Visual Re-Examination in Multimodal Reasoning

• Visual Re-Examination (VRE) is a mechanism where models iteratively recheck visual input to ensure grounding in image evidence during multi-step reasoning.
• It encompasses techniques from reinforced visual attention to structured checkpoints like Chain-of-Questions and forensic inspection protocols.
• Empirical studies reveal that while VRE can bolster reasoning accuracy, models often show illusory reflection, prompting advanced training and benchmark designs.

Searching arXiv for the cited VRE papers to ground the article. Visual Re-Examination (VRE) denotes the recurrent re-consultation of visual evidence during a multi-step reasoning process, so that intermediate and late-stage inferences remain grounded in the image rather than collapsing into self-referential text generation. In recent vision-LLM literature, closely related terms include visual reflection, visual re-checking, and visual introspection. Across these formulations, the central claim is stable: long-form multimodal reasoning is reliable only when the model can “actively verify and refine its reasoning based on the visual input,” rather than merely revising text with more text (Jian et al., 15 Sep 2025). Subsequent work has turned this notion into a training framework, a prompting protocol, a diagnostic probe, and a benchmark design principle, while also exposing a major controversy: many models can say that they are “checking the image again” without actually reallocating computation to visual evidence (Lv et al., 27 Mar 2026).

1. Conceptual scope and relation to multimodal reasoning

In contemporary VLM research, VRE is primarily a response to the failure of naïvely transferring text-only “slow thinking” to multimodal models. The critique is that caption-then-reason pipelines often produce textual reflection: the model revises its chain-of-thought based on its own prior text, not on the image. By contrast, visual reflection is defined as the point at which the model “actively verifies and refines its reasoning based on the visual input,” which is exactly the operational meaning of VRE in long-chain visual reasoning (Jian et al., 15 Sep 2025).

This framing has two immediate consequences. First, VRE is not equivalent to producing longer chains-of-thought. Long responses can intensify what has been described as visual forgetting, where the computation becomes dominated by language context even though visual tokens remain in the sequence (Jian et al., 15 Sep 2025). Second, VRE is not restricted to answer generation. It can also be cast as a decision process over what to inspect next. In "V-REX: Benchmarking Exploratory Visual Reasoning via Chain-of-Questions" (Fan et al., 12 Dec 2025), multi-step exploratory reasoning is represented as a Chain-of-Questions (CoQ), where each intermediate question-answer pair collects evidence needed for the final answer. In that setting, selecting the next question is a discrete form of visual re-examination: the model decides what part of the scene to re-check and in what order.

The idea also appears in prompt-driven forensic reasoning. "REVEAL -- Reasoning and Evaluation of Visual Evidence through Aligned Language" (Praharaj et al., 18 Aug 2025) frames image forgery detection as a prompt-driven visual reasoning task in which a model systematically re-checks an image against generalized constraints such as lighting, reflections, perspective, semantics, and realism. Here, VRE takes the form of a structured checklist plus optional region-wise anomaly inspection.

A broader implication is that VRE is less a single architecture than a family of mechanisms for enforcing repeated grounding. Depending on the paper, that mechanism may be cross-modal attention maintenance, an explicit reflection block, a chain of sub-questions, a forensic checklist, or a benchmark protocol that tests whether the model truly uses the image when it claims to do so (Jian et al., 15 Sep 2025).

2. Mechanistic interpretation and quantitative diagnostics

A major line of work treats VRE as a measurable property of the decoding process. "Look Again, Think Slowly: Enhancing Visual Reflection in Vision-LLMs" (Jian et al., 15 Sep 2025) defines a visual attention weight from response tokens to visual tokens and shows that it decays sharply as more tokens are generated. It also defines a visual dependency measure (VDM) based on the divergence of next-token distributions with and without visual tokens. On MMMU, after roughly 300 generated tokens, attention to visual tokens falls to 20–30% of the initial level, and the VDM drops correspondingly, indicating that later tokens are essentially independent of whether the image is present (Jian et al., 15 Sep 2025).

