1 Introduction↩︎

Recent work on self-evolving large multimodal models (LMMs) has demonstrated that models can improve their visual reasoning capabilities in a purely unsupervised manner, without relying on human-annotated supervision or externally verified reward models [1]–[4]. These approaches frame multimodal reasoning as a self-improving process, instantiating proposer–solver or questioner–reasoner roles within a shared backbone and reinforcing multi-sample agreement, trajectory-aware feedback, and tool-assisted verification to promote more reliable reasoning [1]–[3], [5]. Despite their promise, however, these methods optimize for answer agreement without ensuring the decoder attends to the actual visual content, relying instead on statistical language priors to produce self-consistent outputs. This leads to a persistent failure mode we term visual under-conditioning, which causes hallucination, modality bypass, and unstable visual interpretation even when the vision encoder produces accurate representations. Concretely, this manifests as insufficient attention to visual tokens during decoding, where generations are driven more by language priors than by the image being described.

This limitation arises from two structural features common to existing self-evolving frameworks. First, the Proposer–Solver setup forms an implicit minimax game: the roles are optimized jointly with opposing objectives, making training unstable over long horizons. In practice, one role often dominates; Proposer collapses to trivial or degenerate queries that guarantee agreement, or the Solver overfits to the Proposer’s distribution and fails to generalize, driving the system into local minima that are difficult to correct without external intervention. Second, framing the reward around answer correctness implicitly assumes that self-consistency implies correct visual grounding. A decoder exploiting language priors can achieve high answer agreement through statistical co-occurrences alone, leaving visual under-conditioning unaddressed or even reinforced.

As illustrated in Fig. 1, this failure mode is directly observable. On chart-based or structured scientific queries, prior self-evolving methods produce the same correct answers as ours, suggesting that the required visual cues may be sparse or quickly resolved, after which the decoder can proceed largely from its internal representations. However, on natural scene understanding queries such as identifying what a skateboard is actually resting on, these methods default to statistically common associations (ramp surface, concrete ground), indicating that visual evidence is not being robustly integrated throughout the generation process. In contrast, correctly resolving such cases requires the decoder to remain conditioned on the specific image content, exposing the gap between answer consistency and genuine visual interpretation. Prior self-evolving methods trained on science- and math-centric images show gains on reasoning benchmarks, yet perform at or below the base model on captioning and region-level description tasks that require detailed visual conditioning. This disparity validates the hypothesis that improvements in answer agreement do not translate into stronger visual interpretation, and that visual under-conditioning persists in existing self-evolving frameworks.

To directly address this, we propose VISE (Visual Invariance Self-Evolution), a purely unsupervised self-evolving framework that regularizes the model’s visual conditioning policy rather than answer agreement. Unlike prior multi-role formulations, VISE operates within a single model, motivated by the observation that a well-pretrained LMM already possesses sufficient knowledge to formulate meaningful queries about its visual content. Its training signal comes from two complementary invariance-based rewards: a geometric invariance reward that enforces spatial consistency under known transformations, and a semantic invariance reward that requires the model to recognize the absence of evidence when predicted regions are perturbed. Training on raw images with no annotations, metadata, or external reward models, VISE achieves gains of up to \(+16.85\) CIDEr on COCO and \(+19.66\) CIDEr on TextCaps, reduces object hallucination (Chair-I) by \(5.0\) points, and improves consistently across VQA and reasoning benchmarks, demonstrating that stronger visual conditioning generalizes across tasks rather than trading off against them. Mechanistically, VISE achieves this by increasing attention to visual tokens across decoder layers during generation, reflecting the shift from language-prior-driven to image-conditioned decoding.

In summary, our main contributions are as follows:

• We introduce VISE, a single-model, fully unsupervised self-evolving framework that directly addresses visual under-conditioning in LMMs by regularizing the model’s visual conditioning policy rather than its answer agreement, requiring no annotations, metadata, or external reward models.

• We propose two complementary invariance-based reward signals: geometric invariance and semantic invariance via regional perturbation (ghosting), that jointly enforce spatial consistency and evidence sensitivity, providing a purely self-supervised objective defined entirely from the model’s own predictions.

• We empirically validate VISE across 18 benchmarks spanning image captioning, VQA, reasoning, and hallucination on multiple model scales and four backbone families, demonstrating consistent and substantial improvements over all self-evolving baselines with no tradeoffs between task groups.

