Reducing Hallucination in Vision-Language Models via Stage-wise Preference Optimization under Distribution Shift
arXiv:2605.16411v1 [cs.CV] 13 May 2026
Qinwu Xu1 1
Meta AI April–May 2026
Abstract Hallucination remains a fundamental challenge in vision-language models (VLMs), where autoregressive generation may produce linguistically plausible yet physically inconsistent or visually ungrounded responses due to likelihood maximization under joint probabilistic modeling. We propose a stage-wise preference optimization framework for hallucination reduction through targeted multimodal data construction. Rather than directly optimizing on generic instruction-following data, our approach progressively constructs hallucination-focused preference pairs near known failure boundaries. The framework emphasizes ambiguous spatial orientation, object relationships, OCR uncertainty, and adversarial false-premise training. Hallucinated negatives are generated through minimally perturbed yet visually inconsistent alternatives, enabling Direct Preference Optimization (DPO) to better separate grounded reasoning from plausible hallucination. Experiments on open-source benchmarks and real-world multimodal evaluation scenarios demonstrate improved grounding consistency, reduced hallucination, and more informative grounded responses. Cross-model qualitative evaluation further shows that the proposed multimodal LLM DPO framework produces more visually grounded responses than several frontier proprietary VLMs, such as in ambiguous spatial reasoning and adversarial false-premise settings. The results suggest that hallucination may arise not only from limited model capacity, but also from inherent tendencies of autoregressive probabilistic generation to favor linguistically plausible continuations under weak visual grounding. Future work may explore physical consistency modeling, uncertainty-aware multimodal reasoning, and architectural alternatives beyond standard autoregressive decoding.
1
Introduction (Theoretical Perspective on Hallucination in VLMs)
Vision-language models (VLMs) aim to approximate the conditional distribution pθ (y | xv , xt ), where xv denotes visual input, xt textual input, and y = (y1 , . . . , yT ) the output token sequence. Training is typically performed using maximum likelihood estimation via the cross-entropy objective: " T # X LCE = −E(xv ,xt ,y)∼D log pθ (yt | xv , xt , y<t ) . t=1
Using the chain rule of probability, the autoregressive formulation factorizes as T Y pθ (y | xv , xt ) = pθ (yt | xv , xt , y<t ), t=1
Figure 1: A representative multimodal reasoning failure under iterative questioning. The model initially produces a plausible response, but subsequent reasoning steps progressively introduce unsupported spatial assumptions, nonexistent objects, and incorrect physical relations despite limited visual evidence. This example illustrates how autoregressive generation may drift from grounded perception toward linguistically plausible but visually inconsistent reasoning (photo generated by AI). with, T X log pθ (yt | xv , xt , y<t ). log pθ (y | xv , xt ) = t=1
Therefore, minimizing cross-entropy is equivalent to maximizing the likelihood of the target sequence. Although highly effective for multimodal generation, this objective does not explicitly enforce grounding in visual evidence. Instead, it encourages outputs that are statistically consistent with the joint image–text distribution (Bender2021, Rohrbach 2018). Consequently, VLMs may assign high probability to responses that are linguistically plausible but visually unsupported: pθ (y | xv , xt ) ≫ 0 while y ∈ / Y(xv ), where Y(xv ) denotes the set of visually faithful responses. This issue becomes more severe under weak, ambiguous, or partially observed visual evidence, where generation increasingly relies on learned language priors rather than grounded perception. Due to autoregressive conditioning, early deviations from grounded content can propagate through subsequent token predictions, producing globally coherent yet visually inconsistent responses. Figure 1 highlights several important failure modes of current autoregressive multimodal LLM models. First, the model exhibits weak physical and spatial reasoning, generating conclusions that conflict with plausible human pose and body orientation. Second, the responses become increasingly inconsistent across reasoning turns despite operating on the same visual input. Third, the model progressively favors linguistically plausible continuations over visually verifiable evidence, introducing unsupported objects and spatial relations while maintaining fluent natural language generation. These behaviors suggest that hallucination in multimodal systems is not merely a factual error, but also reflects deeper limitations in grounding consistency, uncertainty modeling, and autoregressive probabilistic generation. Decoding strategies further influence hallucination behavior. Methods such as top-k and nucleus (top-p) sampling introduce stochasticity into generation (Holtzman 2020), while temperature scaling reshapes the output distribution. Higher diversity settings can improve fluency and descriptiveness, but also increase the likelihood of selecting tokens that deviate from grounded visual evidence. Notably, stronger reasoning capability does not necessarily eliminate hallucination. Techniques such as chain-of-thought prompting improve logical coherence and internal consistency (Wei 2022), but 2
may still extrapolate beyond observable evidence. As a result, hallucinations can become more convincing and harder to detect because the generated reasoning remains linguistically coherent despite lacking factual grounding (Dziri 2023, Lui 2023). These observations suggest that standard likelihood-based objectives primarily optimize distributional alignment rather than faithful cross-modal grounding. Addressing hallucination therefore requires supervision signals that explicitly distinguish grounded responses from plausible but unsupported alternatives. In this work, we explore this direction through stage-wise preference optimization and hallucination-targeted data construction.