That paper formalizes attention from a response token $n$ to a visual token $j$ in head $h$ as

$$a^{(h)}_{n j} = \operatorname{softmax}_{j} \left(\frac{q^{(h)}_{n} \cdot k^{(h)}_{j}}{\sqrt{d}}\right),$$

and aggregates it into a token-level visual attention statistic. It also defines VDM through the Hellinger distance between next-token distributions with and without image tokens. The empirical point is not merely that attention decays, but that causal dependence on the image decays as well (Jian et al., 15 Sep 2025).

A more adversarial diagnostic appears in "Are VLMs Seeing or Just Saying? Uncovering the Illusion of Visual Re-examination" (Shi et al., 15 May 2026). That work introduces VisualSwap, where an original image $I_a$ is replaced by a visually similar but semantically different image $I_b$ after a model has already produced reasoning on $I_a$. If a model genuinely re-examines the image when prompted with phrases such as “Wait, let me check the figure again,” its continuation should adapt to $I_b$. In practice, models overwhelmingly miss the swap, with accuracy dropping by up to 60%, and thinking models are nearly 3x more vulnerable than their instructed counterparts (Shi et al., 15 May 2026). The paper formalizes the degradation as

$$\Delta = \text{Acc}_{\text{base}} - \text{Acc}_{\text{probe}}.$$

The same paper defines a Visual Attention Score

$$s_l(t) = \frac{1}{H} \sum_{h=1}^{H} \sum_{v \in V} A_{l,h}(t, v),$$

and shows that user-issued multi-turn instructions substantially elevate attention to visual tokens, whereas self-generated reflective phrases do not (Shi et al., 15 May 2026). This is one of the clearest operational distinctions in the literature between genuine and illusory VRE.

"Reflect to Inform: Boosting Multimodal Reasoning via Information-Gain-Driven Verification" (Lv et al., 27 Mar 2026) adds a complementary diagnostic. It measures the conditional entropy of reflection trajectories,

$j$0

and interprets higher entropy in the <reflection> segment as evidence that the model is injecting non-redundant information rather than paraphrasing its prior text. This suggests that VRE should be evaluated not only by accuracy or attention alone, but by whether a reflection phase contributes new, visually grounded information to the reasoning trace (Lv et al., 27 Mar 2026).

3. Training paradigms for inducing VRE

A first explicit training recipe appears in "Look Again, Think Slowly: Enhancing Visual Reflection in Vision-LLMs" (Jian et al., 15 Sep 2025). Reflection-V is not a new network topology; it is a training strategy applied to Qwen2.5-VL-7B-Instruct and a 3B variant. The core design has two parts: vision-centered reasoning data for cold-start learning, and a visual attention based reward model for reinforcement learning.

The cold-start data are produced by a multi-agent loop. A visual requester LLM receives the question and current textual context but cannot see the image; it emits a JSON-like query for a visual responder VLM that can see the image and returns a concise visual description; a summarizer then attempts to solve the problem from the accumulated context. If the summarizer is correct in the first interaction, the sample is filtered out as “non-reflective.” Only examples requiring at least two visual queries survive. A final cohesion-enhancement pass rewrites the multi-round log into a coherent chain-of-thought with connective phrases such as “Let’s double check” and “Let’s check the image again” (Jian et al., 15 Sep 2025). This procedure teaches when re-examination is useful and associates those textual anchors with spikes in visual attention.

The RL phase uses GRPO and adds a reward for sustained late-stage attention to visual tokens. With

$j$1

the total reward is

$j$2

This explicitly favors correct trajectories that continue to attend to the image in the second half of reasoning (Jian et al., 15 Sep 2025).

A second, more explicitly self-evolving approach is "Reflect to Inform: Boosting Multimodal Reasoning via Information-Gain-Driven Verification" (Lv et al., 27 Mar 2026). There, VRE is defined as a training framework that enables MLLMs to autonomously perform visual introspection during reasoning without additional visual inputs. The model first generates ordinary multimodal reasoning, then a <reflection> block, then <answer>. The training pipeline partitions rollout data into Stable, Unstable, and Intractable samples according to pass rate; synthesizes Elaborative Reflection, Corrective Reflection, and Recheck Reflection traces from the model’s own outputs; fine-tunes on those traces; then applies RLVR with a reward

$j$3

where the reflection term is tied to information gain rather than mere verbosity (Lv et al., 27 Mar 2026).