2 Related Work↩︎

Early work on self-improving LMMs aimed to reduce reliance on human-annotated data through internal preference and alignment signals. Tan  [6] introduce image-driven self-questioning with diffusion-based rejection for DPO optimization. Wang  [7] propose an in-context self-critic with visual metrics to construct preference pairs without external models. Liu  [8] study multimodal self-evolution from a reinforcement learning perspective and introduce entropy-driven exploration to mitigate saturation. While these approaches reduce annotation dependence, they still rely on structured supervision or curated preference construction, limiting fully autonomous self-improvement.

More recent work has shifted to fully unsupervised self-play frameworks with internally generated rewards, yielding strong gains on structured reasoning benchmarks. EvoLMM [1] introduces a Proposer–Solver formulation optimized with continuous self-consistency rewards to enable co-evolution of question generation and reasoning without annotations. Subsequent extensions refine this paradigm: iReasoner [3] incorporates trajectory-aware rewards, VisPlay [2] promotes diversity and difficulty to prevent collapse, and Agent0-VL [5] integrates tool-grounded self-verification. C2-Evo [9] and DoGe [10] further address training instabilities through co-evolutionary data loops and role decoupling mechanisms.

Limitations. Despite their methodological diversity and strong performance on structured reasoning benchmarks, these approaches optimize answer correctness or reasoning consistency as the primary objective. This implicitly assumes that self-consistent outputs reflect improved visual understanding, which is an assumption that breaks down under visual under-conditioning, where a model can remain self-consistent and even correct while relying on statistical language priors rather than visual evidence. VISE addresses this directly by replacing answer-agreement rewards with invariance-based rewards that regularize the model’s visual-conditioning policy itself, operating within a single model on raw unlabeled images without specialist roles or external reward models.

3 Method↩︎

Problem Formulation. Let \(\mathcal{X} = \{x\}\) denote an unlabeled collection of images, with no accompanying queries, bounding box annotations, or category labels. At each training step, the model \(\pi_\theta\) operates in a self-questioning regime: it first generates a natural-language localization query \(q\) by interrogating image \(x\), then predicts a bounding box \(B = (x_1, y_1, x_2, y_2)\) locating the queried object under the same query. Both query generation and bounding box prediction are performed by the same single policy, without separate specialist roles, as a well-pretrained LMM already possesses sufficient visual knowledge to formulate meaningful queries for its own content. All spatial coordinates are represented in a normalized space \([0, S]^4\), where \(S = 1000\), such that a pixel-space coordinate \(c_\text{pix}\) along dimension \(D\) maps to \(\tilde{c} = (c_\text{pix} / D) \cdot S\).

We use the Qwen3-VL family [11] as the base backbone, freezing the vision encoder and updating the multimodal projector, feed-forward layers, and decoder attention projections. This is motivated by the nature of the failure mode: the vision encoder already produces strong visual representations, and the problem lies in how the decoder projects and utilizes them. Propagating noisy, unsupervised reward gradients into the encoder would risk destabilizing representations that are already high quality, without addressing the locus of failure.

Geometric Invariance Reward. Visual under-conditioning manifests most directly in localization instability: a decoder that disengages with visual image evidence will produce predictions that are inconsistent with the actual spatial structure of the scene, and in particular will fail to maintain coherent localization when the image undergoes a known geometric transformation. We exploit this as a self-supervised training signal. If the model correctly conditions its localization on what it sees, then its predicted box on a geometrically transformed image should correspond precisely to the analytically projected version of its prediction on the original. Deviation from this consistency is evidence of visual under-conditioning, and we penalize it directly.

At each training step, the model generates a query \(q\) by conditioning on image \(x\), producing a short natural-language description of a prominent and spatially unambiguous object in the scene. The model then predicts a bounding box \(B_\text{orig}\) on the image under query \(q\). A geometric transformation \(\mathcal{T}\) is sampled uniformly from three types: affine (rotation \(\theta \sim \mathcal{U}(-10^\circ, 10^\circ)\), scale \(s \sim \mathcal{U}(0.9, 1.1)\), translation \((\delta_x, \delta_y) \sim \mathcal{U}(-50, 50)^2\)), crop (ratio \(\rho \sim \mathcal{U}(0.8, 1.0)\), resized to original resolution), or horizontal flip. Each is described by a known \(3{\times}3\) homogeneous matrix \(M\). The model then predicts a second box \(B_\text{new}\) on the transformed image \(x' = \mathcal{T}(x)\) under the same query \(q\). The expected box under the transformation is computed by lifting the four corners of \(B_\text{orig}\) to homogeneous coordinates and applying \(M\) as \(\mathbf{c}'_i = M\mathbf{c}_i\); the axis-aligned box of the resulting corners gives the projected box \(B_\text{proj}\). The geometric invariance reward is:

