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

Introduction

Visual understanding and spatial reasoning are foundational capabilities for multimodal large language models (Tong et al. 2024c,a; Fu et al. 2024). When tackling multi-step visual tasks (e.g., grid navigation, visual search), text-only chains of thought often fall short because natural language struggles to capture fine-grained spatial geometry. Humans naturally overcome this bottleneck by forming intermediate mental images to guide deduction [45, 29, 40, 8, 10]. Motivated by this cognitive process, recent unified multimodal models (UMMs) integrate image generation and visual understanding within a single set of weights [3, 60, 56, 55, 49, 14] (Zhou et al. 2024), enabling interleaved reasoning models to draw intermediate helper images, or visual thoughts, before predicting a final answer [32, 11, 20, 7 d].

In principle, because a unified model can already synthesize images, its generation pathway could serve as a self-contained source of visual supervision to strengthen its understanding pathway. In practice, however, existing visual-thought pipelines such as Mirage [62], ThinkMorph [20], and Visual-OPSD [33] remain tied to external visual supervision, relying on human-designed helper images rendered by external tools (such as shortest-path overlays for navigation or bounding boxes for visual search). Constructing such task-specific templates requires costly manual engineering and confines visual imagination to predefined heuristics, preventing the model from learning from its own generated imagery at scale.

Overcoming this limitation and turning a model’s image generation capability into effective training supervision for visual understanding requires addressing three core challenges: (C1)

Annotation-free learning, replacing human-engineered visual templates with diverse candidate visual thoughts sampled directly from the model’s own generation pathway, without ground-truth bounding boxes, masks, or path overlays; (C2) Intrinsic verification, determining which self-generated images genuinely aid comprehension, since generated candidates are often noisy and standard majority voting over text answers [54, 27, 34, 19] (Zuo et al. 2025) fails when agreeing answers accompany uninformative or corrupted intermediate images [33]; and (C3) Test-time efficiency, internalizing the visual cues of helpful intermediate images back into the understanding pathway, so that deployment neither incurs slow image generation latency nor halts prematurely at the image token.

To address (C1)–(C3), we propose SelfThinker (Fig. 1), a self-evolving framework structured into three stages: Imagine, Verify, and Evolve. In the Imagine stage, given a visual problem, the model samples multiple candidate visual thoughts from its generation pathway and decodes their corresponding answers, supplying diverse visual candidates without human image annotations (resolving C1). In the Verify stage, the model’s frozen understanding pathway acts as an intrinsic verifier, scoring each candidate image by its conditional pointwise mutual information with a target answer direction and calibrating the score against thoughts drawn for other problems to filter out conflicting imagery (resolving C2). In the Evolve stage, the verified visual thoughts serve as privileged teacher context for on-policy self-distillation (SelfThinker), transferring their visual cues into the understanding pathway so the student model reasons directly from the original image in text at test time (resolving C3), while a supporting flow-matching update (SelfThinker-Gen) refines the generator to produce cleaner candidate imagery.

Beyond eliminating human image annotations and test-time drawing overhead, our experiments also reveal a striking property, i.e., self-generated visual supervision transfers broadly across tasks. Although SelfThinker is trained only on small synthetic visual-search and grid-navigation pools, its improvements generalize across seven diverse multimodal benchmarks, spanning fine-grained chart reading and visual search, multi-step spatial planning, and natural-scene perception. Specifically, SelfThinker raises the seven-benchmark average of ThinkMorph-7B from 52.7 to 64.0 (+11.3 points) under ground-truth answer supervision, and to 63.4 (+10.7 points) under label-free majority voting. Controlled image-perturbation probes further confirm that this cross-task transfer stems from stronger visual perception rather than output formatting or language priors. That is, SelfThinker consistently outperforms teacher-context controls that lack verified imagery, surpassing ThinkMorph-7B by up to 40.4 points and its pretrained BAGEL-7B-MoT backbone by 12.0 to 23.2 points across fine-grained and spatial benchmarks.

Our contributions can be summarized as follows.

• We propose SelfThinker, the first self-evolving framework that turns a unified multimodal model’s image generation capability into training supervision for visual understanding without human image annotations.

• We introduce an intrinsic visual verification mechanism based on calibrated conditional pointwise mutual information, enabling the understanding pathway to select informative self-generated helper images.

• We develop a privileged on-policy self-distillation scheme, supported by flow-matching generator refinement, that internalizes verified visual thoughts into the understanding pathway for fast text-only inference without test-time image generation.

• Extensive experiments across multiple benchmarks and controlled image-perturbation probes demonstrate that SelfThinker consistently outperforms strong baselines and its pretrained backbone. Furthermore, our results indicate that SelfThinker generalizes effectively from synthetic training pools to natural-image and spatial perception, highlighting its broad applicability.

Related Work

Unified Multimodal Models

Unified multimodal models (UMMs) integrate visual understanding and image generation within a single architecture. Early-fusion autoregressive models such as Chameleon ([3]) and Emu3 ([55]) represent both modalities as discrete token streams. In contrast, Transfusion (Zhou et al. 2024) and Show-o ([60]) combine next-token prediction with continuous diffusion or discrete denoising. Janus and Janus-Pro decouple the visual encoders used for reading and generation ([56, 5]), whereas MetaMorph aligns an understanding backbone with visual generation via instruction tuning ([49]). BAGEL routes understanding and generation tokens through specialized transformer experts pretrained on interleaved multimodal data ([14]). At the evaluation level, standard multimodal benchmarks often suffer from language shortcuts where questions can be answered without inspecting the image ([4]). This limitation has motivated vision-centric benchmarks such as MMVP, CV-Bench, and BlindTest (Tong et al. 2024c,a; Rahmanzadeh-heravi et al. 2024), alongside image corruption robustness protocols ([24]). SelfThinker builds on a unified architecture to convert its generative capability into stronger visual perception across fine-grained and spatial benchmarks.

Reasoning with Helper Images

To move beyond text-only chains of thought, recent multimodal reasoning methods construct intermediate helper images that supply explicit visual cues during deduction. Tool-augmented systems such as Visual Sketchpad and Whiteboard-of-Thought invoke external code and drawing APIs to render auxiliary lines, boxes, and diagrams [26, 39], while region-focusing methods crop or zoom into informative sub-regions [44, 17] (Zheng et al. 2025). Mirage synthesizes task-specific helper images, such as path-annotated maps, and compresses them into latent visual tokens during training [62]. Within native UMMs, MVoT generates step-by-step spatial visualizations [32], and Chern et al. (2025) sample and critique intermediate images at test time. ThinkMorph fine-tunes BAGEL on interleaved traces across human-annotated helper-image templates, including red bounding boxes for visual search, shortest-path overlays for navigation, and region highlights for chart reading [20]. However, these approaches rely on external rendering tools or human annotation pipelines to prescribe task-specific helper images. In contrast, SelfThinker uses the model’s own generation pathway as a source of visual training supervision without human image annotations, selects informative visual thoughts via the understanding pathway, and distills them into single-image text reasoning.

