Source-linked AI summary
On Robustness and Chain-of-Thought Consistency of RL-Finetuned VLMs
Rosie Zhao, Anshul Shah, Xiaoyu Zhu, Xinke Deng, Zhongyu Jiang, Yang Yang, Joerg Liebelt, Arnab Mondal
TL;DR
RL-finetuned VLMs can achieve stronger visual reasoning while remaining fragile to misleading text and unreliable in their accompanying CoT. The paper stress-tests these models with controlled perturbations and analyzes finetuning and reward designs, finding an accuracy-faithfulness trade-off and persistent difficulty jointly achieving robustness and faithful reasoning.
Problem
RL-tuned VLMs improve visual reasoning benchmarks, but evidence remains limited about their robustness to misleading textual context and the faithfulness of their reasoning traces.
Method
The paper applies controlled misleading captions and CoT perturbations to eight visual reasoning benchmarks, compares open and closed models, and studies RL finetuning, augmentation, and faithfulness-aware rewards.
Results
Controlled perturbations expose brittleness and CoT inconsistency; closed models are more robust, while RL finetuning improves accuracy but can erode faithfulness, and combined interventions can induce shortcut strategies.
Takeaways & Limitations
Reliable multimodal reasoning requires evaluation and training that jointly target correctness, robustness, calibration, visually grounded conflict resolution, and reasoning-answer consistency.
Takeaways & Limitations
The paper’s faithfulness metric measures external reasoning-answer consistency rather than whether the rationale reflects the model’s internal decision-making process.
Abstract
from arXiv · showhide
Reinforcement learning (RL) finetuning has become a key technique for enhancing large language models (LLMs) on reasoning-intensive tasks, motivating its extension to vision-language models (VLMs). While RL-tuned VLMs improve on visual reasoning benchmarks, they remain vulnerable to weak visual grounding, hallucinations, and over-reliance on textual cues. We show that simple, controlled textual perturbations, including misleading captions or incorrect chain-of-thought (CoT) traces, cause substantial drops in robustness and confidence, and that these effects are more pronounced when CoT consistency is taken into account across open-source multimodal reasoning models. In contrast, closed models exhibit similar failure modes but maintain markedly greater robustness and reasoning consistency, suggesting that the gap reflects a shortcoming in current open-source RL finetuning rather than an inherent limitation of the task. To better understand these vulnerabilities, we further analyze RL finetuning dynamics and uncover an accuracy-faithfulness trade-off: finetuning raises benchmark accuracy, but can simultaneously erode the reliability of the accompanying CoT and its robustness to contextual shifts. Although adversarial augmentation improves robustness, it does not by itself prevent faithfulness drift. Incorporating a faithfulness-aware reward can restore alignment between answers and reasoning, but when paired with augmentation, training risks collapsing onto shortcut strategies and robustness remains elusive. Together, these findings highlight the limitations of accuracy-only evaluations and motivate training and assessment protocols that jointly emphasize correctness, robustness, and the faithfulness of visually grounded reasoning.
1 Introduction
The paper tests whether RL-finetuned VLMs remain visually grounded under misleading textual context, finding brittleness, faithfulness drift, and greater robustness in closed models. It evaluates controlled perturbations, RL finetuning dynamics, and reward-based methods to examine correctness, robustness, and reasoning-answer consistency together.
- Motivation: RL-finetuned VLMs improve visual reasoning but remain vulnerable to weak visual grounding, hallucinations, brittle chain-of-thought, and misleading textual context.These vulnerabilities matter because visually grounded systems must preserve basic capabilities under benign textual variations.
- Contributions: Controlled misleading captions and conflicting chain-of-thought traces across eight visual reasoning benchmarks expose accuracy drops, answer-CoT inconsistencies, and altered uncertainty.The entropy-based analysis reveals distinct calibration and robustness profiles across models and training setups.
- Contributions: RL finetuning creates an accuracy-faithfulness trade-off: models gain headline accuracy with more training steps while becoming increasingly inconsistent in their reasoning.This disconnect shows why benchmark accuracy alone does not fully characterize reasoning quality.
- Model comparison: Closed-source models share hallucinations, overthinking, and accuracy drops but generally produce more consistent CoT, sometimes acknowledging misleading context or self-correcting.Their greater robustness suggests the observed gap reflects underdeveloped open-model capabilities rather than an inherent task limitation.