2
Related Work
Hallucination in large language models (LLMs) and vision-language models (VLMs) has been widely studied, with prior work proposing mitigation strategies spanning training-time alignment, external grounding, and inference-time reasoning. Despite substantial progress, hallucination remains persistent, particularly under distribution shift and in multimodal settings, where linguistic plausibility may not correspond to visual evidence (Li et al., 2023; Liu et al., 2023). 2.1
System Prompt Guidance
System prompts are commonly used to reduce hallucination by encouraging models to avoid unsupported claims and rely only on observable evidence \cite{wei2022chain,madaan2023selfrefine}. In multimodal settings, prompts may instruct the model to acknowledge uncertainty, avoid guessing, and focus on visually grounded reasoning. Prompt-based guidance can improve response behavior at inference time with minimal computational overhead; however, its effectiveness is often limited because the underlying model distribution and cross-modal alignment are not fundamentally modified. As a result, models may still generate linguistically plausible but visually unsupported responses under ambiguous or challenging conditions. 2.2
Supervised Fine-Tuning and Data-Centric Methods
Supervised fine-tuning (SFT) reduces hallucination by aligning models to curated datasets emphasizing factual correctness and grounded responses. Under the maximum likelihood objective, SFT improves alignment with observed data distributions and enhances factual accuracy when highquality annotations are available (Ouyang et al., 2022). However, SFT does not explicitly penalize incorrect or hallucinated alternatives. As a result, models often fail to distinguish between plausible but ungrounded outputs and visually supported responses, especially in multimodal settings where linguistic plausibility may not correspond to visual evidence (Zheng 2023). 2.3
Preference-Based Learning
Preference-based methods align model outputs through comparative supervision. Reinforcement Learning from Human Feedback (RLHF) introduces a learned reward model for policy optimization (Christiano et al., 2017; Ouyang et al., 2022), but requires additional training and may suffer from reward’s mis-alignment with human preference or desired behavior. Direct Preference Optimization (DPO) simplifies this pipeline by directly enforcing pairwise preferences (Rafailov et al., 2023): log pθ (y + | x) > log pθ (y − | x). where y + is the preferred response and y − is the rejected response. Group Relative Preference Optimization (GRPO) extends preference optimization from pairwise comparisons to groups of candidate responses by optimizing relative preference scores across multiple sampled outputs (Shao et al., 2024). Instead of comparing only positive and negative response pairs, GRPO models relative advantages within a response group, encouraging the model to assign higher likelihood to more preferred responses while leveraging richer comparative supervision. A simplified objective can be written as: " # X LGRPO = −E(x,{yi }) Âi log pθ (yi | x) , i
3
where Âi denotes the normalized relative preference or advantage score of candidate response yi within the sampled response group. Compared with pairwise preference optimization, GRPO incorporates richer relative ranking information across multiple candidate responses during optimization. Preference optimization methods such as DPO and GRPO can mitigate hallucination when preference data properly reflects visual grounding and factual consistency \cite{ouyang2022training,shao2024deepseekmath}. Compared with GRPO-style group ranking objectives, DPO provides a more direct contrastive learning signal by explicitly separating grounded and hallucinated responses. Nevertheless, hallucinations in vision-language models (VLMs) remain challenging because cross-modal grounding is not explicitly enforced by standard likelihood or preference optimization objectives \cite{liu2023hallucination}. 2.4
External Grounding via Retrieval
Retrieval-augmented generation (RAG) improves factuality by conditioning outputs on external knowledge sources (Lewis et al., 2020; Shuster et al., 2021). By incorporating retrieved evidence, RAG enhances grounding and enables more verifiable responses, particularly in knowledgeintensive tasks. However, it introduces system complexity and depends heavily on retrieval quality. Errors in retrieval can propagate into the generation process, and retrieval-augmented generation (RAG) does not address hallucinations arising from reasoning or perceptual failures, particularly in multimodal settings (Lewis 2020). 2.5
Reasoning and Inference-Time Methods
Reasoning-based approaches such as chain-of-thought (CoT) prompting improve logical coherence by encouraging intermediate reasoning steps \cite{wei2022chain}. Self-consistency further enhances robustness by sampling multiple reasoning paths and selecting consistent outputs \cite{wang2022self}. Multi-turn reasoning and self-refinement frameworks enable iterative correction and clarification \cite{madaan2023selfrefine}, while multi-agent methods introduce crossverification through debate or voting mechanisms \cite{du2024improving}. Although these approaches can reduce hallucination, they primarily operate at inference time, incur substantial computational cost, and do not explicitly guarantee visual grounding or fundamentally alter the underlying model distribution. Table 1 summarizes the strengths and limitations of representative hallucination mitigation approaches. Table 1: Summary of hallucination mitigation methods Method System Prompt SFT RLHF DPO GRPO RAG CoT / Multi-turn Multi-agent 2.6
Core Mechanism Instruction guidance Maximum likelihood training Reward optimization Pairwise preference optimization Group-wise preference optimization External retrieval grounding Structured reasoning Cross-verification
Strengths Simple, low cost Stable, simple Flexible alignment Stable, efficient Richer supervision Strong factuality Improves coherence Robust reasoning
Limitations Limited behavioral control No explicit hallucination penalty Expensive, unstable Depends on data quality Requires diverse candidates Retrieval-dependent Not sufficient for grounding High inference cost
Positioning of Our Approach
Existing approaches primarily modify objectives or inference. In contrast, our work focuses on data-centric hallucination modeling, where preference data is explicitly constructed to capture subtle grounding failures near the model’s decision boundary. We generate minimally perturbed negative responses that introduce hallucinated attributes while remaining close to correct answers. This produces preference pairs near the decision boundary, yielding more informative supervision. Unlike inference-time approaches, our method operates at training time and reshapes pθ (y | x). We also adopt a stage-wise training strategy under controlled distribution shift, progressively refining grounding and reasoning. Overall, our contribution lies in structured data generation and training for hallucination reduction, complementing existing advances in preference optimization. 4
Figure 2: Overview of the model and training pipeline. A ViT encoder extracts visual features, which are compressed via a Perceiver resampler and fused with text inputs for LLaMA decoding. Our method operates at the training level through stage-wise preference optimization and structured data construction.
3
Methodology
3.1
Model Architecture
We instantiate our framework on a large-scale vision-language model (VLM) composed of a visual encoder, a Perceiver-based cross-modal alignment module, and a LLaMA-3 70B language decoder. The underlying architecture is kept fixed throughout all experiments. Rather than modifying the model itself, our approach focuses on improving hallucination robustness through training methodology and data-centric refinement, specifically via stage-wise preference optimization and hallucination-targeted data construction. The overall architecture and training pipeline are illustrated in Figure 2. We propose a stage-wise preference optimization framework that progressively adapts a visionlanguage model (VLM) from basic visual grounding to more challenging multimodal reasoning under distribution shift. The core idea is to first establish robust perceptual grounding using large-scale supervised data, and then refine the model through targeted preference optimization on hallucination-prone scenarios. The overall framework is illustrated in Figure 2. 3.1.1
Problem Setup
Let x = (xv , xt ) denote a multimodal input consisting of an image xv and a textual query xt , and let y denote the generated response. We consider two related training distributions: • D1 : a large-scale multimodal dataset emphasizing basic grounding tasks, including object recognition, OCR, and standard VQA; • D2 : a substantially smaller but more challenging distribution focused on spatial reasoning, compositional understanding, and hallucination-prone scenarios. The second-stage distribution D2 is derived from D1 through a data-centric augmentation pipeline that transforms grounded examples into more challenging reasoning-oriented training samples. These include long-form responses, adversarial false-premise examples, and grounded-versushallucinated preference pairs. Compared with D1 , D2 is intentionally smaller and more targeted, reflecting its role in hallucination-focused refinement rather than broad capability acquisition. Each sample in D2 contains a preference pair (y + , y − ), where y + ≻ y − . Our objective is to learn a policy pθ (y | x) that improves hallucination robustness while preserving strong multimodal grounding performance. 5
3.1.2
Preferred-Conditioned Data Augmentation
A key component of our framework is the construction of D2 . Rather than sampling it independently, we generate it through structured augmentation of grounded multimodal examples from D1 . In addition to long-form response enrichment and adversarial false-premise generation, we also construct hard preference pairs in which negative responses remain linguistically plausible and semantically close to the preferred responses. Many samples involve ambiguous spatial relationships, partial visual evidence, or subtle grounding errors that are difficult to distinguish. This design encourages the preference optimization stage to focus on fine-grained hallucination boundaries rather than trivial negative cases. Given x ∼ D1 with grounded response y + , we generate a perturbed sample: x′ ∼ A(x; y + ),where A denotes a structured augmentation operator that increases reasoning difficulty while preserving semantic consistency. The augmentation process includes, but is not limited to: i) introducing spatial reasoning challenges (e.g., relative positioning and occlusion-aware counting), ii) adding compositional and contextual constraints, and iii) constructing ambiguity-aware and hallucination-prone query conditions. We then construct preference pairs (y ′+ , y ′− ) for x′ , where negative responses are minimally perturbed variants of grounded answers that introduce subtle hallucinations while remaining linguistically plausible. This induces a controlled distribution shift of D1 → D2 , with D2 concentrated near decision boundaries between grounded and hallucinated outputs. Figure 3 presents examples of VQA data constructed through our multi-stage curation pipeline, where progressively more challenging hallucination scenarios are introduced, including spatial reasoning, contextual understanding, and fine-grained object attribute recognition. Although D1 consists of large-scale and diverse SFT data with varying task complexity and response lengths, selected grounded samples are further transformed through data-centric augmentation into the more targeted preference-oriented distribution D2 .
