

V* (191 items)
Method	real	scr.	$\times 2$	$\times 4$	$\times 8$	$\times 16$	$\sigma 2$	$\sigma 4$	$\sigma 8$	$2\times 2$	$4\times 4$	$8\times 8$	$16\times 16$
BAGEL-7B	48.7	37.7	48.2	–	42.4	40.3	52.4	44.5	42.4	43.5	44.5	43.5	40.8
ThinkMorph-7B	34.0	36.6	35.6	–	34.6	36.1	36.1	35.1	35.1	37.2	36.1	37.7	36.1
SelfThinker (Nav)	56.0	38.2	57.6	–	–	38.2	56.5	53.9	–	52.9	54.5	50.8	44.0
SelfThinker	67.0	34.0	66.5	–	61.8	46.1	66.5	56.5	46.1	57.1	57.1	53.9	47.6
MMVP (300 items)

Appendix

Table 12: Image-condition probe at 512 thinking tokens across mosaic, Gaussian blur, and tile-permutation ladders, including the pretrained BAGEL-7B backbone. scr.: pixels in random order.

\times k

: mosaic with

k

-pixel blocks.

\sigma k

: Gaussian blur with

\sigma=k

pixels.

g\times g

: permutation of a

g\times g

grid of intact image tiles. Dashes: not run

Figure 12: Accuracy across mosaic block sizes (top), Gaussian blur

\sigma

(middle), and

g\times g

tile permutations (bottom) on V*, MMVP, CV-Bench-2D, and CV-Bench-3D for BAGEL-7B (512 tokens), ThinkMorph-7B, SelfThinker (Nav), and SelfThinker . Squares on the right mark the pixel-scramble baseline

Reasoning for VLMs

See all 64 papers

Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards
Jun 25, 2026Ritesh Thawkar, Shravan Venkatraman, Omkar Thawakar +5Unified Multimodal Models Multimodal Generation
1Mohamed bin Zayed University of Artificial Intelligence · 4Linköping University · 3Australian National University +1
Thinking Without Images: Internalizing Visual Manipulation with On-Policy Self-Distillation
Jun 7, 2026Yishuo Cai, Jiahui Liu, Yuanxin Liu +9Tool Use in VLMs Efficient VLM Inference
1State Key Laboratory of Multimedia Information Processing, School of Computer Science, Peking University · 2Central South University · University of Science and Technology of China +2
Large Language Models are Universal Reasoners for Visual Generation
May 5, 2026Sucheng Ren, Chen Chen, Zhenbang Wang +5Diffusion Model Guidance Image Generation
1Johns Hopkins University · 2Apple

Jun 25, 2026J/K move · Enter open · S save

Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards

Ritesh Thawkar, Shravan Venkatraman, Omkar Thawakar, Abdelrahman Shaker+4

1Mohamed bin Zayed University of Artificial Intelligence · 4Linköping University · 3Australian National University+1

Unified Multimodal Models Multimodal Generation Multimodal Foundation Models Image Generation Evaluation Multimodal Learning Multimodal Understanding

Open paper AskPDF arXiv

Most unified large multimodal models (LMMs) that support both visual understanding and image generation still rely on curated post-training supervision, such as human annotations, preference labels, or external reward models. We ask whether a unified LMM can improve both abilities autonomously using only unlabeled images. We propose a self-evolving training framework with three internal roles: a Proposer that generates visual questions, a Solver that answers and evaluates them, and a Generator that synthesizes images. Training uses only self-derived consistency signals, without human annotations, preference labels, or task-trained external reward/judge models. To stabilize learning, we introduce Solver Token Entropy (STE), a continuous difficulty signal based on token-level prediction uncertainty that remains useful even when sample-level consistency becomes unreliable. For image generation, we design a multi-scale internal evaluation scheme that combines question-answer fidelity scoring with cycle-consistent captioning. This creates a solver-mediated coupling, where better visual understanding enables more reliable generation assessment and stronger internal training signals. The framework preserves the same role decomposition, reward logic, and training schedule across diffusion-based BLIP3o, rectified-flow BAGEL, and autoregressive VARGPT-v1.1 architectures, requiring only each backbone's native prompting and generation interface. Across eight understanding metrics, our method consistently improves over the corresponding base models. On BAGEL, it achieves a

+3.5\%

absolute gain on MMMU and improves GenEval image generation performance from

82\%

85\%

. Code and models are publicly released.

Abstract

Reasoning for VLMs

Explore similar work

Ask, Solve, Generate: Self-Evolving Unified Multimodal Understanding and Generation via Self-Consistency Rewards

Thinking Without Images: Internalizing Visual Manipulation with On-Policy Self-Distillation

Large Language Models are Universal Reasoners for Visual Generation