- Training dynamics: Adversarial augmentation improves robustness, but accuracy gains do not prevent faithfulness drift, which remains decoupled from robustness improvements.Maintaining data diversity helps robustness without guaranteeing reliable reasoning-answer alignment.
- Training dynamics: Faithfulness-aware rewards can align reasoning with answers, but combining them with augmentation can produce unstable dynamics and limited robustness gains.The result highlights the difficulty of jointly enforcing robustness and faithfulness in current VLM training regimes.
2 Evaluating Robustness of Vision-Reasoning Models
Textual perturbations substantially affect visual-reasoning performance, confidence, and chain-of-thought faithfulness. Closed models are generally more robust and faithful than open-source models, but still hallucinate and can misprioritize conflicting modalities.
- Robustness to textual perturbations: Misleading captions and incorrect thinking strings substantially reduce performance, while disclaimers often recover performance toward baseline.These perturbations also expose differences between models that appear similar under unperturbed conditions.
- Faithfulness: Many correct answers under Wrong-Think are nevertheless unfaithful, revealing a disconnect between accuracy and chain-of-thought reliability.The most perturbation-robust models can exhibit the highest inconsistency, including under Base conditions.
- Abstention: Abstention does not resolve failures: performance decreases when abstention is enabled, while adversarial text produces frequent abstentions and active misdirection.Corrective cues can increase abstention frequency, indicating continued sensitivity to disambiguating signals.
- Open versus closed models: Closed models show less performance degradation and higher faithfulness than open-source models under misleading textual interventions.Their reasoning more often acknowledges misleading context or self-corrects, although a persistent performance gap remains across models.
- Residual failure modes: Closed models remain vulnerable to hallucinations and modality conflicts despite their greater robustness and faithfulness.Examples include fabricated image URLs and reinterpretation of visual objects to align with misleading captions.
- Confidence and calibration: Adversarial prompts can confidently mislead models, sharply reducing P(Correct Letter), especially for VLAA-Thinker and Vision-R1 under Wrong-Think.Correcting cues restore substantial probability mass toward the ground-truth letter, while spatially guided models are comparatively more stable.
- Confidence and calibration: Pbase is a more precise predictor of future robustness than negative entropy across nearly all models and perturbations.In SpaceR, Pbase reaches AUROC 0.958 for Stop-Think robustness versus 0.732 for entropy.
3 Effect of RL-Finetuning on Robustness and CoT Consistency
The study examines how RL finetuning changes robustness and reasoning faithfulness in VLMs under textual perturbations. Accuracy generally improves, but reasoning consistency can decline, while augmentation improves caption robustness without reliably preventing faithfulness drift.
- Setup: RL finetuning is evaluated across Base, visual-math, and caption/thinking-augmentation settings using verifiable rewards for format and final-answer correctness.The training mixtures combine SAT2, Pixmo-Count, and optionally Geometry3K; augmentations are applied with 10% probability each for wrong or correct captions and thinking strings.
- Accuracy: Adding Geometry3K consistently improves baseline accuracy across most benchmarks, especially on mathematical or geometric reasoning tasks.The effect is most visible in Base and Wrong-Think conditions, suggesting benefits from explicit visual-math reasoning exposure.
- Robustness: Augmentation maintains comparable clean-prompt accuracy and improves robustness to misleading captions, bringing performance closer to Base-level accuracy.The improvement is stronger for Wrong-Caption than for Wrong-Think perturbations.
- Robustness: Accuracy rises across perturbation types during RL training, but models can increasingly decouple correct answers from interpretable reasoning traces.Wrong-Think remains particularly difficult despite exposure to both caption and thinking augmentations.
- Confidence: Entropy steadily decreases during training, while misleading or truncated reasoning produces higher entropy than Base and Wrong-Caption conditions.The authors interpret this as a global confidence-narrowing effect of accuracy-driven RL rather than prompt-specific adaptation.
- Caveat: Results vary substantially across random seeds, sometimes more than across dataset compositions.The study therefore emphasizes reporting multiple seeds in RL-finetuning experiments.
- Faithfulness: Faithfulness generally decreases during RL finetuning, including augmentation runs, even when accuracy improves under Wrong-Caption perturbations.The result indicates that exposure to perturbed contexts does not implicitly preserve consistent reasoning.