Figure 3: Stage-wise data curation. In the first stage (SFT), training data emphasizes concise and visually grounded responses. In the second stage (DPO), detailed preferred responses are paired with hallucinated alternatives to model challenging scenarios, including spatial reasoning, contextual inference (e.g., price estimation), and fine-grained object attributes such as material recognition (2nd photo was AI generated, all others were captured by the author) 3.1.3
Theoretical Interpretation
More generally, the optimization objective can be conceptually interpreted as : max ED2 [pref θ ] θ
s.t. 6
KL(pθ ∥ pθ1 ) ≤ ϵ,
where pref θ denotes the preference score assigned to grounded responses over hallucinated alternatives, and KL(pθ ∥ pθ1 ) constrains the updated policy pθ to remain close to the Stage-1 SFT policy pθ1 . Intuitively, the objective encourages improved hallucination discrimination while avoiding large deviations from the original grounded behavior learned during supervised fine-tuning. Thus, Stage 1 acts as an implicit regularizer, constraining updates to remain within a parameter region that preserves fundamental grounding behavior. From a complementary perspective, multimodal reasoning can be viewed as involving progressively more difficult latent factors z = (zeasy , zhard ), where zeasy corresponds to basic perception and grounding (e.g., object recognition and OCR), and zhard represents higher-order reasoning factors such as spatial understanding, ambiguity resolution, and hallucination-sensitive reasoning. Under this view, the conditional generation process can be conceptually expressed as p(y | x, zeasy , zhard ). Our stage-wise optimization approximates a progressive alignment process: pθ1 (y | x, zeasy ) → pθ2 (y | x, zeasy , zhard ), where Stage 1 first establishes stable multimodal grounding on simpler distributions, while Stage 2 further adapts the model to harder hallucination-prone reasoning scenarios. This staged decomposition simplifies optimization and improves robustness under distribution shift by avoiding direct optimization over highly complex reasoning behaviors from the beginning of training. 3.2
Auxiliary Mechanisms for Hallucination Reduction
In addition to stage-wise preference optimization, we incorporate two complementary mechanisms that further improve grounding: visual prompting and adversarial false-premise training. These mechanisms operate at different levels—conditioning and data construction—but share the goal of aligning generation with observable visual evidence. 3.2.1
Visual Prompting with Few-Shot Guidance
We employ a visual prompting strategy as an initial guard layer that encourages structured examination of the image prior to answer generation. The system prompt explicitly instructs the model to attend to visual evidence and avoid unsupported guessing, while few-shot examples demonstrate grounded reasoning patterns for challenging multimodal tasks such as counting, spatial understanding, and object relationships. Formally, the prompting strategy can be viewed as introducing an intermediate latent reasoning trajectory r = (r1 , . . . , rK ), representing a sequence of implicit grounding and reasoning steps before answer generation: X pθ (y | x) = pθ (y | x, r) pθ (r | x). r
Here, pθ (r | x) models the probability of selecting a particular reasoning trajectory r conditioned on the multimodal input x, while pθ (y | x, r) denotes answer generation conditioned on that reasoning process. For example, in an object-counting task, a reasoning trajectory may involve sequential grounding steps such as identifying visible objects in different image regions, checking partially occluded areas, and progressively aggregating counts before producing the final answer. The prompting strategy reshapes the distribution pθ (r | x) toward more visually grounded reasoning trajectories, thereby reducing hallucination-prone generation paths. From a probabilistic perspective, the prompting mechanism reshapes the distribution pθ (r | x), suppressing reasoning trajectories that omit important visual evidence and thereby reducing hallucination. However, as an inference-time guidance method, its effectiveness is inherently limited by the coverage and diversity of the few-shot examples, making it difficult to generalize across broad multimodal reasoning scenarios. 3.2.2
Adversarial False-Premise Training
We introduce an adversarial training scheme targeting hallucinations induced by incorrect or misleading queries. Specifically, we construct inputs x′ where the question contains a false premise 7
(e.g., referencing an object absent from the image). For each such input, we define: 1) a preferred response y + that correctly rejects or questions the premise, and 2) a negative response y − that complies with the premise and hallucinates content. These augmented examples are incorporated into the preference dataset D2 for hallucinationtargeted DPO training. This directly penalizes the model’s tendency to generate plausible but visually unsupported completions, a known failure mode of autoregressive models. By training on such adversarial cases, the model learns to better distinguish factual validity from linguistic plausibility, improving robustness under ambiguous or misleading conditions. Figure 3, case (4), presents an example of adversarial false-premise training data, where the question assumes the presence of an object (a car) that does not actually exist in the image, despite the scene context (a garage environment with dark-colored structures near the entrance) making such an assumption appear visually plausible. 3.2.3
Integration with Stage-wise Optimization
These mechanisms integrate naturally with the stage-wise framework. • Visual prompting operates at inference time, improving conditioning by structuring intermediate reasoning. • Adversarial training operates at training time, enriching D2 with targeted hallucination cases. Together with preferred-conditioned augmentation, they expose the model to diverse failure modes, including: • weak or distant visual evidence, • compositional reasoning challenges, • and false or misleading premises. This combination improves alignment between generated outputs and observable inputs, leading to consistent reductions in hallucination without requiring additional inference-time computation beyond prompting. 3.3
Controlling Response Detail under Grounding Constraints
To improve response completeness while maintaining factual grounding, we introduce data-centric strategies that explicitly reshape the distribution of response detail during training. We employ two complementary augmentation strategies to increase response richness: • Long-form duplication: We apply a length-aware resampling strategy based on exponential tilting. Specifically, each sample with length x is assigned a weight: w(x) = eβ(x−µa ) where µa is the mean length of the original distribution and β > 0 controls the degree of up-weighting for longer responses. This formulation increases the relative contribution of longer answers while preserving the overall distributional structure. The parameter β is chosen to match the empirical shift in mean length between the source and target distributions. In practice, we additionally cap w(x) to avoid over-amplifying extreme long-tail samples. • Iterative expansion: We generate second-turn responses that progressively refine and elaborate the initial answers used in SFT training, adding attributes, spatial relations, and contextual details. These strategies shift the distribution of preferred responses toward greater length and specificity and complexity such including spatial relationship and reasoning. As shown in Figure 4 , the augmented dataset exhibits a consistent increase in response length, along with improved coverage of finegrained visual attributes. 8