Self-Evolving Training

Self-evolving training frameworks iteratively improve model capability without human rationales by filtering self-generated trajectories [21, 15] (Zelikman et al. 2022; Yuan et al. 2023) or optimizing against outcome rewards and self-play [6, 13]. Related work also evaluates whether intermediate steps genuinely contribute to predictions via pointwise mutual information (PMI) and counterfactual interventions [12, 28, 31, 52]. A complementary paradigm is learning using privileged information

[53, 25, 37], which supplies a teacher with training signals that are unavailable to the student at test time. Combining privileged context with on-policy context distillation [2, 46, 1] allows a student policy to internalize rich teacher conditioning along its own sampled trajectories. Recent methods instantiate this teacher context with annotated helper images in Visual-OPSD [33], image crops in Vision-OPD [63], annealed contexts in ATESD [22], and hindsight skills in SEED [57].

When ground-truth labels are unavailable during self-training, majority voting across rollouts provides a label-free supervisory signal [54, 27] (Zuo et al. 2025). Recent self-distillation methods such as u-OPSD [34] and Canon [19] condition the teacher on a consensus text solution, while UniGame pits generation against understanding in an adversarial loop [47]. As summarized in Table 1, prior training pipelines either require human-annotated helper images and external crops (ThinkMorph, Visual-OPSD, Vision-OPD) or distill text-only consensus solutions without visual thoughts (u-OPSD, Canon). SelfThinker bridges this gap through a unified self-evolving training loop. Rather than distilling dataset-provided visual annotations, we use conditional PMI against mismatched thoughts to verify helper images synthesized by the model itself, and distill the verified imagery into fast text-only inference.

Method	Visual thoughts	Self-generated imagery	No image annotations	Thought verification	Text-only inference
Visual Sketchpad ([26])	✓	×	×	×	×
MVoT ([32]), ThinkMorph ([20])	✓	✓	×	×	×
Visual-OPSD ([33]), Vision-OPD ([63])	✓	×	×	×	✓
u-OPSD ([34]), Canon ([19])	×	×	✓	×	✓
SelfThinker (ours)	✓	✓	✓	✓	✓

Methodology

Problem Formulation

A unified multimodal model (UMM) $\pi_\theta= (\pi_\theta^U,\pi_\theta^G)$ shares parameters across an autoregressive understanding pathway $\pi_\theta^U$, which encodes images and generates text tokens, and a continuous generation pathway $\pi_\theta^G$, which synthesizes images in latent space [14, 20]. Given a visual reasoning problem $x = (I,q)$ consisting of an input image $I$ and a query $q$, the model operates in the Imagine stage by sampling

$K$ interleaved trajectories $\{\tau^{(k)}\}_{k=1}^{K}$:

Here $r_{\mathrm{pre}}^{(k)} \sim\pi_{\theta}^{U}(\cdot\mid I,q)$ is the initial text thought before the image token <image_start>, $v^{(k)} \sim\pi_{\theta}^{G}(\cdot\mid I,q,r_{\mathrm{pre}}^{(k)})$ is a self-generated helper image (an intrinsic visual thought), and $(r_{\mathrm{post}}^{(k)},a^{(k)}) \sim\pi_{\theta}^{U}(\cdot\mid I,q,r_{\mathrm{pre}}^{(k)},v^{(k)})$ is the text continuation ending in a decoded answer $a^{(k)}$.

Directly training on self-generated trajectories faces two obstacles. First, a correct text answer $a^{(k)} = a^\star$ can arise despite an uninformative or corrupted helper image $v^{(k)}$ ([33]). Second, running image generation $\pi_{\theta}^{G}$ at test time incurs substantial latency, whereas disabling it causes premature termination at <image_start>. As illustrated in Figure 1 and Algorithm 1, SelfThinker addresses both issues through two subsequent stages. In the Verify stage (Sect. 3.2), the understanding pathway scores the task utility of each candidate helper image $v^{(k)}$ without human image annotations. In the Evolve stage, the model internalizes the verified visual context into the understanding pathway via on-policy self-distillation (Sect. 3.4), supported by flow-matching refinement of the generation pathway (Sect. 3.3).

Intrinsic Visual Verification

Directed utility score.

To evaluate whether a candidate helper image $v$ provides useful visual evidence for solving $(I,q)$, we use the frozen understanding pathway $\pi_{\theta}^{U}$ as an intrinsic verifier. Let $d=(d_1,\ldots,d_L)$ be a token sequence encoding a target answer direction. We define the directed utility of $v$ as the log-likelihood gain of generating $d$ when $v$ is revealed relative to the plain problem context:

where $\log p_{\theta}(d \mid C)=\sum_{\ell=1}^{L}\log\pi_{\theta}^{U}(d_{\ell}\mid C,d_{<\ell})$. Here $C(I,q)=[I,q]$ is the single-image problem context, and $C(I,q,v)=[I,q,\rho,v]$ appends a neutral context header $\rho$, the helper image $v$, and a fixed transition instruction $\tau$ (prompt templates in Appendix § 6). Measured in nats, $u(v)$ equals the conditional pointwise mutual information $\operatorname{PMI}_{\theta}(v;d\mid I,q)$ ([12]) and requires only two forward passes without gradients.

Equation (2) has three desirable properties for verification without human image annotations. First, the score is signed, rewarding helper images that increase the probability of $d$ ($u(v)>0$) and penalizing conflicting imagery ($u(v)<0$). Second, omitting the rollout text $(r_{\mathrm{pre}}^{(k)},r_{\mathrm{post}}^{(k)})$ from $C(I,q,v)$ isolates the visual contribution of $v$; because all candidates for item $i$ share the same header $\rho$ and instruction $\tau$, any prompt offset cancels in within-item comparisons $u_i(v_i^{(k)})-u_i(v_i^{(k')})$. Third, the score requires no ground-truth helper images or handcrafted pixel rules.

Within-item calibration.