4 Related Work
Related work positions this study within RL-based multimodal reasoning, explicit visual grounding, and chain-of-thought faithfulness research. Prior findings motivate testing whether verbal reasoning remains aligned with visual evidence under contextual conflict.
- RL-based finetuning for VLMs: RLVR has been adapted from text-only reasoning to multimodal models to elicit stepwise visual-language reasoning.The related models include SpaceR, Video-R1, Vision-R1, VLAA-Thinker, and ViGoRL-Spatial, all finetuned from Qwen-2.5-VL-7B-Instruct.
- Perception-reasoning decomposition: Recent perception-first approaches strengthen visual grounding with curated data, intermediate representations, or verifiable proxy tasks before applying RL or CoT finetuning.These methods use representations such as scene graphs, boxes, and temporal segments.
- Spatial reasoning in MLLMs: Spatial-reasoning systems inject explicit grounding through regions, coordinates, grounding tokens, 3D scene graphs, or perception steps.Examples include Ferret, Shikra, Kosmos-2, SpatialRGPT, and Pink.
- CoT faithfulness: Research on language-model CoT questions whether plausible explanations faithfully reflect internal reasoning and whether they omit or misrepresent influential factors.This literature frames generated rationales as potentially unreliable explanations rather than guaranteed accounts of computation.
- Multimodal reasoning reliability: Prior multimodal studies report diminishing visual attention on longer traces, hallucinations, and weak correspondence among visual evidence, attention, and verbalized reasoning.These findings motivate evaluating explanation self-consistency and visual grounding together.
5 Conclusion and Future Work
The conclusion argues that controlled textual perturbations reveal weaknesses in visual grounding, uncertainty calibration, and modality-conflict resolution. It recommends evaluation and training protocols that jointly target robustness, faithfulness, and reward design.
- Conclusion: Controlled perturbations show that open-source RL-finetuned VLMs remain sensitive to language priors and struggle to arbitrate between textual and visual evidence.Stronger closed models exhibit related failure modes but greater robustness, indicating an underdeveloped open-model capability for resolving modality conflict.
- Training implications: Robustness and faithfulness do not reliably follow from simple augmentation or auxiliary rewards, which can induce shortcuts based on superficially helpful captions or reasoning prefixes.The proposed training target is genuine reconciliation of conflicting evidence rather than reliance on contextual cues.
- Faithfulness: Under RL, CoT can become an output degree of freedom optimized for reward while remaining weakly coupled to the model’s actual decision process.The authors therefore argue that CoT supervision or rationale rewards alone are unlikely to produce faithfulness.
- Future work: The authors recommend modality-conflict perturbations, conflict-specific calibration, source-aware reasoning, stronger-model or human distillation, and visual-evidence verification.They also propose extending perturbations to multi-turn and interactive settings.
- Overall implication: Trustworthy multimodal reasoning requires protocols that explicitly assess interactions among modalities, reasoning traces, calibration, and reward design.The perturbations are presented as practical probes rather than merely toy stress tests.
A Evaluation Details
The evaluation spans spatial, visual-question-answering, and general real-world benchmarks, with controlled wrong-caption and wrong-thinking augmentations. These perturbations pair incorrect auxiliary text with questions whose visual answers remain fixed.
- Spatial benchmarks: 3DSRBench evaluates height, location, orientation, and multi-object reasoning across real and synthetic images, including common and uncommon camera viewpoints.Its questions use 2,100 MS-COCO images and 672 synthetic multi-view pairs.
- Spatial benchmarks: Spatial-MM combines object-relation questions with roughly 310 open-ended multi-hop questions carrying detailed reasoning-path annotations.The Spatial-Obj subset contains around 2,000 multiple-choice questions spanning 36 relation types.
- Augmentation: The augmentation procedure supplies captions and initial thinking strings that consistently imply incorrect answers.Wrong captions follow a declarative template, while Wrong-Think initializes generation with a misleading reasoning trajectory and may append a corrective marker.
- Augmentation: For structured multiple-choice datasets, an incorrect option is selected and used to generate matching wrong captions and thinking strings.SpatialMM Multihop instead uses an alternative free-form answer, with manually verified generated samples because the dataset is relatively small.
A.5 Faithfulness Evaluation
The evaluation tests whether a model’s final answer is consistent with its preceding reasoning, using a structured judge prompt and Qwen3-32B judgments. Examples show that correctness and reasoning-answer consistency are distinct properties.