Figure 4: DPO response length of words distribution: a) original one; b) new one with data duplication and augmentation Qualitatively, the resulting responses contain richer object descriptions, more precise spatial relations, and more explicit grounding in visual evidence. 3.3.1
Trade-off Between Detail and Hallucination
Increasing response detail introduces an inherent risk of hallucination. The model learns a conditional distribution pθ (y | x), which combines evidence-supported generation with prior-driven completion. When visual evidence is incomplete or ambiguous, generating longer and more specific responses increases reliance on learned priors, thereby raising the likelihood of unsupported details. This phenomenon can be interpreted through conditional entropy: increasing response richness effectively enlarges the space of plausible continuations, increasing the probability of ungrounded generations. To mitigate this effect, our preference construction emphasizes detailed yet visually grounded responses while contrasting them against similarly detailed but subtly hallucinated alternatives. Rather than rewarding verbosity itself, the optimization encourages response richness only when supported by observable evidence. This preference construction and optimization strategy encourage the model to generate richer and more informative responses while suppressing unsupported elaborations. Rather than enforcing a simple trade-off between detail and hallucination, the optimization reallocates probability mass toward responses that are both detailed and visually grounded. Although residual ambiguities may persist in highly challenging scenarios, the overall effect is improved informativeness together with reduced hallucination. We empirically analyze this trade-off and its mitigation in Section 4.3.2. 3.4
Practical Implications
The proposed framework yields several practical benefits: 1) Improved optimization stability: initialization from θ1 stabilizes preference optimization on more challenging hallucination-targeted data distributions; 2) Better distributional coverage: large-scale Stage-1 training provides broad multimodal grounding and linguistic coverage prior to targeted refinement; 3) Targeted robustness: preference-oriented augmentation concentrates learning on difficult reasoning boundaries involving ambiguity, spatial understanding, and hallucination-prone conditions; and 4) Improved detail– grounding balance: the preference construction encourages richer responses while discouraging unsupported elaborations. Together, these components form a progressive data-centric refinement pipeline that adapts multimodal models to increasingly complex reasoning scenarios while improving hallucination robustness without modifying the underlying model architecture.
4
Experimental Results and Analysis
4.1
Experimental Design
We evaluate our approach using a combination of open-source benchmarks and curated evaluation protocols, comparing model behavior before and after hallucination-targeted training. The primary evaluation tasks include visual question answering (VQA) (Antol 2015), document visual question
9
answering (DocVQA) (Mathew 2021), and multimodal reasoning benchmarks such as MMBench (Liu 2023). To capture different aspects of model quality, we employ an LLM-as-a-judge framework for pairwise evaluation. In this setting, responses from two models are compared on the same input, and a judge model selects the preferred answer. In contrast, pointwise evaluation assigns an independent scalar score to each response, with model performance measured through averaged scores across samples. Prior work has shown that pairwise evaluation can reduce common scalar-scoring biases, particularly verbosity bias, where longer responses are systematically favored regardless of factual correctness or grounding (Zheng 2023). In addition to quantitative evaluation, we conduct qualitative analysis to examine improvements in visual grounding, multimodal reasoning, hallucination robustness, and response informativeness. We further evaluate performance on long-tail and hallucination-prone scenarios, comparing our approach against strong baselines including LLaMA3 multimodal (MM) DPO (ours), Gemini Flash, and GPT-series models. It is worth noting that we also include a small subset of AI generated stylized images, such as watercolor-like renderings generated by Nano Banana, as a stylized-image transformation for visual question answering (VQA). These transformations obscure Personally Identifiable Information (PII) while preserving key spatial and semantic content, consistent with prior work onidentity-obscured VQA and vision-language modeling (Bara et al., 2022; Anonymous, 2026). Additionally, such stylization introduces mild distribution shifts, allowing us to probe whether models rely on superficial visual cues or more robust semantic understanding, as explored in prior stylized-image visual tasks (Omoniyi et al. 2025). 4.2
Quantitative Results
Performance on Open-Source Benchmarks We first evaluate the Stage-1 SFT checkpoint on several open-source benchmarks to verify that the model maintains strong general multimodal capability prior to hallucination-targeted refinement. As shown in Table 2, the model achieves competitive performance across standard visual understanding tasks, including DocVQA, VQA, and MMBench. For the Stage-2 DPO-trained model with the new stage-wise and hallucination-targeted refinement, we further conduct pairwise evaluation on approximately 839 diverse multimodal evaluation scenarios, including indoor scenes, outdoor environments, food, landscapes, and plants. In this evaluation, responses from two models are compared on the same input and judged by an LLM-based evaluator. Results reveal substantial improvements in grounding quality and response informativeness. As shown in Table 3, the DPO-trained model achieves consistent gains across multiple judge models, with win rates increasing by +6.2% to +8.2% over the baseline. Table 2: Performance comparison on standard multimodal benchmarks. Model LlaMA-3 MM base LlaMA-3 MM trained
DocVQA 82.30 83.00
VQA 81.45 82.18
MMBench 78.99 80.14
Table 3: Comparison of model performance across evaluation metrics. Metric LlaMA-3 MM base Wins LlaMA-3 MM trained Wins Equivalent (TIE/C) Both Wrong (D) Error Rate ∆ Win Rate (%)
GPT-4o 207 (24.7%) 276 (32.9%) 294 (35.0%) 49 (5.8%) 13 (1.5%) +8.2%
10
Gemini-2.5-Flash 256 (30.5%) 316 (37.7%) 221 (26.3%) 20 (2.4%) 26 (3.1%) +7.2%
LLaMA-4-Marvrick 265 (31.6%) 317 (37.8%) 220 (26.2%) 19 (2.3%) 18 (2.1%) +6.2%
4.3 4.3.1
Qualitative Analysis Context Richness and Hallucination Reduction
Figure 5 presents representative examples comparing SFT and DPO model outputs. The DPOtrained model consistently produces richer and more actionable responses while maintaining stronger visual grounding. Beyond describing visible content, the model learns to provide contextually useful information aligned with query intent. For example, in the shopping scenario (Example 2), the model not only identifies storefront elements but also reasonably infers operational context and the likely availability of purchasable items, improving contextual completeness and task usefulness . Similar patterns are observed across multiple examples, where preference optimization improves contextual completeness without introducing unsupported details. These results suggest that DPO encourages responses that better balance informativeness and grounding.