A third paradigm is training-free latent intervention. "Revis: Sparse Latent Steering to Mitigate Object Hallucination in Large Vision-LLMs" (Wu et al., 12 Feb 2026) interprets VRE as internal re-activation of suppressed visual information. REVIS constructs a purified visual direction orthogonal to a hallucination-prone language-prior direction and performs sparse intervention at a single calibrated layer: $j$4 This is described as a latent-space equivalent of “look at the image again, but don’t rely on your textual biases to fill in details” (Wu et al., 12 Feb 2026). It is not reflection in the textual sense, but it is an internal VRE primitive.

4. Prompting protocols, benchmarks, and evaluation settings

Prompt-based VRE is exemplified by REVEAL. Its Holistic Scene-Level Evaluation asks a model to score an image along eight forensic dimensions—lighting/shadows, reflections/transparency, perspective/geometry, repetition/patterns, edge/boundary consistency, contextual/semantic consistency, anomaly/artifact detection, and human/object realism—each with a Likert score from 1 to 5 and a short explanation (Praharaj et al., 18 Aug 2025). Its Region-wise Anomaly Detection overlays a 3×3 labelled grid on the image and asks the model to first provide a holistic assessment and then inspect each cell individually, explicitly referring to cells by number. In this formulation, VRE is a guided checklist plus coarse localization.

V-REX turns VRE into an evaluable multi-step exploration problem. It contains 702 multi-step problems, 2–6 steps each (avg 3.57), across 4 reasoning categories and 15 scenarios, and decomposes capability into Planning and Following (Fan et al., 12 Dec 2025). Planning chooses the next helpful question in a Chain-of-Questions; Following answers a curated question sequence. The benchmark defines

$j$5

This operationalizes VRE as a controlled sequence of re-examinations rather than a single answer (Fan et al., 12 Dec 2025).

VISREAS addresses a different but adjacent requirement: a model should verify whether a question is answerable with respect to the image before answering it. The task maps $j$6 to a validity label and an answer, and the dataset contains 2.07M semantically diverse queries generated automatically using Visual Genome scene graphs (Akter et al., 2024). In its pseudocode-based reasoning pipeline, any function can return NONE if an object, attribute, or relation is absent; if any intermediate step returns NONE, the final answer becomes “the question itself is problematic” (Akter et al., 2024). This makes answerability checking itself a form of VRE.

ReVSI extends the same logic to benchmark construction for 3D spatial intelligence. It argues that each QA pair must be answerable and correct under the model’s actual visual inputs, re-annotates objects and geometry across 381 scenes from 5 datasets, and provides frame-budget-specific variants at 16/32/64/all (Zhang et al., 27 Apr 2026). In that work, VRE is applied to the benchmark itself: the evaluation pipeline is re-examined so that models are not scored against labels unsupported by the frames they actually receive.

5. Empirical findings, gains, and controversies

Reflection-V reports that baseline VRMs’ attention to visual tokens decays sharply to ~20–30% of its initial magnitude by token 300, while RL-enhanced models using only textual rewards show even faster decay. Reflection-V-7B starts with higher visual attention, decays more slowly, and retains ~30–40% of initial visual attention after several hundred tokens, where baselines drop below 10% (Jian et al., 15 Sep 2025). It also outperforms Qwen2.5-VL-7B substantially across all reported benchmarks and, as a 7B model, beats GPT-4o and InternVL-2.5-38B on MathVision and MathVista while surpassing InternVL-2.5-38B and LLaVA-OneVision-72B on MMMU and M3CoT (Jian et al., 15 Sep 2025).