- Evaluation protocol: The judge evaluates consistency between the final answer in <answer> tags and the reasoning in <think> tags, independently of answer correctness.The protocol permits answer changes during reasoning but checks the model’s final judgment before </think>; missing answers receive “no,” while absent reasoning is vacuously “yes.”
- Evaluation protocol: Multiple-choice consistency is checked by matching either the answer text or its corresponding option letter.The prompt explicitly maps outputs such as an option letter to the associated answer choice.
- Judgment reliability: Qwen3 in Thinking mode provided reliable judgments across examples involving both incorrect-but-consistent and correct-but-inconsistent responses.The evaluation illustrates these two cases to separate final-answer correctness from reasoning consistency.
- Illustrative examples: In one example, the reasoning concludes that there is one curtain while the final answer selects the option for two, so the judge outputs “no.”The inconsistency arises because the reasoning supports option A, whereas the final answer is C.
C.1 Additional Results – Correct Thinking and Caption Ablations
Correct auxiliary captions and thinking prefixes improve accuracy, whereas dummy text has little effect and disclaimers reduce the benefit. These results implicate semantic reliance on added text and show that answer correctness does not eliminate reasoning-answer decoupling.
- Accuracy under correct auxiliary text: Dummy captions leave performance largely unchanged, while correct captions and correct thinking prefixes substantially improve accuracy across models and datasets.The contrast indicates that semantic content, rather than merely adding text or changing prompt format, drives the ablation effects.
- Accuracy under correct auxiliary text: Correct auxiliary text helps models, whereas misleading auxiliary text harms them, showing that perturbation effects depend on textual meaning.The dummy-caption control rules out a simple prompt-format explanation.
- Disclaimer effects: Disclaimer language partially reverses the accuracy gains from correct captions and thinking prefixes.The effect appears in both Correct-Caption With Disclaimer and Correct-Think With ‘But’ conditions.
- Faithfulness under correct auxiliary text: Correct thinking prefixes can support the correct answer, yet incorrect final answers often remain inconsistent with the accompanying reasoning.Thus, making the injected reasoning correct changes the accuracy direction without eliminating reasoning-answer decoupling.
- Faithfulness under correct auxiliary text: The disclaimer and faithfulness patterns indicate that models do not reliably distinguish visual evidence from auxiliary textual context.Models remain sensitive to whether added captions or thinking prefixes are presented as trustworthy or qualified.
C.2 Additional Results - Real World Understanding / General VQA Benchmarks
Additional real-world and general VQA benchmarks reproduce the vulnerability to misleading text, while the effects of suppressed reasoning vary by model and task. Faithfulness can remain low even when final-answer accuracy appears robust.
- Additional benchmark accuracy: Wrong-Think and Wrong-Caption substantially reduce accuracy across MMBench, MME-RealWorld-Lite, and V*Bench, with disclaimers only partially recovering performance.These results extend the perturbation pattern beyond the original five benchmarks.
- Additional benchmark accuracy: Wrong-Think can nearly collapse several models on MME-RealWorld and V*Bench, producing greater variation across models than Wrong-Caption.Wrong-Caption also causes significant drops, but its performance curves are more tightly clustered.
- Suppressed reasoning: Stop-Think effects are highly model- and task-dependent, improving some models while degrading others.On MMBench, fluctuations are generally small; other benchmarks show larger swings tied to model and training recipe.
- Faithfulness: Models can retain robust final-answer accuracy while producing reasoning traces frequently judged unfaithful under Wrong-Think.General-purpose VQA shows more faithful answers than tightly controlled spatial tasks, possibly because dataset priors or superficial cues suffice.
- Evaluation reliability: Agreement among judge models remains consistently high, with Qwen3-32B and GPT-OSS-120B particularly well aligned.A small human study also found high agreement between humans and judge models.
C.3 Additional Results – InternVL3 Results
InternVL3-based models show the same sensitivity to misleading textual context as the main model family. Wrong-Think additionally increases inconsistency between final answers and stated reasoning.
- Evaluation scope: The InternVL3 evaluation covers 3DSRBench, CV-Bench, Spatial-MM, WhatsUp, and SpatialMM-Multihop.This protocol applies the same evaluation design across the original five benchmarks.