Figure 5: Representative comparisons between SFT and DPO model outputs across diverse multimodal scenarios (1st photo captured by the author, 2nd by Krüger 2024, Unsplash, 3rd and 4th ones are AI generated) Figure 6 further illustrates hallucination mitigation across spatial reasoning, OCR understanding, and contextual interpretation tasks. In several cases, the queried information is not directly observable from the image. For example, when asked to determine the number of pages in a book (Example 1) or the prices of food items (Example 3), the baseline model generates plausible but unsupported guesses. In contrast, the DPO-trained model correctly abstains, explicitly indicating that the information is not visible or suggesting alternative ways to obtain the answer. Similarly, under difficult OCR conditions where text is only partially readable, the SFT model tends to infer unsupported details from scene context, whereas the DPO model avoids such ungrounded extrapolation (Figure 6, Example 4). We also compare the proposed approach against visual prompt-based hallucination mitigation. Although prompting occasionally reduces fabricated details, its effectiveness remains limited and inconsistent due to the restricted generalization capability of few-shot prompt guidance (Example 1). In some cases, the prompting strategy makes the model overly conservative, resulting in vague or partially irrelevant responses rather than grounded descriptions (Example 2). In contrast, the DPObased approach achieves more reliable alignment through preference-based training, enabling the model to suppress hallucinations while preserving descriptive richness and task relevance. Overall, these results highlight the limitations of prompt-level control and demonstrate the advantage of data-centric preference optimization for robust hallucination reduction.
11
Figure 6: Representative examples of hallucination mitigation under spatial reasoning, OCR, and contextual understanding scenarios (1st/2nd photos were captured by authors and 3rd/4th were AI generated). 4.3.2
Cross-Model Comparison
To further analyze the trade-off between response richness and hallucination, we conduct a comparative study on multimodal examples across multiple models, including our LLaMA3-MM DPO model, Gemini Flash, and GPT-series models. The representative examples are demonstrated in the Appendix - Across Model Comparison and Analysis. Figure 7 presents a representative adversarial example involving fine-grained bug classification. The queried object is visually ambiguous and lacks sufficient evidence for reliable species-level identification. In this case, the DPO-trained model responds conservatively and avoids unsupported categorization. In contrast, both GPT-series and Gemini Flash models generate specific bug labels despite limited visual evidence. Notably, the Gemini model produces inconsistent predictions (e.g., silverfish versus seed bug) under different prompting constraints, even though these categories correspond to visually distinct insect types. This behavior suggests that the generated responses may be influenced more strongly by linguistic priors or token-level completion preferences than by stable visual grounding. These observations further highlight the importance of hallucination-targeted preference optimization for improving consistency and grounding in ambiguous multimodal scenarios. Figure 8 presents a comparative evaluation of three multimodal models across four challenging visual question answering scenarios, highlighting differences in hallucination behavior, grounding reliability, and multimodal reasoning consistency. Overall, LLaMA-3 MM-DPO produces more conservative and visually grounded responses, avoiding unsupported inferences under ambiguous conditions. In contrast, GPT-5.3 and Gemini Flash more frequently introduce additional details that are not strictly supported by the image, reflecting a stronger tendency toward language-driven completion and contextual extrapolation. Example 1 reveals limitations in physical reasoning and spatial grounding. In this case, Gemini Flash hallucinates the presence of a scooter despite the implausible physical configuration required for the child to carry such an object horizontally. GPT-5.3, while avoiding the scooter hallucination, incorrectly interprets the spatial relationship between the person’s left and right hands under partial occlusion. Example 2 illustrates ambiguity sensitivity and reasoning instability, where Gemini Flash produces inconsistent interpretations of the same scene across different prompting conditions (e.g., sunset with visible evergreen trees versus a low-light nighttime scene). In Examples 3 and 4, all models capture the overall scene semantics; however, weaker grounding leads to varying degrees of over-interpretation, such as inferring brake-light states or contextual details beyond observable evidence. These results suggest that stage-wise preference optimization improves grounding consistency and reduces hallucination under challenging multimodal conditions, although physical reasoning and uncertainty handling remain difficult failure modes for current vision-language models.
12
Figure 7: Adversarial hallucination example for fine-grained object identification. The DPO-trained model avoids unsupported classification and remains grounded in visible evidence, whereas GPTseries and Gemini Flash models generate specific but unverified bug categories (photo captured by author). Overall, the results highlight differences in how multimodal models balance visual evidence against language-driven inference. While all models perform reasonably well on simpler and well-grounded queries, larger differences emerge in scenarios involving ambiguity, long-form reasoning, spatial relationships, or incomplete visual evidence. Additional challenging VQA examples are provided in the Appendix. LLaMA-3 MM (DPO) tends to produce more conservative responses, particularly under ambiguous or adversarial conditions. Its outputs remain closely tied to observable evidence, resulting in a lower incidence of unsupported inference and improved grounding consistency. However, this conservatism can occasionally reduce descriptive richness in highly uncertain scenarios. GPT-series model (API version accessed at evaluation time) demonstrates comparatively balanced behavior between descriptiveness and grounding. The model generally produces structured and cautious interpretations while maintaining reasonable alignment with visual content. In some cases, however, it appears less consistent in leveraging fine-grained visual cues, such as small or partially occluded text, which may be omitted or only partially captured without explicit prompting. Gemini Flash model (API version accessed at evaluation time) generates the most detailed and context-rich responses among the evaluated systems, capturing a broad range of objects, scene relationships, and contextual cues. However, this stronger descriptiveness is accompanied by a higher tendency toward over-interpretation, where additional attributes or contextual details are inferred without sufficient visual evidence or beyond the scope required by the query.