Because $\log p_{\theta}(d\mid\cdot)$ scales with the token length $L$ of $d$, we calibrate utilities within each item $i$ against mismatched thoughts. Specifically, we pair each candidate $v_i^{(k)}$ with a mismatched thought $v_{\mathrm{mis}}^{(k)}=v_{i+\delta}^{(k)}$ and $N$ generated by the same model for a different problem, and evaluate $s_{ik}=u_i(v_{\mathrm{mis}}^{(k)})$ under item $i$’s context and direction. A mismatched thought matches a genuine candidate in style, resolution, and token count while depicting an unrelated scene, which controls for

Algorithm 1 SelfThinker: Self-Evolving Training
Require: Model $\pi_\theta = (\pi_\theta^G,\pi_\theta^U)$, problem pool $\mathcal{D} = \{(I_i,q_i)\}_{i=1}^N$, rollouts $K$, quantile $\alpha = 0.95$
Ensure: Updated generator SelfThinker-Gen $(\pi_\theta^G)$ and distilled student SelfThinker $(\pi_\theta^U)$
1: $\mathcal{D}_{\mathrm{ver}} \leftarrow \varnothing$
2: for $i = 1,\ldots,N$ do ▷ Stages 1–2: Imagine & Verify
3: Sample traces $\{(r_{\mathrm{pre},i}^{(k)},v_{\mathrm{post},i}^{(k)},a_i^{(k)})\}_{k=1}^K \sim \pi_\theta(\cdot\mid I_i,q_i)$
4: Set $d_i \leftarrow a_i^\star$ (answer mode) or $\tilde{a}_i = \operatorname{mode}(\{a_i^{(k)}\}_{k=1}^K)$ (vote mode)
5: for $k = 1,\ldots,K$ do
6: $u_{ik} \leftarrow u_i(v_i^{(k)}),\quad s_{ik} \leftarrow u_i(v_{\mathrm{mis}}^{(k)})$ ▷ {xref}`Eq. (2)`
7: end for
8: $k^\star \leftarrow \arg\max_k u_{ik},\quad Q_i \leftarrow s_{i,(\lfloor\alpha(K-1)\rfloor+1)}$
9: if $u_{ik^\star} > Q_i$ then
10: $\mathcal{D}_{\mathrm{ver}} \leftarrow \mathcal{D}_{\mathrm{ver}} \cup \{(I_i,q_i,v_i^{(k^\star)},d_i)\}$ ▷ {xref}`Eq. (3)`
11: end if
12: end for
13: for minibatch $\mathcal{B} \subset \mathcal{D}_{\mathrm{ver}}$ do ▷ Stage 3: Evolve
14: SelfThinker-Gen: update $\pi_\theta^G$ on $[I,q,v^\star,d]$ via $\nabla_\theta\mathcal{L}_{\mathrm{Gen}}(\theta)$ ▷ {xref}`Eq. (4)`
15: SelfThinker: sample $y \sim \pi_\theta^U(\cdot\mid c_S)$, update $\pi_\theta^U$ via $\nabla_\theta\mathcal{L}_{\mathrm{Und}}(\theta)$, update EMA $\bar{\theta}$ ▷ {xref}`Eq. (6)`
16: end for

baseline conditioning bias. Sorting $\{s_{i1},\ldots,s_{iK}\}$ as $s_{i,(1)} \leq\cdots\leq s_{i,(K)}$ defines an item-adaptive order-statistic threshold $Q_i = s_{i,(\lfloor\alpha(K-1)\rfloor+1)}$ ($\alpha= 0.95$; Appendix § 6). The top candidate is verified whenever it exceeds $Q_i$:

Target directions. We instantiate $d_i$ in two settings, neither of which uses human helper-image annotations. The answer direction sets $d_i = a_i^\star$ using the ground-truth text answer. The label-free vote direction sets $d_i = \tilde{a}_i = \operatorname{mode}(\{a_i^{(k)}\}_{k=1}^K)$ using the rollout majority vote, measuring how strongly each candidate image visually supports the consensus answer without requiring image or answer labels.

Supporting Generative Refinement

Given the verified dataset $\mathcal{D}_{\mathrm{ver}} = \{(I_i,q_i,v_i^\star,d_i)\}$, we can refine the generation pathway $\pi_\theta^G$ via rectified flow matching (Lipman et al. [35]; Liu et al. [36]; Deng et al. [14]) (SelfThinker-Gen) to improve the quality of candidate visual thoughts. Let $z_1 = E_{\mathrm{VAE}}(v^\star)$ denote the clean latent tokens of the verified helper image encoded by the variational autoencoder (VAE), $z_0 \sim\mathcal{N}(0,I)$ Gaussian noise, and $z_t = (1-t)z_0 + tz_1$ their linear interpolation at $t \sim\mathcal{U}(0,1)$. The generation expert $v_\theta^G$ regresses the target velocity $z_1-z_0$ jointly with a cross-entropy loss $\mathcal{L}_{\mathrm{CE}}(\theta)$ on the answer tokens:

with $\lambda_{\mathrm{FM}} = 1.0$ and $\lambda_{\mathrm{CE}} = 0.25$ (Appendix § 6). Filtering training targets by Eq. (3) shifts the generative prior toward problem-consistent trajectories and localized target boxes without human image annotations.

Internalizing Visual Supervision into Understanding

To convert verified visual thoughts into stronger visual comprehension without test-time drawing overhead, our primary update (SelfThinker) distills the verified helper images back into the understanding pathway via privileged on-policy self-distillation ([1, 33]). For each verified item $(I,q,v^\star) \in\mathcal{D}_{\mathrm{ver}}$, we construct asymmetric student and teacher contexts:

The student $\pi_\theta^U$ observes only the test-time context $c_S$ and samples an on-policy text completion $y = (y_1,\ldots,y_T) \sim\pi_\theta^U(\cdot\mid c_S)$ without image generation. An exponential moving average (EMA) teacher $\bar{\pi}_{\bar{\theta}}^U$ ($\bar{\theta} \leftarrow\mu\bar{\theta} + (1-\mu)\theta$) observes the privileged context $c_T$ containing the verified helper image $v^\star$. The teacher is initialized from $\theta$ and evaluated with stop-gradient, whereas the Stage 2 verifier remains fixed at the base ThinkMorph-7B weights. Letting $P_t^T = \bar{\pi}_{\bar{\theta}}^U(\cdot\mid c_T,y_{<t})$, $P_t^S = \pi_\theta^U(\cdot\mid c_S,y_{<t})$, and $M_t = \beta P_t^T + (1-\beta)P_t^S$, we minimize the token-level generalized Jensen–Shannon divergence (JSD) along $y$ using the Kullback–Leibler (KL) divergence:

Minimizing Eq. (6) teaches the student to bypass the <image_start> bottleneck and extract the spatial cues highlighted by $v^{\star}$ directly from the original image $I$. At test time, SelfThinker reasons purely in text from $c_S = [I, q]$ without generating intermediate images.

Experiments

Experimental Setup

Compared models.

We build on ThinkMorph-7B (Gu et al. [20]), an interleaved reasoning model fine-tuned from BAGEL-7B-MoT (Deng et al. [14]) (denoted BAGEL-7B). The model routes understanding and generation tokens through specialized Qwen2.5 transformer experts (Qwen Team [41]) paired with a SigLIP vision encoder (Zhai et al. 2023). All intrinsic utility scores are evaluated by the frozen understanding pathway of ThinkMorph-7B. We evaluate our self-distilled understanding models across training pools and supervision settings: SelfThinker (Nav) (distilled from verified navigation thoughts), SelfThinker (Mix) (distilled from combined navigation and visual-search thoughts), SelfThinker (Vote) (distilled from visual-search thoughts verified via label-free majority voting), and our primary model SelfThinker (distilled from visual-search thoughts verified via ground-truth answers). In our ablation studies (Sect. 4.5), we also evaluate two teacher-context controls under identical visual-search training schedules. Specifically, Noise-OPSD replaces the teacher’s verified thought with uniform RGB noise, and Prompt-OPSD omits the second image so the teacher observes only the problem and transition instruction. For the generation pathway, we compare our supporting generative update (SelfThinker-Gen) against random-thought fine-tuning (Rand-RFT).