- Accuracy under misleading text: InternVL3-8B and SenseNova-SI both lose accuracy under misleading captions and thinking prefixes.The image and original question remain unchanged, indicating that the effect is not specific to the Qwen2.5-VL family.
- Faithfulness under misleading reasoning: Wrong-Think increases reasoning-answer inconsistency in InternVL3-based models alongside reducing accuracy.The same qualitative faithfulness pattern appears under Base, Wrong-Think, and Wrong-Think With ‘But’.
C.4 Additional Results – RefCOCOg Grounding Perturbation Evaluation
The RefCOCOg evaluation tests whether misleading spatial text can disrupt visual grounding while leaving referring expressions unchanged. Across models, Wrong-Caption and Wrong-Think cause measurable, sometimes systematic localization shifts toward the injected misinformation.
- Evaluation setup: The proof-of-concept evaluates 9,602 referring expressions across 2,600 images under default, Wrong-Caption, and reasoning-only Wrong-Think conditions.Wrong-Caption prepends misleading directional text, while Wrong-Think injects a misleading spatial belief into the reasoning prefix.
- Evaluation setup: Perturbations reverse the target’s approximate spatial direction, preserving the original referring expression while making the injected cue inconsistent with the gold box.Robustness is measured using mean-IoU changes, relative IoU drops, and directional follow rate.
- Results: Video-R1-7B follows the misleading caption direction in 0.70 of cases, indicating systematic steering rather than random localization noise.Qualitative inspection suggests that the injected caption can function as a spatial override, sometimes displacing the referring expression.
- Results: Wrong-Think exposes a different vulnerability: VLAA-Thinker drops 22% under Wrong-Think despite only a 6% Wrong-Caption drop, while SpaceR drops 39% under Wrong-Caption.These patterns suggest differing coupling between grounding behavior, textual context, and generated reasoning traces.
- Scope: The results extend the central vulnerability beyond VQA, but the RefCOCOg experiment remains a proof of concept rather than a full grounding-faithfulness benchmark.IoU and directional-following metrics capture localization degradation and shifts, whereas determining whether decisions were driven by visual evidence would require manual annotation.
C.5 Additional Results - Wrong Caption Faithfulness Evaluation
Under Wrong-Caption, models often remain accurate by ignoring the caption, yet other generations anchor on it; therefore, accuracy can coexist with inconsistent visual-textual reasoning.
- Wrong-Caption faithfulness: Correct but unfaithful generations comprise roughly 10–15% of outputs under Wrong-Caption, with and without a disclaimer.Some incorrect answers are also unfaithful because they explicitly condition on the misleading caption.
- Wrong-Caption faithfulness: Most incorrect generations are consistent but wrong, indicating that systematic failures usually arise from adopting the misleading caption at face value.Conversely, some models answer correctly by disregarding the caption and re-describing the image.
C.6 AdditionalResults–HumanStudyComparisonwithLLM-as-JudgeforFaithfulnessEvaluation
A human study evaluates whether reasoning-answer consistency can be annotated reliably and whether Qwen3-32B provides a usable automated judge. Humans agree strongly, and the judge closely matches human majority labels while showing a conservative bias.
- Study design: The study labels 600 examples from ViGoRL and SpaceR across CVBench and 3DSRBench under Base, Wrong-Think, and Wrong-Think-With-But prompts.Each example received independent labels from three of eight annotators.
- Human agreement: Human agreement is high: all agreement metrics exceed 0.85, and annotators unanimously agree on 554 of 600 examples, or 92.3%.Agreement remains high across datasets, models, and prompt conditions, including adversarial settings.
- Judge comparison: Qwen3-32B agrees with the human majority on 91.8% of examples, with Cohen’s κ of 0.813.Humans label 70.8% of examples consistent, compared with 65.3% for the judge.
- Judge comparison: The judge is slightly more conservative than humans: most disagreements label human-consistent outputs as inconsistent, while only 8 examples show the reverse.This pattern suggests reported consistency rates may slightly underestimate true reasoning-answer faithfulness.
- Implication: Together, the results support using both human annotation and Qwen3-32B for large-scale reasoning-answer faithfulness evaluation.Strong inter-annotator agreement indicates that the consistency criterion is well-defined and reliably assessable.