13
Figure 8: Cross-model comparison across challenging multimodal reasoning scenarios involving ambiguity, spatial understanding, and contextual interpretation (1st/4th photos generated by AI, 2nd/3rd ones captured by authors). Table 4: Strengths and limitations observed across evaluated multimodal models. Model LLaMA-3 MM (DPO)
GPT-5.3 Instant
Gemini 3 Flash
Strengths Strong grounding to visible evidence; lower tendency toward unsupported inference; more consistent handling of hallucination-prone scenarios Balanced reasoning and interpretation; generally structured and cautious responses; good overall grounding quality Rich and detailed scene descriptions; strong semantic and contextual interpretation capability
Limitations Can be comparatively conservative in ambiguous cases, occasionally reducing descriptive richness Less consistent utilization of finegrained visual cues (e.g., small text or partially occluded objects); occasional incomplete extraction of visual details Higher tendency toward overinterpretation and unsupported contextual inference under ambiguous conditions
Regarding OCR-related performance, the GPT-series model shows comparatively weaker robustness under challenging text conditions, including small, blurred, or partially occluded text regions. We note that comparisons involving proprietary systems (GPT and Gemini) are based on publicly available API versions at the time of evaluation and may evolve with future model updates. Furthermore, no external OCR or computer-vision toolkits were integrated into any of the evaluated systems during testing. Importantly, our method shifts this trade-off frontier. Relative to the LLaMA-3 MM (SFT) baseline, the stage-wise DPO-trained model improves response richness while maintaining or reducing hallucination, indicating more effective allocation of probability mass toward evidence-supported details. A detailed comparison across multiple evaluation judges is summarized in Table 4. Across the manually analyzed challenging examples, Gemini Flash exhibited the highest frequency of hallucination cases, followed by GPT-series models, while LLaMA-3 MM (DPO) showed the lowest frequency of unsupported inference. Note that the qualitative examples are intended as illustrative case studies rather than exhaustive benchmarking. Future work may explore larger-scale human evaluation, statistical analysis, and standardized annotation protocols.
14
4.4 4.4.1
Discussion Data Scaling and Preference Efficiency
A notable finding is the strong data efficiency of preference-based optimization. With only approximately 50k–120k hallucination-targeted DPO samples, the model exhibits consistent improvements in grounding consistency and hallucination reduction. This contrasts with pretraining and supervised fine-tuning (SFT), which typically require substantially larger datasets to establish broad multimodal capabilities. While large-scale training remains essential for general perception and language understanding, our results suggest that carefully constructed preference data can produce disproportionately large alignment gains during later refinement stages. We attribute this efficiency primarily to data structure rather than raw scale. The proposed hallucination-targeted preference pairs are constructed near difficult reasoning and grounding boundaries, yielding high-information training signals that directly target failure modes of multimodal generation. These findings suggest that, for alignment-oriented objectives, improving data quality and preference structure may be more impactful than simply increasing training scale. Interestingly, these findings are closely related to the observed detail–grounding trade-off. The effectiveness of relatively small preference datasets suggests that hallucination behavior is concentrated near ambiguous reasoning boundaries, where carefully structured supervision can strongly influence the balance between descriptive richness and visual grounding. 4.4.2
Trade-offs Between Detail and Hallucination
Our experiments reveal a consistent relationship between response richness and hallucination risk. As model outputs become longer and more descriptive, the probability of introducing unsupported attributes or contextual inferences increases, particularly under incomplete, ambiguous, or weak visual evidence. This behavior is illustrated in Figure 8. Models favoring highly detailed responses (e.g., Gemini Flash) exhibit stronger tendencies toward over-interpretation, whereas more conservative models reduce hallucination at the cost of descriptiveness and contextual richness. This phenomenon can be interpreted as a shift from evidence-conditioned generation toward priordriven completion. Increasing response length expands the space of plausible continuations, increasing reliance on learned priors and language-level completion behavior. Our approach mitigates this effect through the preference construction strategy introduced in Section 3.4.2, where detailed grounded responses are contrasted against similarly detailed but hallucinated variants. This isolates grounding—not verbosity—as the primary optimization signal. As a result, the proposed stage-wise preference optimization reshapes the generation distribution toward evidence-supported detail rather than merely suppressing response length. Empirically, this shifts the detail–grounding trade-off frontier, enabling richer responses while maintaining or improving factual consistency. Nevertheless, residual trade-offs persist in highly ambiguous scenarios, indicating that fully disentangling descriptiveness from hallucination remains an open challenge for multimodal reasoning systems. 4.4.3
Limitations of Autoregressive Multimodal Generation
Despite substantial improvements in grounding and hallucination reduction, several fundamental limitations remain difficult to address within autoregressive likelihood-based generation frameworks. First, current multimodal models still exhibit weaknesses in physical and spatial reasoning, particularly under occlusion, ambiguity, or complex human-object interactions. In such cases, token prediction may favor statistically plausible completions that violate physical consistency. Second, hallucination is closely tied to the language modeling objective itself. Because autoregressive models optimize joint token likelihood, generation can become dominated by linguistic plau-
15
sibility and contextual priors rather than strict visual evidence, especially when the image provides incomplete or ambiguous information. Third, model outputs can vary noticeably across prompting conditions, decoding strategies, or response-length constraints. As illustrated in several adversarial examples, relatively small prompt variations may lead to substantially different interpretations of the same visual input, indicating instability in the underlying reasoning trajectories. These observations suggest that hallucination is not solely a data or alignment problem, but also reflects intrinsic limitations of autoregressive probabilistic generation. Future work may require stronger uncertainty modeling, explicit physical reasoning mechanisms, or alternative multimodal generation paradigms beyond standard next-token prediction objectives.
5
Conclusion
We present a data-centric framework for reducing hallucination in vision-language models through stage-wise preference optimization under distribution shift. Rather than modifying the underlying model architecture, our approach focuses on structured supervision and hallucination-targeted preference construction. By generating detailed grounded responses together with minimally perturbed hallucinated alternatives, the proposed framework enables the model to better distinguish between linguistic plausibility and evidence-supported generation. Our results demonstrate several important findings. First, carefully structured preference data can produce disproportionately large alignment gains despite relatively small training scale, suggesting that data quality and supervision structure may be more critical than raw volume during post-training refinement. Second, pairwise evaluation provides a more reliable signal for hallucination assessment than traditional pointwise metrics, particularly under long-form generation where verbosity bias can obscure grounding quality. Third, the relationship between descriptiveness and hallucination is not purely a fixed trade-off. Through hallucination-targeted preference optimization, the model can shift the generation distribution toward richer yet more grounded responses, improving informativeness while maintaining factual consistency. Qualitative comparisons further reveal that hallucination often emerges near ambiguous reasoning boundaries where autoregressive generation becomes increasingly influenced by linguistic priors rather than visual evidence. Although the proposed approach substantially improves grounding robustness, challenges remain in physical reasoning, uncertainty handling, and consistency under ambiguous visual conditions. These observations suggest that hallucination is not solely a limitation of model capacity, but also reflects intrinsic characteristics of autoregressive probabilistic generation. Overall, this work highlights the importance of data-centric alignment for multimodal reasoning systems and demonstrates that targeted preference optimization can effectively reshape grounding behavior without requiring architectural modification. Future work will explore stronger mechanisms for physical reasoning, uncertainty-aware multimodal generation, and ambiguity-aware grounding under challenging visual conditions.