Synthesized problem pools.

To evaluate candidate visual thoughts on layouts unseen during ThinkMorph-7B training, we synthesize fresh Navigation (Frozen-Lake grid maps (Towers et al. [51])) and Visual Search (multi-object shape canvases) pools. Each pool is split into 1,500 training and 500 diagnostic items with disjoint layouts. For each problem, we sample $K = 6$ interleaved traces with one visual thought per trace. Programmatic reference thoughts (shortest-path overlays and target boxes) are rendered only for post-hoc diagnostics and are never used in verification or training. Complete pool specifications, sampling parameters, and optimization schedules are provided in Appendix § 6.

Evaluation protocols.

Following the benchmark suite of Li et al. [33], we evaluate multimodal reasoning across seven benchmarks: VSP (600 maze items) (Wu et al. [59]), ChartQA (444 horizontal and 382 vertical bar items) (Masry et al. [38]), V* (191 items) (Wu and Xie [58]), MMVP (300 items) (Tong et al. [50]), SAT (150 circular split items) (Ray et al. [43]), BLINK (1,901 items across 14 subtasks) (Fu et al. [18]), and CV-Bench (1,438 2D and 1,200 3D items) (Tong et al. [48]). All benchmarks are evaluated in VLMEvalKit (Duan et al. [16]) using greedy text-only decoding (up to 1,024 thinking tokens) and a unified Qwen3-30B-A3B-Instruct-2507 judge (Yang et al. [61]; Kwon et al. [30]). On the 500-item diagnostic splits, we evaluate intrinsic verification using four metrics: verification yield, Top-1 AUC (area under the receiver operating characteristic curve) (Hanley and McNeil [23]), paired win rate, and utility margin (nats) (formal definitions in Appendix § 7). Finally, to test whether accuracy gains stem from visual perception rather than language priors, we evaluate a multi-layer image-condition probe across image swaps, pixel scrambles, tile permutations, mosaics, and Gaussian blurs (Appendix § 11).

Main Benchmark Results

Improving visual understanding without test-time image generation.

Table 2 and Figure 2 present the main benchmark comparison. By distilling verified visual thoughts into the understanding pathway, Self-Thinker raises the seven-benchmark average from 52.7 to 64.0 (+11.3 points). It outperforms ThinkMorph-7B on every benchmark while operating purely in text mode at test time, avoiding the 59–75 s per-trace diffusion overhead required by interleaved image generation. Notably, even though training uses only a small pool of synthetic navigation or visual-search problems, the distilled visual supervision transfers broadly across natural-image and spatial benchmarks. Specifically, SelfThinker (Nav) lifts the seven-benchmark average to 60.0 (+7.3 points, reaching 60.3 on BLINK), SelfThinker (Mix) reaches 60.8 (+8.1 points, reaching 83.5 on CV-Bench), label-free SelfThinker (Vote) reaches 63.4 (+10.7 points, reaching 81.9 on ChartQA), and SelfThinker reaches 64.0 (+11.3 points).

Method	VSP	ChartQA	$V^*$	MMVP	SAT	BLINK	CV-B.	Average	$\Delta$
ThinkMorph-7B ([20])	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	–

Cross-task gains on vision-centric reasoning.

Grouping the benchmarks by task type shows how self-generated visual supervision transfers to diverse visual domains. On fine-grained visual reading and search tasks that require locating small objects or reading dense charts, SelfThinker improves ChartQA by +27.4 points (to 80.1, and 81.9 for SelfThinker (Vote)) and $V^*$ by +16.3 points (to 70.2). On multi-step spatial planning, SelfThinker lifts VSP maze navigation from near-failure (1.2) to 15.8 (+14.6 points) without drawing intermediate paths at test time, while improving perspective reasoning on SAT by +4.7 points (to 60.0). On 2D and 3D natural-scene perception, SelfThinker consistently advances CV-Bench (+8.3, to 81.7), BLINK (+4.6, to 58.2), and MMVP (+3.3, to 82.3).

Cross-Task Perception Transfer and Internalization

Unlocking multi-scale visual perception beyond the pretrained BAGEL-7B backbone.

To verify that these cross-task gains reflect stronger visual perception rather than language shortcuts, we compare SelfThinker against both ThinkMorph-7B and its pretrained BAGEL-7B backbone under the 512-token image-condition probe (Figure 3b; full 256- and 512-token corruption tables in Appendix § 11). Across all four spatial and object-perception benchmarks, SelfThinker surpasses ThinkMorph-7B by 33.0 to 40.4 points and outperforms the pretrained BAGEL-7B backbone by 12.0 to 23.2 points. Specifically, SelfThinker reaches 67.0% on V* (+18.3 over BAGEL-7B), 78.0% on MMVP (+12.0), 83.4% on CV-Bench-2D (+20.9), and 82.0% on CV-Bench-3D (+23.2), while also lifting ChartQA-h/v under the 256-token probe from 32.2%–32.5% in ThinkMorph-7B to 67.3%–68.6% (Table 10). Because pixel-scramble accuracy remains at a similar baseline across all three models, the entire improvement over BAGEL-7B stems from stronger visual evidence extraction from the real problem image. Graded mosaic, Gaussian blur, and tile-permutation ladders in Appendix § 11 further show that SelfThinker preserves global interobject geometry across $2 \times2$ to $8 \times8$ mosaics and mild blur on 2D and 3D spatial tasks, while extracting fine local glyphs on ChartQA.

Method	real	real†	wrong†	gray	scr.	×2	×4	×8	×16	σ1	σ2	σ4	σ8
ChartQA, horizontal bars (444 items, 222 with a different previous image)
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
ChartQA, vertical bars (382 items, 191 with a different previous image)
ThinkMorph-7B	32.5	14.1	4.2	3.4	6.0	26.7	8.6	3.7	3.9	26.4	18.3	6.5	4.5
SelfThinker (Nav)	42.9	30.4	4.2	7.9	8.4	40.8	16.2	–	–	42.7	29.3	14.1	7.9
Prompt-OPSD	53.4	41.9	6.8	11.0	10.7	46.6	18.3	11.0	–	52.4	33.0	15.7	–
Noise-OPSD	61.0	52.9	5.8	8.6	10.2	54.7	18.1	11.3	–	60.2	39.5	15.7	11.8
SelfThinker	68.6	54.5	4.7	11.5	13.1	63.1	19.9	12.8	11.8	63.4	43.2	15.4	14.7
V* (191 items)
ThinkMorph-7B	34.0	34.0	34.0	32.5	36.6	35.6	37.2	34.6	36.1	36.1	36.1	35.1	35.6
SelfThinker (Nav)	56.0	56.0	40.3	38.2	38.2	57.6	50.3	52.4	38.2	59.2	56.5	53.9	45.0
Prompt-OPSD	50.8	50.8	36.6	34.6	36.1	51.3	49.7	45.5	–	52.4	51.8	47.1	41.4
Noise-OPSD	51.3	51.3	31.4	35.6	34.0	51.8	50.3	49.7	–	49.2	49.7	51.3	35.6
SelfThinker	67.0	67.0	38.2	35.1	33.0	–	63.9	61.8	46.1	64.4	66.5	56.5	46.1

Eliminating premature termination at the image slot.