\[\mathcal{R}_\text{geo} = \frac{\,\text{GIoU}(B_\text{proj},\, B_\text{new}) + 1\,}{2} \label{eq:r95geo}\tag{1}\]

where GIoU is the Generalized Intersection over Union [12]:

\[\text{GIoU}(B_1, B_2) = \text{IoU}(B_1, B_2) - \frac{|\mathcal{C}| - |B_1 \cup B_2|}{|\mathcal{C}|} \label{eq:giou}\tag{2}\]

with \(\mathcal{C}\) denoting the smallest axis-aligned box enclosing both \(B_1\) and \(B_2\). The linear normalization in Eq. (1 ) maps GIoU \(\in [-1, 1]\) to \(\mathcal{R}_\text{geo} \in [0, 1]\). This reward is maximized when the model’s localization on the transformed view agrees precisely with the geometric projection of its original prediction, and degrades smoothly as spatial consistency deteriorates. In this way, \(\mathcal{R}_\text{geo}\) directly targets the spatial dimension of visual under-conditioning: such a model cannot maintain such consistency across views and is therefore penalized, even if its individual predictions appear plausible in isolation.

Figure 4 provides representational evidence that \(\mathcal{R}_\text{geo}\) achieves its intended effect. We compute per-layer Centered Kernel Alignment (CKA) similarity [13] between representations of original and geometrically augmented views for the base model and VISE across 100 random COCO images. On Qwen3-VL-2B, gains are confined to the final decoder layers, localizing geometric under-conditioning to the stages where generation decisions are formed. On Qwen3-VL-4B, the \(\Delta\)CKA advantage is distributed more broadly across the decoder, consistent with the scale effects in Table 1: concentrated failures yield more direct downstream gains under invariance-based correction.

Semantic Invariance Reward. Geometric consistency is a necessary but insufficient condition for faithful visual conditioning. A model could achieve high \(\mathcal{R}_\text{geo}\) by predicting large, spatially stable regions without its predictions being meaningfully driven by the semantic content they enclose. Hence, we also address the complementary dimension of visual under-conditioning: evidence sensitivity. A model whose responses are conditioned on the image should recognize that removing the predicted region removes the evidence for the queried object. If instead the decoder is shortcutting from language priors, it would remain insensitive to that removal. We reward the opposite: the model should judge the object as visible when the region is intact, and as absent when it is obscured.

We do this by introducing a regional perturbation procedure termed ghosting. Given the predicted box \(B_\text{orig}\), the corresponding pixel region in the original image \(x\) is identified and its contents replaced by a Gaussian-blurred version with kernel \(\sigma = 25.0\), producing a ghosted image \(\tilde{x}\) in which the localized region is visually degraded while the surrounding context is fully preserved. The model then assesses the visibility of the queried object under \(q\) on both \(x\) and \(\tilde{x}\) via greedy decoding, yielding binary judgments \(v = \texttt{vis}(x, q) \in \{0, 1\}\) and \(\tilde{v} = \texttt{vis}(\tilde{x}, q) \in \{0, 1\}\). The semantic invariance reward is: \[\mathcal{R}_\text{sem} = \begin{cases} 1.0 & \text{if } v = 1 \;\text{ and }\; \tilde{v} = 0 \\ 0.0 & \text{otherwise} \end{cases} \label{eq:r95sem}\tag{3}\]

A prediction that is geometrically consistent but semantically arbitrary (enclosing a region that does not contain the queried object) receives \(\mathcal{R}_\text{sem} = 0\) even if it satisfies \(\mathcal{R}_\text{geo}\), ensuring the two signals are complementary and jointly necessary for the model to improve its evidence binding. Figure 3 provides direct evidence of the intended behavioral change. We measure generation-time visual attention, which is the fraction of attention each generated token assigns to image tokens across transformer layers, for the base model and VISE averaged over 100 random COCO images. Unlike prefill attention, which is dominated by the visual embedding layer, generation-time attention captures how the decoder references visual evidence while producing each token. VISE consistently assigns more attention to visual tokens across mid-to-late decoder layers on both model scales, indicating that the semantic invariance reward encourages the decoder to maintain visual conditioning throughout generation rather than reverting to language priors after initial image encoding.