References Alayrac, Jean-Baptiste, Jeff Donahue, Pauline Luc, et al. Flamingo: A Visual Language Model for Few-Shot Learning. Advances in Neural Information Processing Systems, 2022. Antol, Stanislaw, Aishwarya Agrawal, Jiasen Lu, et al. VQA: Visual Question Answering. Proceedings of the IEEE International Conference on Computer Vision, 2015. Bai, Jinze, Shuai Bai, Shusheng Yang, et al. Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond. arXiv preprint arXiv:2308.12966, 2023. Bara, C.-P., et al. Privacy Preserving Visual Question Answering. AAAI Workshop on PrivacyPreserving Artificial Intelligence, 2022. Christiano, Paul F., Jan Leike, Tom Brown, et al. Deep Reinforcement Learning from Human Preferences. Advances in Neural Information Processing Systems, 2017.
16
Dosovitskiy, Alexey, Lucas Beyer, Alexander Kolesnikov, et al. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale. International Conference on Learning Representations, 2021. Du, Yilun, Shuang Li, Antonio Torralba, Joshua B. Tenenbaum, and Igor Mordatch. Improving Factuality and Reasoning in Language Models through Multiagent Debate. International Conference on Machine Learning, 2024. Jaegle, Andrew, Sebastian Borgeaud, Jean-Baptiste Alayrac, et al. Perceiver IO: A General Architecture for Structured Inputs and Outputs. arXiv preprint arXiv:2107.14795, 2021. Krüger, Kevin. A Dark City Street at Night with a Store Front. Unsplash, 2024. Lewis, Patrick, Ethan Perez, Aleksandra Piktus, et al. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks. Advances in Neural Information Processing Systems, 2020. Li, Junnan, Dongxu Li, Silvio Savarese, and Steven Hoi. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models. International Conference on Machine Learning, 2023. Liu, Yuan, Haotian Li, Yuhang Wu, et al. MMBench: Is Your Multi-modal Model an All-around Player? arXiv preprint arXiv:2307.06281, 2023. Madaan, Aman, Niket Tandon, Peter Clark, et al. Self-Refine: Iterative Refinement with SelfFeedback. Advances in Neural Information Processing Systems, 2023. Mathew, Minesh, Dimosthenis Karatzas, and C. V. Jawahar. DocVQA: A Dataset for VQA on Document Images. Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, 2021. Omoniyi, A., et al. Privacy-Preserving Image Captioning with Partial Encryption and Deep Learning. Mathematics, 2025. Ouyang, Long, Jeffrey Wu, Xu Jiang, et al. Training Language Models to Follow Instructions with Human Feedback. arXiv preprint arXiv:2203.02155, 2022. Rafailov, Rafael, Archit Sharma, Eric Mitchell, Stefano Ermon, Christopher D. Manning, and Chelsea Finn. Direct Preference Optimization: Your Language Model is Secretly a Reward Model. Advances in Neural Information Processing Systems, 2023. Shao, Zhihong, Peiyi Wang, Qihao Zhu, et al. DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models. arXiv preprint arXiv:2402.03300, 2024. Shuster, Kurt, Spencer Poff, Moya Chen, Douwe Kiela, and Jason Weston. Retrieval Augmentation Reduces Hallucination in Conversation. Findings of EMNLP, 2021. Wang, Xuezhi, Jason Wei, Dale Schuurmans, et al. Self-Consistency Improves Chain of Thought Reasoning in Language Models. International Conference on Learning Representations, 2022. Wei, Jason, Xuezhi Wang, Dale Schuurmans, et al. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models. Advances in Neural Information Processing Systems, 2022. Xu, Hang, Saining Xie, Xiaokang Tan, et al. Democratizing Contrastive Language-Image Pretraining: A CLIP Model Trained on 400 Million Image-Text Pairs. arXiv preprint, 2023. Zheng, Lianmin, Wei-Lin Chiang, Ying Sheng, et al. Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. arXiv preprint arXiv:2306.05685, 2023.
17
Appendix - Mini-Benchmark of different multimodal LLM models - examples LLaMA3-MM Image Gemini 3 Flash
Q: What do you see? (photo captured by author)
Q: What kind of bug is this?(photo captured by author)
SFT: It is painted from Mennet painting a The painting visible in the top-left feathery. corner of the second image is by Claude Monet, as indicated by the text on the SFT+ Prompt: This is a box of SHIN print. Based on the visible colors and Ramyun instant noodles, a popular Korean style, it appears to be one of his famous spicy noodle brand produced by Nongshim. Water Lilies (Nymphéas) series. Q: What painting is this? (photo captured by author)
DPO: This is a painting by Claude Monet. SFT: To run on a treadmill, first warm up by walking for 2 minutes at a speed of 4–5 km/h. Then jog for 10 minutes at 6–7 km/h and finally cool down by walking again for 2 minutes. You can adjust this routine according to your pace and needs.
Q: How is this game played? (photo - AI generated)
GPT-5.3 Instant
Analysis
SFT: You’re looking at a dog standing on a You are looking at an asphalt surface A cracked asphalt road or path in a wooded Gemini and LLaMA-3 MM paved area, with trees and vegetation behind with visible black sealant cracks in the area, with trees and dense green bushes (DPO) both correctly identify it. foreground. In the background, a small, along the edge. the key object (dog) and tan-colored dog is standing near a lush provide reasonable spatial DPO: You’re looking at an asphalt road or green bush at the edge of a wooded area grounding. GPT misses the driveway, with trees and foliage ahead. with tall, slender trees. dog entirely, indicating a There are cracks and patches in the road, and failure in salient object a dog is standing further ahead. detection. While Gemini is more detailed, both Gemini and LLaMA are grounded and accurate. SFT: This looks like an insect, possibly a fly Based on the image, that appears to be a The insect appears slender with a long, thin This is an adversarial case. or beetle. Western Conifer Seed Bug. They are body and narrow wings folded along its LLaMA-3 MM correctly common "accidental invaders" that enter length—consistent with a small moth rejects the premise (no bug DPO: This is not a bug; it’s an envelope or homes seeking warmth. They are (possibly a clothes moth or similar present). Both GPT and package that may contain some documents harmless to humans and structures, species). Gemini hallucinate specific or contents. though they can release a pungent odor insect types, demonstrating if handled or crushed. Just gently over-reliance on prior relocate it outside. knowledge instead of visual evidence.