As shown in Figure 3a, when test-time image generation is disabled, ThinkMorph-7B frequently halts prematurely at the <image_start> token, completing a final text answer on only 32% to 73% of items across V*, CV-Bench-2D, ChartQA-h, MMVP, and SAT. SelfThinker raises direct text answer completion to 77%–100% (reaching 100% on V* and 99% on SAT). Crucially, on the paired subsets of items where ThinkMorph-7B already states an answer before <image_start>, SelfThinker still improves conditional accuracy under the 512-token probe from 35.3% to 66.8% on V* (+31.5 points), from 46.6% to 82.9% on CV-Bench-2D (+36.3), from 37.2% to 68.4% on ChartQA-h (+31.3), from 48.9% to 77.4% on MMVP (+28.5), and from 50.4% to 60.0% on SAT (+9.6). This confirms that self-distillation strengthens visual reasoning itself rather than merely repairing answer formatting.

Internalizing visual thoughts without test-time drawing.

To verify that the student has absorbed the privileged teacher conditioning into single-image reasoning, we perform controlled context interventions on the 1,092 verified visual-search items (Table 9 and Figure 9 in Appendix E). Whereas ThinkMorph-7B drops by 6.1 points when the verified visual thought is removed at test time (45.9% with the thought versus 39.7% with the problem image alone), SelfThinker achieves virtually identical accuracy whether given the problem plus its verified thought (99.0%) or the problem image alone (98.9%, a difference of +0.1 points). Label-free SelfThinker (Vote) likewise matches at 95.4% versus 95.3% (+0.1 points). Conversely, when the problem image is withheld and only the verified thought is shown, SelfThinker and SelfThinker (Vote) reach 96.5% and 93.0% accuracy (versus 35.4% for ThinkMorph-7B), showing that distillation sharpens the model’s visual reader. To evaluate generalization on unseen problem types, we also test SelfThinker (Nav), which is distilled solely on grid navigation and never exposed to visual-search items. SelfThinker (Nav) transfers zero-shot to the 1,092 visual-search problems at 92.9% single-image accuracy (with a 0.0-point gap between with and without the visual thought, up from 39.7% in ThinkMorph-7B).

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

Intrinsic Verification Analysis

Why the understanding pathway must select what is worth learning.

We first examine how much answer evidence the frozen understanding pathway extracts from ThinkMorph’s human-annotated helper images across its training tasks (Figure 4a). While shortest-path overlays carry strong utility on both training maps (AUC 0.95) and unseen fresh maps (AUC 0.81), fixed human templates on visual search and chart refocus yield near-chance discrimination between matched and mismatched thoughts (AUC 0.52–0.53). Prescribing a fixed helper-image template per task does not guarantee that the model’s reader can extract useful evidence from it; scoring candidate thoughts directly through the understanding pathway provides a task-agnostic criterion for selecting informative visual expressions.

Separating helpful visual cues from conflicting imagery.

Table 3 and Figure 4b evaluate intrinsic verification on self-generated thoughts across the unseen diagnostic splits. On navigation, unverified candidate rollouts average −9.1 nats and mismatched thoughts from other maps reduce target answer log-likelihood by −11.8 nats on average. In contrast, verified intrinsic thoughts $v^\star$ achieve a positive mean utility of +2.0 nats, exceeding programmatic GT path overlays (+0.6 nats; Figure 4b), while reaching a Top-1 AUC of 0.821, a verification yield of 0.674, and a utility margin of 2.10 nats (Table 3). On visual search, appending a second image introduces a constant baseline log-likelihood offset across candidates (rows 3–4 of Figure 5). Calibrating against same-item mismatched thoughts $u_i(v_{\mathrm{mis}}^{(k)})$ removes this offset and achieves a Top-1 AUC of 0.660 and a yield of 0.578. Comparing generated red boxes against the ground-truth target coordinates on the 500 unseen visual-search diagnostic items further confirms that thoughts boxing the true queried target achieve a +0.31-nat utility margin over mismatched thoughts, whereas boxes placed on distractor objects average −0.04 nats.

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

Label-free verification via rollout majority voting.

We next test whether visual thoughts can be verified without ground-truth answer labels. On visual search, 47.4% of diagnostic items exhibit high text answer agreement across visibly diverse visual thoughts, and trace answer correctness is nearly orthogonal to the visual utility of the accompanying thought ($\rho= 0.022$; full analysis in Appendix § 9). Consequently, text consensus alone cannot certify image quality. However, when the rollout majority vote $\tilde{a}$ is used to set the target direction $d$ in Eq. (2), label-free vote verification closely matches ground-truth answer verification (Table 3), achieving a 0.895 Spearman rank correlation with ground-truth utilities, a 54.5% paired win rate, and a 63.4 seven-benchmark average (+10.7 points over ThinkMorph-7B, Table 8).

Method	VSP	ChartQA	V*	MMVP	SAT	BLINK	CV-B.	Average	Δ
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
Noise-OPSD	13.2	78.3	65.4	80.0	56.0	58.5	81.2	61.8	+9.1
SelfThinker (Vote, label-free)	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
Δ vs. ThinkMorph-7B	+14.6	+27.4	+16.3	+3.3	+4.7	+4.6	+8.3	+11.3	–

Ablation Studies

Privileged visual thoughts versus context distillation.

Table 4 disentangles the factors driving SelfThinker’s benchmark gains (complete per-benchmark breakdown in Appendix § 8). First, distilling verified thoughts into the understanding pathway via on-policy self-distillation yields large gains across all task pools. Performance scales consistently from navigation (SelfThinker (Nav) at 60.0, +7.3) and mixed pools (SelfThinker (Mix) at 60.8, +8.1) to visual search (SelfThinker at 64.0, +11.3), and label-free SelfThinker (Vote) at 63.4, +10.7). Across 13 independent training seeds for SelfThinker (62.4 ± 1.0, range 60.7–64.0) and 11 seeds for label-free SelfThinker (Vote) (61.9 ± 0.8, range 61.1–63.4), every run matches or outperforms Prompt-OPSD (60.7). Second, paired item-level bootstrap tests (10,000 resamples) confirm that conditioning the teacher on verified visual thoughts (SelfThinker at 64.0) significantly outperforms both Prompt-OPSD (+3.3 points, 95% CI [+1.7, +5.0], $p < 10^{-4}$) and Noise-OPSD

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
SelfThinker	Visual-search distillation with verified visual thoughts	64.0	+11.3

(+2.2 points, 95% CI [+0.8, +3.8], $p$ = 0.0015). Label-free SelfThinker (Vote) (63.4) likewise significantly exceeds both controls (+2.7 points, CI [+1.0, +4.2]; +1.6 points, CI [+0.1, +3.1]), while verified visual thoughts further provide +10.7 to +23.3 point gains on fine-grained 2D and 3D image-condition probes (Appendix § 11).