The self-evolving VRE framework in "Reflect to Inform" reports consistent improvements across visual reasoning and perception-heavy tasks. On its base Qwen2.5-VL-7B setup, MathVista rises from 68.2 to 71.2, MathVerse from 49.2 to 53.1, WeMath from 62.1 to 68.7, V*-Bench overall from 76.4 to 83.8, ChartQA overall from 83.7 to 88.8, OCRBench_v2 English from 56.3 to 62.6, and Chinese from 57.2 to 64.7 (Lv et al., 27 Mar 2026). The paper attributes the largest reasoning gains to RLVR and the final stabilization gains to Post-RL self-distillation.

REVIS shows that latent-space VRE can reduce hallucinations without degrading general reasoning. On Qwen2.5-VL, CHAIR$j$7 drops from 31.00% to 25.00%, which the paper describes as a 19.35% relative reduction in sentence-level hallucination, while MME rises from 2327.52 to 2345.30 and MM-Vet from 70.18 to 72.16 (Wu et al., 12 Feb 2026). This suggests that VRE need not always be implemented as an explicit textual reflection block.

The strongest controversy concerns whether self-reflective language is trustworthy evidence of actual re-examination. VisualSwap answers largely in the negative. Across Qwen3-VL, Kimi-VL, and ERNIE-VL families, models overwhelmingly miss the image swap, with accuracy dropping by up to 60%; thinking models are nearly 3x more vulnerable than their instructed counterparts; and scaling offers no mitigation (Shi et al., 15 May 2026). Yet explicit multi-turn user instructions restore visual grounding, and attention analysis shows why: user instructions substantially elevate attention to visual tokens, whereas self-reflection does not (Shi et al., 15 May 2026). The controversy is therefore not whether VLMs can re-examine, but whether autonomous textual self-reflection reliably triggers that capability.

Benchmark validity is another contested area. ReVSI shows that open models’ performance can drop sharply on a corrected, frame-budget-aware benchmark relative to prior spatial reasoning evaluations, and that some fine-tuned spatial VLMs hallucinate catastrophically on dummy videos where humans would answer 0 with 100% accuracy (Zhang et al., 27 Apr 2026). This indicates that apparently strong VRE behavior may partly reflect benchmark bias or label noise rather than grounded visual reasoning.

6. Limitations, misconceptions, and future directions

Several limitations recur across the literature. Reflection-V explicitly notes no architectural memory for re-examination: the image is encoded once, and re-examination relies purely on softattention, which may be insufficient in very long contexts (Jian et al., 15 Sep 2025). Its visual attention reward is also a proxy: high attention does not guarantee correct interpretation, and failed trajectories do not directly penalize spurious attention patterns (Jian et al., 15 Sep 2025). VRE’s self-evolving RL framework similarly shows sensitivity to over-optimization: performance improves up to about 100 RL steps and then degrades beyond about 150 steps (Lv et al., 27 Mar 2026). REVIS, despite being training-free, depends on the base model already having usable perceptual representations; it cannot create missing visual information de novo (Wu et al., 12 Feb 2026).

A common misconception is that reflective wording implies reflective computation. VisualSwap directly contradicts that assumption: phrases like “Wait, let me check the figure again” are often just learned textual patterns (Shi et al., 15 May 2026). Another misconception is that more chain-of-thought is necessarily better. Multiple papers instead argue that longer reasoning can amplify textual inertia, causing the model to drift farther from image evidence unless VRE is explicitly trained or externally triggered (Jian et al., 15 Sep 2025).

Future directions stated across the papers include replaying or refreshing visual tokens with learned triggers rather than fixed intervals, integrating OCR, segmentation, or symbolic solvers into the reflection loop, extending VRE to video and dynamic scenes, and designing reward functions that depend on mutual information or more fine-grained process verification rather than attention alone (Jian et al., 15 Sep 2025). ReVSI points toward semi-automated visibility-aware benchmark construction, while V-REX suggests learning explicit exploratory policies for question selection in a finite decision space (Fan et al., 12 Dec 2025). This suggests that the next phase of VRE research will likely combine three threads: stronger internal triggers for re-grounding, more diagnostic evaluations that distinguish genuine from illusory reflection, and hybrid systems that couple latent or attention-based reactivation with explicit, verifiable inspection steps.