Composite Reward and Optimization. The total reward combines both invariance signals as \(\mathcal{R}_t = \lambda_\text{geo} \mathcal{R}_\text{geo} + \lambda_\text{sem} \mathcal{R}_\text{sem}\), where \(\lambda_\text{geo} = \lambda_\text{sem} = 0.5\). At each step, the model produces a completion \(y\) containing the predicted box coordinates for \((x, q)\). We maintain an exponential moving average baseline \(b_t \leftarrow 0.9\,b_{t-1} + 0.1\,\mathcal{R}_t\) to reduce gradient variance, giving advantage \(A_t = \mathcal{R}_t - b_t\). Letting \(\Delta_t = \log p_\theta(y \mid x, q) - \log p_\text{ref}(y \mid x, q)\) denote a KL-like divergence proxy between the current policy and the frozen reference \(\pi_\text{ref}\), we optimize:

\[\mathcal{L}(\theta) \;=\; -\,A_t \cdot \log p_\theta(y \mid x, q) \;+\; \beta_t \cdot \Delta_t \label{eq:loss}\tag{4}\]

where the first term is a REINFORCE-style update and the second regularizes against the reference model. We adapt the KL Coefficient \(\beta_t\) dynamically to maintain a target divergence level:

\[\beta_{t+1} = \begin{cases} \beta_t\,(1 + \eta) & \text{if } |\Delta_t| > \tau \\ \beta_t\,(1 - \eta) & \text{otherwise} \end{cases} \label{eq:kl95adapt}\tag{5}\]

where \(\tau = 0.020\) is the target divergence budget, \(\eta = 0.10\) is the adaptation rate, and \(\beta_t\) is clipped below at \(10^{-6}\). This tightens regularization when the policy drifts beyond the target and relaxes it when updates are conservative, providing stable self-evolution without a fixed regularization strength.

4 Experiments↩︎

Implementation Details. We fine-tune each base model using LoRA [14] while keeping the vision encoder frozen. For the smaller models (2B and 4B), we use rank \(r{=}16\) and \(\alpha{=}32\), whereas for the larger models (8B and 32B), we increase the rank to \(r{=}32\) and \(\alpha\) to \(64\) (dropout \(0.05\) in all cases). Training uses the AdamW optimizer [15] with weight decay \(0.01\) and gradient clipping at \(1.0\). We set the learning rate to \(10^{-6}\) and the KL regularization target to \(0.020\) (adaptive rate \(0.10\)) for smaller models, and reduce them to \(1.5\times10^{-7}\) and \(0.004\) (adaptive rate \(0.15\)), respectively, for larger models. Equal reward weights of \(0.5\) are applied to both the geometric and semantic invariance terms. All models are trained for 4000 steps on 8\(\times\) AMD MI250X GPUs using bfloat16 precision. No question–answer pairs, annotations, metadata, or external reward models are used.

Training Data. We use 4000 raw, unlabeled images sampled from the COCO dataset [16], with no captions, bounding boxes, or semantic labels retained. Spatial transformations (affine, crop, and flip) are applied online during training to generate geometric invariance targets. Supp. Sec. 8 reports Objects365 training results showing consistent gains beyond COCO-specific image exposure.

Baselines and Evaluation. We compare against five self-evolving baselines on the same base models: VisPlay [2], EvoLMM [1], and iReasoner [3], which are fully unsupervised and require no external rewards or annotations, and VisionZero [4] (CLEVR, Chart, and Real-World [RW] variants), where only CLEVR is label-free, while Chart and Real-World use GPT-4o during dataset construction. All evaluations are run on AMD MI250X GPUs using the lmms-eval framework [17], with HuggingFace Transformers v4.38 and bfloat16 precision for consistency with training. We evaluate on four image captioning benchmarks (COCO [2014/2017 average] [16], NoCaps [18], Flickr30k [19], TextCaps [20]), twelve VQA and reasoning benchmarks (GQA [21], OK-VQA [22], VQAv2 [23], AI2D [24], ChartQA [25], InfoVQA [26], ScienceQA [27], MMMU [28], CaptionQA [29], RWQA [30], ESB [31], MMBench [32]), and two hallucination benchmarks (POPE [33] and COCO Cap Chair [34]).