D.1 Wrong-Caption - “Ignoring" the Caption
Qualitative traces show that Wrong-Caption and Wrong-Think can produce brittle reasoning strategies: models may ignore misleading text, anchor on it, or self-correct inconsistently. Correct answers therefore do not always indicate stable visual grounding.
- Caption handling: VLAA-Thinker can ignore a misleading caption and answer correctly, but other generations accept the caption and describe the scene according to incorrect text.The resulting oscillation is between caption neglect and caption anchoring rather than consistent visual grounding.
- Caption handling: ViGoRL hallucinates spatial relations under Wrong-Caption despite grounded reinforcement learning intended to bias reasoning toward visual evidence.Its example reasons that shelves are closer to the camera in accordance with the misleading caption.
- Reasoning consistency: SpaceR continues a misleading Wrong-Think trajectory about the apples’ location but outputs the correct label, demonstrating answer-reasoning inconsistency.The trace claims “top” while the final answer is “bottom right.”
- Self-correction: Vision-R1 sometimes self-corrects after detecting a contradiction, reversing an initially misleading spatial conclusion and answering that the objects are next to each other.Other corrections rely on general world knowledge rather than explicit visual grounding.
- Closed-model comparison: Gemini-2.5-Pro more often recognizes conflicts between visual evidence and text, but its final choice is not consistent across inputs.The paper illustrates this behavior with CVBench examples.
E.2 Wrong-Caption: Hallucinations
Wrong-Think conditioning can push models toward hallucinated, text-driven answers, while Stop-Think offers only modest and nonmonotonic gains. Data augmentation can instead encourage a resetting strategy that re-analyzes the image before answering.
- Hallucinations: Wrong-Think conditioning sometimes drives hallucinated outputs, including nonexistent image-crop links, rather than reliable visual reasoning.Gemini-2.5-Pro occasionally corrected the misleading reasoning, but in other cases accepted it and generated hallucinated links.
- Visual re-evaluation: After re-evaluating the image, a model can replace an initially incorrect spatial assessment with a scene-based explanation identifying the printer as closest to the stationery.The model contrasts the printer’s adjacent workstation position with the bin’s floor position.
- Wrong-Think propagation: The Wrong-Think example follows misleading spatial reasoning through to the final answer, illustrating how textual cues can determine the response.The model repeats the supplied claim that the chair is closer and outputs option (B) chair.
- Stop-Think: Stop-Think yields only modest gains over the base model, with nonmonotonic performance trends across training steps.Runs containing math data generally perform better even in the Stop-Think setting.
- Data augmentation: Data-augmented RL can produce a “resetting” strategy that begins neutrally and re-describes the scene, improving robustness under Wrong-Caption evaluation.The example first states that the spatial arrangement must be considered before reconstructing the relevant object positions.
F.3 Adding Faithfulness-as-a-Reward
Faithfulness-aware rewards can align reasoning consistency with answers, but combining them with adversarial augmentation creates unstable dynamics. The resulting models may copy misleading text or collapse into short, unfaithful outputs, limiting robustness gains.
- Reward effects: Faithfulness-aware rewards align answer correctness with reasoning consistency, countering the decoupling observed in standard RL post-training.Without the reward, accuracy remains above faithfulness; with it, the two metrics remain more closely aligned under perturbations.
- Training stability: Combining faithfulness rewards with data augmentation can destabilize training and produce generation collapse instead of jointly improving robustness and consistency.The authors attribute this interaction to changes in the reward landscape that make certain text-consistent trajectories disproportionately rewarding.
- Wrong-Caption behavior: The Aug+Faithfulness model copies misleading captions, whereas the Aug-only model uses a resetting trace to reconsider the image and answer correctly.The contrast shows that faithfulness enforcement can favor copying when the misleading caption is treated as a reward-maximizing reasoning path.
- Wrong-Think behavior: Under Wrong-Think, the Aug+Faithfulness model copies the misleading reasoning and answers “right,” while the Aug-only model re-describes positions and answers “no.”The Aug+Faithfulness output skips a substantive chain of thought, unlike the corrective trace produced by augmentation alone.
- Remaining limitation: Faithfulness rewards show promise for alignment but remain insufficient for robustness because models do not reliably distinguish valid from invalid textual reasoning cues.The paper contrasts recaptioning for Wrong-Caption with the deeper reasoning-validity discrimination required by Wrong-Think.