Supporting generative refinement improves candidate training imagery.

Table 5 evaluates the supporting generative update (SelfThinker-Gen) on candidate thought generation across the 500-item unseen diagnostic splits. Training on randomly selected thoughts (Rand-RFT) fails to improve utility margins on unseen navigation maps (2.06 versus 2.10 nats). In contrast, SelfThinker-Gen raises the navigation utility margin to 2.96 nats (+0.90 nats over Rand-RFT, lifts Top-1 AUC from 0.821 to 0.847, and increases yield from 0.674 to 0.694. On visual search, SelfThinker-Gen improves every verification metric (lifting yield from 0.578 to 0.632 and Top-1 AUC from 0.660 to 0.675) while increasing the localized bounding-box generation rate from 40.8% to 50.8% (Appendix § 9), providing cleaner visual supervision for understanding evolution.

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

Conclusion

We introduced SelfThinker, a self-evolving framework that turns a unified multimodal model’s image generation capability into training supervision for visual understanding without human image annotations. By scoring candidate visual thoughts through the model’s frozen understanding pathway via calibrated conditional pointwise mutual information, SelfThinker selects informative helper images under both ground-truth answers and label-free majority votes. Distilling these verified visual thoughts into the understanding pathway via privileged on-policy self-distillation, supported by flow-matching generator refinement, enables the deployed model to reason directly from the original image in text without generating images at test time. Although trained only on small synthetic visual-search and navigation pools, SelfThinker improves the seven-benchmark average of ThinkMorph-7B by +11.3 points (and by +10.7 points under label-free majority voting). Controlled image-perturbation probes further show that this self-generated visual supervision transfers broadly across tasks, unlocking fine-grained and 2D/3D spatial perception that surpasses ThinkMorph-7B by up to 40.4 points and the pretrained BAGEL-7B backbone by up to 23.2 points.

Implementation Details

Rollout sampling.

Table Table 6 summarizes the synthesized navigation and visual-search problem pools, and Table 7 lists the training hyperparameters for SelfThinker-Gen and SelfThinker. For each item, we sample $K = 6$ interleaved traces with one visual thought each using the interleaved inferencer of ThinkMorph-7B. Sampling uses 10 denoising steps per thought, text temperature 0.7, classifier-free guidance scales of 4.0 for text and 2.0 for images, and at most 256 thinking tokens. Generating one interleaved trace takes 59 s on navigation and 75 s on visual search on one NVIDIA H20 GPU, whereas scoring a candidate thought with its controls takes 2.3–2.8 s. Navigation thoughts are sampled from ThinkMorph-7B (and from each fine-tuned generator for generator comparisons), while visual-search thoughts are sampled from the navigation SelfThinker-Gen checkpoint. All understanding students are initialized from ThinkMorph-7B.

Pool	Problem image	Question and answer	Reference thought
Navigation	Frozen-Lake grid, $4 \times4$ to $8 \times8$, 12–28% holes, $192 \times192$ 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), $384 \times384$ px	Colour of a target shape or shape of a target colour (13 classes)	Red bounding box

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
Sequence budget	16,384 tokens per packed batch (8,192 expected)	10,240 tokens per forward pass
Parallelism	Fully sharded data parallel over 8 GPUs	Fully sharded data parallel over 8 GPUs
Image transforms	Vision Transformer (ViT) stride 14, 378–980 px; variational autoencoder (VAE) stride 16, 512–1,024 px	ViT stride 14, 378–980 px, at most 2,007,040 pixels

Threshold calibration.

For intrinsic verification in Eq. (2), the plain context $C(I,q)$ and privileged context $C(I,q,v) = [I,q,\rho,v,\tau]$ are formatted identically to the student and teacher contexts of Eq. (6). We use the transition instruction $\tau$ (“Now, using your own independent reasoning, answer the problem above. Think step by step.”) and the privileged context header $\rho$ (“The following images are additional visual context. Use them silently as grounding context, and do not describe or echo them.”). The mismatched partner $v_{\mathrm{mis}}^{(c)}$ of each thought is assigned by a fixed cyclic offset across items in the scoring order. This ensures that every candidate is compared against an in-distribution thought generated by the same model for a distinct problem. For training set construction, the 95th-percentile threshold $Q_i$ uses the order-statistic form, i.e., the $(\lceil0.95K - 1\rceil+ 1)$-th smallest element of the mismatched utilities $s_i$ (the second largest for $K = 6$), whereas diagnostic yield calculations use the linearly interpolated 95th percentile.

Teacher-context controls.

Navigation training uses 1,179 verified thoughts from 1,472 scored training items. Visual search uses 1,092 verified thoughts (or 1,074 under the vote direction) from 1,500 items, and the mixed pool combines both verified sets. In SelfThinker-Gen, the flow-matching loss on latent thought tokens has weight 1.0 and the cross-entropy loss on answer tokens has weight 0.25. Rand-RFT uses the exact same 1,179 navigation items and optimization schedule as SelfThinker-Gen, but selects one thought per item uniformly at random without reading $u$. The teacher-context controls on visual search (Noise-OPSD and Prompt-OPSD) reuse the exact items and optimization settings in Table 7 (1,000 steps on eight NVIDIA H20 GPUs with learning rate $10^{-5}$ and cosine decay, with navigation distillation trained for 2,000 steps). They modify only the teacher context $c_T$: replacing $v^{\star}$ with a $512 \times512$ image of independent uniform RGB noise in Noise-OPSD, or omitting $v^{\star}$ in Prompt-OPSD.

Diagnostic Definitions

Verification metrics.

On the 500-item diagnostic splits, the utility margin is the mean difference $u(v) - u(v_{\mathrm{mis}})$ in nats across all candidate pairs, and the win rate is the percentage of pairs with $u(v) > u(v_{\mathrm{mis}})$. The Top-1 AUC computes the area under the ROC curve comparing each item’s top intrinsic thought against the mean of its mismatched utilities ([23]), and the yield is the fraction of items whose top thought exceeds $Q_i$. As an analytical diagnostic for label-free verification, we also report metrics on the accurate-vote subset where the rollout majority vote $\tilde{a}$ matches $a^*$.

Bounding-box detector.