Image Captioning.

In Table 1, we compare VISE against all baselines across four captioning benchmarks and four Qwen3-VL scales. Consistency-based methods trained on math and scientific images show a clear pattern at 2B: EvoLMM drops \(-0.70\) CIDEr on COCO, \(-0.77\) on NoCaps, and \(-0.94\) on Flickr30k, with iReasoner following a similar trend. This persists across scales, and cannot be explained by distribution mismatch alone; when answer agreement is the reward, the model is never required to describe what it actually sees, so captioning degrades. VisionZero-RW is the strongest captioning baseline at 2B (\(+4.04\) COCO CIDEr) owing to closer alignment with natural scenes, though its Chart and CLEVR variants show smaller and less consistent gains. VisPlay improves COCO by \(+2.31\) but regresses on NoCaps (\(-0.38\)) and TextCaps (\(-0.09\)), suggesting diversity and difficulty rewards prioritize question complexity over visual faithfulness. At larger scales, VisionZero variants recover further (e.g., RW reaches \(+8.41\) COCO CIDEr at 8B) while EvoLMM and iReasoner remain inconsistent, pointing to distribution proximity rather than visual conditioning as the reason.

VISE improves CIDEr on Qwen3-VL-2B by \(+16.85\) on COCO, \(+14.73\) on NoCaps, \(+16.55\) on Flickr30k, and \(+19.66\) on TextCaps with gains \(4{\times}\)–\(7{\times}\) larger than the strongest baseline on each benchmark and no regressions across any dataset or scale. Gains decrease with model size (\(+16.85\) at 2B to \(+8.72\) at 32B on COCO), in line with larger models entering training with stronger visual conditioning already consolidated during pretraining. Figure 5 shows what these numbers look like in practice. Where baselines either stay vague ("large animals near a river") or commit to plausible-but-wrong details ("wolves," "obelisk"), VISE reads the image: it catches that the animals are three bears of different sizes walking in order, that there is a hand-drawn figure on the car window, and that the square is Trafalgar rather than a generic city landmark. The gap between these outputs is precisely what the CIDEr gap in Table 1 reflects.

VQA and Reasoning. Table 2 reports performance across twelve VQA and reasoning benchmarks. The baseline results exhibit a consistent generalization tradeoff that is not observed with VISE. VisionZero-Chart improves ChartQA by \(+0.96\) at 2B but drops \(-0.45\) on OK-VQA and \(-0.50\) on CaptionQA; VisionZero-RW shows the reverse, gaining on natural-language tasks (\(+0.90\) RWQA) but dropping on ChartQA (\(-0.47\)); VisionZero-CLEVR improves structured reasoning (\(+2.54\) ScienceQA, \(+1.91\) InfoVQA) while remaining inconsistent on open-ended VQA. This bidirectional pattern suggests that domain-specific consistency training picks up the statistical regularities of the training distribution alongside any conditioning signal, imposing a ceiling that cannot be lifted without changing the reward itself. EvoLMM and iReasoner follow the same logic: strong gains on ScienceQA (\(+3.59\), \(+3.70\) at 2B) but drops on OK-VQA (\(-2.73\), \(-2.63\)) and VQAv2 (\(-0.51\), \(-0.43\)), in line with the captioning dips in Table 1.

VISE improves all twelve benchmarks simultaneously at 2B, with gains of \(+4.19\) on ScienceQA, \(+2.41\) on InfoVQA, \(+1.75\) on MMMU, and \(+2.12\) on CaptionQA, and no regressions at any scale. At 4B, the MMMU gain grows to \(+3.72\), the largest improvement on that benchmark across all methods and scales, indicating that stronger conditioning particularly benefits tasks requiring multi-discipline visual reasoning. Across all four scales, VISE exhibits no structured-versus-open-ended tradeoff observed in the baselines. Instead of adapting to a specific domain, the invariance reward improves the model’s visual conditioning behavior, and these improvements generalize across task formats.