Each visual-search thought is converted to RGB, and a pixel is classified as red if $r \ge160$, $g \le90$, $b \le90$, and $r - \max(g,b) > 50$. A thought is flagged as containing a red bounding box if its red pixel fraction lies in $[0.0015, 0.18]$ with at least 30 red pixels, its red bounding box is at least $10 \times10$ pixels, and the red density along the perimeter band of the bounding box is at least 0.08 and exceeds the interior red density by at least 0.04. teacher controls, and SelfThinker. Figure 6 visualizes the overall and per-benchmark gains relative to ThinkMorph-7B.

Dual consensus.

On visual search, an item has high answer agreement if the majority vote accounts for at least 80% of the $K = 6$ rollout answers. It has high visual agreement if the mean pairwise cosine similarity of its six thoughts (after resizing to $48 \times48$ RGB arrays and subtracting each array’s mean) exceeds the median across items.

Pixel deviation.

Pixel deviation measures the mean absolute difference between a visual thought and its problem image after resizing both to $96 \times96$ RGB in $[0, 1]$. Programmatic path overlays deviate by 0.015 on average, and self-generated navigation thoughts have quartile boundaries of 0.012, 0.017, and 0.043.

Per-Benchmark Breakdown of Distillation Variants

Table 8 reports the complete per-benchmark scores of all task-pool variants, visual-search teacher controls, and SelfThinker. Figure 6 visualizes the overall and per-benchmark gains relative to ThinkMorph-7B.

Across training pools (Figure 6a), navigation distillation in SelfThinker (Nav) lifts the seven-benchmark average by +7.3 points (reaching 60.0 and achieving the highest score of 60.3 on BLINK). Combining navigation with visual-search thoughts in SelfThinker (Mix) raises the gain to +8.1 points (60.8) and achieves the highest score on CV-Bench (83.5). Moving to the visual-search pool yields a further jump: Prompt-OPSD (+8.0, 60.7) and Noise-OPSD (+9.1, 61.8) establish strong context-distillation baselines, and conditioning the teacher on verified visual thoughts raises the gain to +10.7 points for label-free SelfThinker (Vote) (63.4, reaching 81.9 on ChartQA) and +11.3 points for SelfThinker (64.0). Across individual benchmarks (Figure 6b), SelfThinker consistently improves over Prompt-OPSD and Noise-OPSD on ChartQA (80.1 versus 77.7 and 78.3), V* (70.2 versus 68.6 and 65.4), VSP (15.8 versus 3.3 for Prompt-OPSD and 13.2 for Noise-OPSD), MMVP (82.3 versus 79.7 and 80.0), SAT (60.0 versus 56.7 and 56.0), and CV-Bench (81.7 versus 80.6 and 81.2).

Additional Verification Analyses

Acceptance across pixel deviation quartiles.

On navigation, verified thoughts span the full range of pixel deviation from the problem map, with acceptance rates of 11.4%, 13.9%, 11.6%, and 15.1% across the four deviation quartiles. The highest-deviation quartile, which includes complete scene re-renderings such as row 2 of Figure 5, achieves the highest acceptance rate (15.1%).

Answer consensus versus visual consensus.

Figure 7a shows the dual-consensus breakdown across the 500 diagnostic visual-search items: 48.4% of items have both high answer agreement and visually similar thoughts, while 47.4% have high answer agreement produced from visibly diverse thoughts (matching the 48.5% and 47.3% split on the 1,500 training items). In this diverse-imagery regime, intrinsic verification identifies a validated helper image on 66% of items. Figure 7b,c further shows that the label-free vote direction matches the answer direction in paired win rate (54.5% versus 54.1%, and 54.9% on accurate-vote items), while trace answer correctness is orthogonal to the visual utility of the thought ($\rho= +0.022$).

Generation pathway gains and target-box selectivity.

Figure 8 illustrates how SelfThinker-Gen improves the generation pathway on both tasks. On navigation (Figure 8a), SelfThinker-Gen lifts the utility margin on unseen diagnostic maps from 2.10 nats (ThinkMorph-7B) and 2.06 nats (Rand-RFT) to

2.96 nats, confirming that verification drives the improvement in generated imagery. On visual search (Figure 8b), SelfThinker-Gen increases the share of thoughts that draw a localized red bounding box from 40.8% to 50.8%, while raising the utility-threshold pass rate among both box thoughts (from 26.4% to 29.7%) and non-box thoughts (from 25.0% to 29.9%). Evaluating against the ground-truth target coordinates on the 500 unseen diagnostic items further confirms that thoughts boxing the true queried object achieve a mean utility margin of +0.31 nats over mismatched thoughts, compared with −0.04 nats for boxes on distractor objects.

Context Interventions on Visual Search

We evaluate how ThinkMorph-7B and the SelfThinker students respond to five controlled context interventions on the 1,092 verified visual-search items (Table 9 and Figure 9). Whereas ThinkMorph-7B loses 6.1 percentage points when the helper image is removed at test time (45.9% with “Own” versus 39.7% with “None”), Self-Thinker and label-free SelfThinker (Vote) achieve virtually identical accuracy whether given the item’s own verified thought (99.0% and 95.4%) or no second image at all (98.9% and 95.3%,) a difference of +0.1 points). This confirms that the distilled students absorb the privileged visual conditioning without requiring a second image at deployment. Appending a neutral sentence leaves accuracy unchanged (98.9% and 95.3%). When the problem image is withheld and only the verified visual thought is shown (“Thought only”), SelfThinker and SelfThinker (Vote) achieve 96.5% and 93.0% accuracy (compared with 35.4% for ThinkMorph-7B), demonstrating that their understanding pathway directly extracts the solution from the verified thought. Furthermore, to test generalization on unseen problem types, SelfThinker (Nav), trained solely on navigation problems, reaches 92.9% zero-shot on these 1,092 visual-search items both with and without the visual thought (0.0-point gap, up from 45.9% and 39.7% for ThinkMorph-7B).

Multi-Layer Image-Condition Probe

Tables 10, 11, and 12 report the complete results of the image-condition probe at 256 and 512 thinking tokens. Figures 10, 11, and 12 illustrate the visual transformations, the per-benchmark visual-sensitivity heatmaps, and the three degradation ladders. Unlike the main benchmark evaluation in Table 2 (which uses 1,024 thinking tokens and semantic grading by Qwen3-30B-A3B-Instruct-2507), the image-condition probe uses a strict 256- or 512-token budget and deterministic rule-based extraction (the last standalone option letter for multiple-choice items or the last number within 5% relative error for ChartQA) to isolate visual sensitivity under controlled image corruptions.