Hallucination. Table 3 evaluates all methods on POPE and COCO Cap Chair. VisPlay increases hallucination at 2B (Chair-I \(+0.003\), Chair-S \(+0.23\)), consistent with diversity rewards prioritizing question complexity over visual fidelity. EvoLMM and iReasoner reduce Chair scores modestly (\(-0.23\), \(-1.74\)) but drop POPE accuracy (\(-1.42\), \(-1.31\)) simultaneously. Sentence-level hallucinations decline while binary object-presence reliability weakens, pointing to inconsistent rather than substantive improvement. VisionZero-RW is the strongest baseline (Chair-I \(-2.99\), Chair-S \(-4.05\)), due to broader real-world visual coverage. VISE reduces Chair-I by \(-5.00\) and Chair-S by \(-5.45\) at 2B, surpassing the best baseline by \(+2.01\) and \(+1.40\) while also improving POPE by \(+1.02\). This joint improvement across both metrics is unique to VISE: penalizing confident predictions under regional perturbation discourages generating objects that are statistically plausible but not visually present. Gains attenuate with model size (Chair-I \(-0.37\), \(-0.44\) at 8B, 32B), in-line with the captioning trend in Table 1.

Effect of Model Scale. Gains from VISE are largest at smaller scales and attenuate with model size, consistent across all task groups. We attribute this to a capacity ceiling effect in the self-evolving setting: larger models enter post-training with stronger conditioning consolidated during pretraining and instruction tuning, leaving less headroom for invariance-based correction. This aligns with prior observations that self-evolving methods exhibit diminishing returns at scale [2], and suggests that invariance rewards are most effective where visual under-conditioning remains pronounced, particularly in smaller models that have not yet developed robust evidence-binding behavior.

Backbone Generalization. Table 4 evaluates VISE on four architecturally diverse backbones trained under an identical setup using the same 4000 unlabeled COCO images. Captioning and hallucination gains are consistent across families, with NoCaps showing the largest absolute improvements. Reasoning and VQA also improve without regressions, including on Llama-3.2-11B (\(+1.31\) POPE) despite its weaker baseline grounding, indicating that the reward is insensitive to pretraining distribution. The consistency of improvements across architectures confirms that the invariance reward is architecture-agnostic and that visual under-conditioning is a general phenomenon addressed by VISE regardless of backbone.

4pt

Ablation Study. Table 5 isolates the contribution of each reward component on Qwen3-VL-2B and Qwen3-VL-8B. Training with \(R_\text{geo}\) yields moderate captioning gains (\(+4.83\) COCO CIDEr, \(+4.33\) Flickr30k at 2B) and modest hallucination reductions (Chair-I \(-1.35\), Chair-S \(-1.45\)), highlighting that spatial consistency provides a meaningful visual signal. However, because it cannot penalize visually unsupported but plausible generations, its improvements remain well below those of the full model. \(R_\text{sem}\) accounts for most of VISE’s gains, delivering \(+13.99\) COCO CIDEr and reducing Chair-I by \(-4.15\) and Chair-S by \(-4.45\) at 2B. This is consistent with our findings that hallucination reduction stems directly from the semantic reward design. By penalizing confident predictions under regional perturbation, the ghosting signal emerges as the primary driver of the captioning and hallucination gains in Tables 1 and 3.

Full model (\(R_\text{geo} + R_\text{sem}\)). Combining both rewards yields additional and consistent gains, with \(R_\text{geo}\) contributing complementary benefits to \(R_\text{sem}\). At 2B, the full model adds \(+2.86\) COCO CIDEr and a further \(-0.85\) Chair-I reduction beyond \(R_\text{sem}\) alone. The same pattern holds at 8B: while \(R_\text{sem}\) dominates captioning (\(+6.26\) vs.\(+2.83\) COCO CIDEr), the combined model delivers broader improvements across VQA and reasoning tasks. Together, these results show that spatial consistency and evidence sensitivity address distinct facets of visual under-conditioning and are both necessary for the full benefits of VISE. Supp. Sec. 8 further validates this interpretation: a random-reward control remains near base captioning performance, confirming that the gains come from the invariance rewards rather than fine-tuning alone.

5 Conclusion↩︎

We introduced VISE, a fully unsupervised self-evolving framework for large multimodal models that directly addresses visual under-conditioning without relying on human annotations, external reward models, or multi-role formulations. By training within a single model using geometric consistency under spatial transformations and semantic sensitivity under regional perturbation, VISE encourages the decoder to pay more attention to visual tokens rather than relying on statistical language priors. Our experiments show consistent gains across captioning, VQA, reasoning, and hallucination benchmarks with no task tradeoffs, and these improvements hold across four model scales and four architecturally diverse backbones. Ablations confirm that semantic invariance drives most gains while geometric invariance contributes complementary improvements, together covering distinct dimensions of the failure mode. These results highlight that answer-consistency rewards are insufficient for genuine visual improvement: directly increasing attention to visual tokens during decoding is both necessary and sufficient to produce broad, robust gains. This work suggests a promising direction for self-evolving multimodal training, shifting the objective from output agreement to evidence-conditioned generation.

6 Acknowledgement↩︎

The computations were enabled by resources provided by LUMI hosted by CSC (Finland) and LUMI consortium, and by Berzelius resource provided by the Knut and Alice Wallenberg Foundation at the NSC.

7 Hyperparameter Sensitivity↩︎

Table 6 reports how VISE performs under different reward weight ratios and KL divergence targets on Qwen3-VL-2B and 8B. The results are stable across both axes. Varying the \(\lambda\) ratio between the geometric and semantic rewards (\(0.75/0.25\), \(0.50/0.50\), \(0.25/0.75\)) produces differences well under 0.5 CIDEr on captioning and under 0.3 on Chair-I, with no clear winner across all benchmarks simultaneously. The equal-weight default lies near the center of this range and was selected on principled grounds rather than tuned on evaluation data. One pattern worth noting is that shifting weight toward \(\mathcal{R}_\text{sem}\) tends to slightly improve reasoning metrics while marginally worsening hallucination, and the reverse holds for \(\mathcal{R}_\text{geo}\), which is consistent with the complementary roles each reward plays as described in the ablation.

The KL target \(\tau\) shows similarly low sensitivity. Tightening to \(\tau = 0.010\) slightly improves hallucination metrics (Chair-I \(-0.29\) relative to default at 2B) at the cost of marginally reduced captioning gains, while relaxing to \(\tau = 0.050\) has the opposite effect, allowing slightly more policy drift and producing small improvements on reasoning at the expense of hallucination. Neither direction degrades performance meaningfully, suggesting the adaptive KL mechanism is doing its job of keeping updates stable regardless of the target value. Taken together, these results indicate that VISE is not sensitive to precise hyperparameter choices within reasonable ranges, and that the reported gains are a robust property of the invariance-based training objective rather than an artifact of careful tuning.

4pt

8 Additional Validation Experiments↩︎

Here, we provide targeted supplementary validations for the main experimental claims: training-domain robustness, the choice of LoRA over full fine-tuning, reward causality, transformation and ghosting design choices, and training efficiency.

8.1 Training Domain and COCO Split Separation↩︎

Table 7 reports Qwen3-VL-8B results over three seeds. VISE is trained either on COCO or on Objects365 images, always without captions, boxes, or category labels. The COCO training images are taken from train2014/train2017 and evaluated on disjoint validation splits with zero image overlap. Training on Objects365 gives nearly identical gains, showing that the improvements are not due to COCO-specific image exposure.

4pt

8.2 LoRA versus Full Fine-Tuning↩︎

Table 7 also compares full fine-tuning (FFT) with LoRA under the same training domains. FFT improves over the base model but consistently underperforms LoRA across captioning, hallucination, VQA, and reasoning metrics. This supports our training design: noisy unsupervised encoder updates can hurt visual conditioning, whereas updating the cross-modal and decoder components is sufficient and more stable.

8.3 Reward Causality via Random-Reward Control↩︎

r0.55

5pt

To isolate the role of the invariance rewards from generic exposure to unlabeled images, we train a random-reward control using the same images, optimization setup, and policy update, but replace the VISE reward with \(\mathcal{R}\sim\mathcal{U}(0,1)\). Table 8 shows that random rewards stay near base captioning performance, while VISE produces large gains. This confirms that the improvements are driven by the geometric and semantic invariance signals rather than by fine-tuning alone.

8.4 Transformation and Ghosting Design Choices↩︎

Table 9 evaluates design choices in the two invariance branches on Qwen3-VL-2B. Moderate transformations such as affine, crop, and flip all improve over the base model, while overly large affine perturbations degrade performance due to degenerate border boxes. For semantic perturbation, Gaussian ghosting performs best: the default \(\sigma=25\) kernel induces the intended visible-to-not-visible change, while weaker blur, stronger blur, zero masking, and Gaussian noise underperform.

5pt

Noisy localization steps are naturally down-weighted by the reward design: failed localization gives near-zero REINFORCE advantage \(A_t=R_t-b_t\approx0\), while inconsistent visibility judgments yield \(\mathcal{R}_\text{sem}=0\). Thus, imperfect self-generated boxes do not provide a strong positive learning signal.