The prompt presents a single image and the question, and multiple-choice items append “Answer with a single letter.” Final answers are extracted from explicit <answer> tags or post-</think> text when present, falling back to the last standalone option letter or number before <image_start>. Image swaps use the previous row of the benchmark file on the subset of items whose previous row depicts a distinct image (222 horizontal and 191 vertical ChartQA items, 1,198 CV-Bench-3D items, and all items elsewhere). Pixel scrambles permute all pixels deterministically, and $g \times g$ tile permutations partition the image into $g^2$ equal rectangular patches and shuffle those patches, preserving local detail inside each patch while disrupting global spatial layout. Mosaics apply box downsampling by factor $k \in\{2, 4, 8, 16\}$ followed by nearest-neighbour upsampling, and blurs apply a Gaussian kernel with $\sigma\in\{1, 2, 4, 8\}$ pixels at native resolution.

Visual perception advantage over teacher controls.

At 256 thinking tokens (Figure 11 and Table 10–Table 11), SelfThinker achieves the highest accuracy on every benchmark: 67.3%–68.6% on ChartQA-h/v, 67.0% on $V^*$, 78.0% on MMVP, 83.5% on CV-Bench-2D, and 81.8% on CV-Bench-3D, outperforming ThinkMorph-7B (32.2%–46.8%) by wide margins. The real problem image provides a 55–56 point accuracy gain over pixel scramble on ChartQA, 51 points on CV-Bench-2D, 32 points on MMVP, 34 points on $V^*$, and 28 points on CV-Bench-3D (Figure 11b). Crucially, SelfThinker achieves substantially higher real-image accuracy than both teacher controls (Noise-OPSD and Prompt-OPSD), exceeding them by +15.7 and +16.2 points on $V^*$ (67.0% versus 51.3% and 50.8%), +4.7 and +11.0 points on MMVP (78.0% versus 73.3% and 67.0%), +15.2 and +10.7 points on CV-Bench-2D (83.5% versus 68.3% and 72.8%), and +23.3 and +13.6 points on CV-Bench-3D (81.8% versus 58.5% and 68.2%).

Method	real	real†	wrong†	gray	scr.	×2	×4	×8	×16	$\sigma$1	$\sigma$2	$\sigma$4	$\sigma$8
MMVP (300 items)
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
CV-Bench-2D (1,438 items)
ThinkMorph-7B	40.2	40.2	36.4	32.6	33.4	41.5	42.6	41.9	41.0	41.2	43.1	42.9	42.4
SelfThinker (Nav)	78.1	78.1	39.4	32.3	32.4	76.7	74.6	70.8	61.1	79.1	78.0	75.2	69.0
Prompt-OPSD	72.8	72.8	39.1	31.8	30.7	71.1	69.1	63.4	–	71.8	72.3	66.2	61.1
Noise-OPSD	68.3	68.3	40.5	31.6	33.8	68.6	66.1	55.4	–	67.6	65.6	62.1	56.3
SelfThinker	83.5	83.5	42.4	33.9	32.4	82.5	80.9	75.0	67.7	84.1	81.2	78.4	72.6
CV-Bench-3D (1,200 items, 1198 with a different previous image)
ThinkMorph-7B	46.8	46.7	45.2	47.5	48.0	49.2	48.3	48.2	43.9	46.8	48.2	49.8	46.4
SelfThinker (Nav)	66.5	66.5	52.9	53.2	50.2	66.8	68.1	67.4	62.3	68.2	69.5	68.2	67.3
Prompt-OPSD	68.2	68.2	51.8	48.3	51.3	65.1	67.4	63.7	–	68.8	67.8	62.0	59.5
Noise-OPSD	58.5	58.6	52.5	47.5	48.7	60.0	61.2	55.5	–	58.9	58.8	57.7	54.3
SelfThinker	81.8	81.8	54.8	52.3	54.3	83.8	82.2	77.2	69.6	83.0	81.8	81.1	74.2

Multi-scale spatial perception.

Comparing the transformation families across Table 10–Table 12 and Figure 12 highlights two complementary visual perception capabilities. On ChartQA horizontal and vertical bars (Table 10), SelfThinker preserves high accuracy on $2 \times2$-pixel mosaics (64.9% and 63.1%) and $\sigma= 1$ blur (68.9% and 63.4%), demonstrating strong local glyph reading. On $V^*$, MMVP, CV-Bench-2D, and CV-Bench-3D, at 512 thinking tokens (Table 12 and Figure 12), SelfThinker preserves its large advantage over both ThinkMorph-7B and pretrained BAGEL-7B across $2 \times2$ to $8 \times8$-pixel mosaics, mild blur ($\sigma= 2, 4$), and $2 \times2$ to $16 \times16$ tile permutations. Specifically, on CV-Bench-2D, SelfThinker achieves 82.8%, 80.7%, and 75.2% across ×2, ×4, ×8 mosaics (versus 32.5% on scramble), and on CV-Bench-3D it achieves 82.2% and 77.2% across ×4 and ×8 mosaics (versus 54.5% on scramble). Together with the 12.0 to 23.2 point improvement over pretrained BAGEL-7B across all four spatial and object-level benchmarks (Table 12), these results confirm that SelfThinker unlocks high-precision visual perception across both fine local details and global scene geometry.

Method	real	scr.	×2	×4	×8	×16	σ2	σ4	σ8	2 × 2	4 × 4	8 × 8	16 × 16
$V^*$ (191 items)
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)
BAGEL-7B	66.0	50.3	58.7	–	52.7	53.0	52.7	52.3	49.0	59.0	52.3	51.3	50.3
ThinkMorph-7B	44.7	45.7	45.3	–	–	47.0	51.7	46.0	–	48.0	47.3	49.3	51.3
SelfThinker (Nav)	70.0	45.3	69.3	–	53.0	50.3	68.7	63.3	57.7	64.7	61.0	59.3	50.0
SelfThinker	78.0	46.0	74.3	–	53.3	52.7	72.7	63.0	62.7	75.7	63.3	59.3	53.3
CV-Bench-2D (1,438 items)
BAGEL-7B	62.5	36.7	58.6	57.2	53.0	47.1	59.5	56.7	53.1	46.9	42.8	42.9	42.5
ThinkMorph-7B	43.0	33.5	41.7	42.6	41.9	41.2	43.1	43.0	42.4	45.0	43.4	43.3	40.6
SelfThinker (Nav)	78.0	32.4	76.6	74.8	70.9	61.1	78.0	75.2	69.1	59.3	55.5	51.3	44.9
SelfThinker	83.4	32.5	82.8	80.7	75.2	67.8	81.2	78.4	72.7	64.0	57.4	55.1	52.4
CV-Bench-3D (1,200 items)
BAGEL-7B	58.8	51.4	–	57.1	55.6	54.3	59.5	57.8	55.0	55.8	52.8	50.0	50.9
ThinkMorph-7B	47.6	48.8	–	49.8	49.1	45.0	49.4	50.7	47.5	50.0	52.7	54.2	54.1
SelfThinker (Nav)	71.2	50.7	–	73.9	72.1	65.2	72.2	70.3	69.2	68.2	63.1	57.4	55.2
SelfThinker	82.0	54.5	–	82.2	77.2	69.5	81.9	81.2	74.2	76.6	68.8	61.6	57.1