8.5 Training Efficiency↩︎

Each VISE step uses seven forward passes: query generation, box prediction on the original and transformed views, visibility prediction on the original and ghosted views, and policy/reference log-probability evaluation. Training Qwen3-VL-2B for 4000 steps takes 16 hours on 8\(\times\) AMD MI250X GPUs using bfloat16, about \(\sim2\times\) faster to converge than multi-role self-evolving baselines such as EvoLMM.

9 Prompts Used in VISE Training↩︎

VISE uses three distinct prompts at each training step, one for each stage of the self-supervised pipeline. We describe each below and provide the corresponding prompt in the figures that follow.

9.1 Prompt Used for Grounding Query Generation↩︎

At the start of each training step, the model is prompted to generate a natural-language grounding query \(q\) for the input image \(x\). The prompt instructs the model to identify a single prominent, spatially unambiguous object in the scene and describe it concisely. The generated query is then used as input for both the bounding box prediction and the semantic visibility verification stages within the same step. No external query bank, template, or category list is used — the query is produced entirely by the model itself from the raw image.

9.2 Prompt Used for Bounding Box Prediction↩︎

Given the image \(x\) (or its geometrically transformed version \(x' = \mathcal{T}(x)\)) and the generated query \(q\), the model is prompted to predict a bounding box localizing the queried object. The prompt specifies the normalized coordinate space \([0, 1000]^4\) and the expected output format. This same prompt structure is used for both the original and transformed views, ensuring that any difference in predicted boxes reflects the model’s visual grounding behavior rather than prompt variation.

9.3 Prompt Used for Semantic Visibility Verification↩︎

After predicting \(B_\text{orig}\), the model is prompted to assess whether the queried object is clearly visible in both the original image \(x\) and the ghosted image \(\tilde{x}\). The prompt asks for a binary yes/no judgment and is kept deliberately minimal to avoid leading the model toward a particular answer. The semantic invariance reward \(\mathcal{R}_\text{sem}\) is computed from the pair of visibility judgments returned by this prompt, as described in the Method Section of the main paper.

10 Extended Qualitative Results↩︎

10.1 Generation-Time Visual Attention↩︎

Figures 7–11 show additional per-sample, generation-time visual attention comparisons between the base Qwen3-VL-2B model [11] and VISE across a range of scene types using the prompt, "What is happening in this scene?" For each example, we plot the fraction of attention allocated to image tokens at each decoder layer during generation, alongside the corresponding text outputs from both models. The attention advantage of VISE is consistent across all examples shown, concentrating in the mid-to-late decoder layers where semantic generation decisions are made, and is accompanied by noticeably more specific and visually grounded output text.

10.2 Image Description Comparisons↩︎

Figures 12 to 15 provide additional qualitative comparisons of image descriptions produced by VISE and four self-evolving baselines: VisionZero (RealWorld) [4], VisPlay [2], EvoLMM [1], and iReasoner [3]. All methods are evaluated on the same images under the same prompts using their respective Qwen3-VL-2B checkpoints. Baseline methods tend to produce descriptions that are either vague and category-level, relying on statistically common scene descriptions, or occasionally confident but incorrect about specific visual details. VISE consistently produces more fine-grained and accurate descriptions, correctly identifying object-level details such as clothing attributes, vehicle types, spatial relationships, and scene-specific context that the baselines miss or misattribute.

10.3 Per-Sample Generation-Time Attention Breakdown↩︎

Figures 16–18 and Figures 19–21 show per-sample generation-time attention breakdowns for randomly selected samples from the COCO dataset, evaluated on Qwen3-VL-2B and Qwen3-VL-4B respectively. Each figure shows the input image, a per-layer line plot of the visual token attention ratio during generation, and a Token\(\times\)Layer heatmap comparing the base model (top) and VISE (bottom) side by side. The line plots show that VISE consistently allocates a higher fraction of attention to image tokens throughout generation, with the advantage most pronounced in mid-to-late layers where semantic content is produced. The heatmaps make this difference concrete at the token level: VISE shows broader and more intense red regions across both the layer and token axes, indicating that individual generated tokens are more strongly anchored to visual evidence. This pattern holds across all six samples and both model scales, and is particularly strong in cases requiring spatial layout description and fine-grained visual detail. All samples are evaluated using the prompt: “Describe in detail the spatial layout and positions of all objects in this image.”

References↩︎