DPO: To play this game, players should sit together, and one player should spin the clock’s hand. Wherever the hand stops, that time indicates how many minutes the player must perform a task.
This clock rug is an educational tool for learning to tell time. Children can physically jump or place objects on the numbers. They can also use their arms or long toys to mimic clock hands, practicing terms like "Quarter To" and "Half Past" printed on the rug.
A framed print referencing Claude Monet GPT gives a cautious and is visible in the background, likely an grounded answer, avoiding Impressionist-style artwork rather than a overcommitment. LLaMA-3 clearly identifiable specific painting. MM (DPO) identifies the artist correctly but lacks specificity. Gemini over-specifies by hallucinating a specific series (“Water Lilies”) without sufficient evidence. Players spin the center arrow. When it The image does not provide a stops, they move to or interact with the clearly defined rule set. All number it points to—such as stepping on it, models infer plausible counting, or doing a simple task. Turns gameplay mechanics, but none rotate among players. The game focuses on can be verified as correct. This learning numbers, counting, and basic is an inherently ambiguous turn-taking through play. case where hallucination cannot be strictly judged.
1
Image
LLaMA3-MM
Gemini 3 Flash
GPT-5.3 Instant
SFT: This book is 120 pages.
The visible page numbers are 102 and 103. Given the visible thickness of the remaining pages underneath, the book likely contains at least 200 to 250 pages in total. It appears to be a large treasury or collection of children's stories.
I can’t determine the exact number of LLaMA-3 MM (DPO) and pages from this image alone. It only shows GPT correctly refuse to infer one open spread of an illustrated children’s beyond visible evidence. book, without visible page numbers, spine Gemini hallucinates a page thickness, or cover details. number based on a visible page index, which does not indicate total page count.
This image shows a commemorative plaque or historical marker on a stone base. It is situated on a grassy hill overlooking a road and residential neighborhood. Such markers typically identify local historical sites, memorials, or geographic points of interest in public parks.
A ground utility box (likely for irrigation LLaMA-3 MM and Gemini or electrical access) set in a grassy area both correctly identify the near a road. object as a memorial/plaque. GPT initially misclassifies it as 2nd attempt: It appears to be a memorial or a utility box but corrects itself commemorative plaque mounted on a in a second attempt. Gemini stone base. It’s placed outdoors on a grassy provides richer context but hill, overlooking a road and nearby area, slightly overextends beyond likely marking a historical event, location, the question. or honoring a person or group.
SFT + Prompt: by looking at the page number, this book appears to have 68 pages.
Q: Tell me how many pages this book has by looking at it (photo captured by author).
DPO: The number of pages in the book is not visible.
SFT: This is a memorial plaque that includes a person's name, dates of birth and death, and some personal information or messages. DPO: This is a memorial or plaque, perhaps for a person or event. Q: What is this? (photo - AI generated)
Analysis
SFT: You are looking at a dark screen or You are looking at a completely black, The image is too dark to make out any clear All three models appropriately surface, possibly a view of a room or area solid image. There are no visible objects or details. express uncertainty without that is dimly lit. objects, colors, or patterns present in the hallucinating content. file. It appears to be either a blank frame Differences are stylistic rather DPO: You are looking at a dark-colored or an image taken in total darkness. than substantive. screen with no visible scene or text. Q: What am I looking at? SFT: In front of you is a large aquarium with many types of fish and other aquatic animals. DPO: In front of you is a large aquarium with various types of fish.
You are looking into a large aquarium or fish tank. A school of small, silvery fish with dark spots near their tails is swimming in front of you. The tank features a sandy bottom, large rocks or driftwood, and some green aquatic plants. (note: dark spots are at the center)
An aquarium tank with fish and a sea turtle All models correctly identify swimming underwater. an aquarium. Gemini provides the most detailed description but includes minor spatial inaccuracies. GPT hallucinates a turtle, which is not clearly supported. LLaMA-3 MM is concise and fully grounded.
Q: What is in front of me? (photo captured by author)
2
Image
LLaMA3-MM
Gemini 3 Flash
GPT-5.3 Instant
SFT: The price of this cake is ₹450, with GST extra.
Q: What is its price? (AI generated)
Analysis
You are looking at an assortment of six The image shows a box with desserts (cake cake slices in a white bakery box. The slices and a grilled sandwich), but no price box contains two pink-frosted slices, is visible, so the cost cannot be determined. DPO: Prices vary depending on location, two dark chocolate or black forest slices, size, and type of cake. For the best prices, and two triangle-shaped honey or milleyou can contact local bakeries or online cake feuille slices. While prices vary by shops. bakery, a set like this typically costs between $20.00 and $35.00.
LLaMA-3 MM and GPT correctly refuse to infer price without visual evidence. Gemini hallucinates a plausible price range, which is not grounded in the image.
SFT: I don’t know exactly how the instructor is, but they are working hard to engage their students and provide valuable information.
There’s no clear way to judge the teacher’s quality from this image alone. It just shows a classroom setting with a teacher presenting and students listening; effectiveness depends on things like teaching style, clarity, engagement, and student outcomes, which aren’t visible here.
All models appropriately avoid making unsupported judgments. Differences lie only in verbosity and explanation depth.
Sides listed: French fries (“tater tots”), hush puppies, coleslaw, beer-battered onion rings, side Caesar salad, side house salad, and Chef’s veg.can. Follow-up Q: give price details. A: French fries (“tater tots”): about $3, Hush puppies (5): about $6 Coleslaw: about $3, Beer-battered onion rings: about $6, Side Caesar salad: about $6, Side house salad: about $6, Chef’s veg: about $6
LLaMA-3 MM provides the most accurate OCR extraction including prices. Gemini captures items but partially misses or distorts price details. GPT omits prices unless explicitly prompted and later produces inaccurate estimates.
The instructor uses visual aids and maintains an orderly environment. While students appear focused, assessing teaching quality requires hearing their explanations and seeing student-teacher interaction. Currently, the setup looks professional and wellorganized for a lecture-style class.
DPO: Evaluating the quality of an instructor is difficult because it depends on many factors, such as teaching style, subject matter expertise, and interaction with Q: Is this teacher a good students. In this image, the instructor is instructor? (AI generated) shown giving a lecture, but there’s no specific information about their teaching skills or interaction with students. Therefore, I cannot conclude whether they are good instructors or not. SFT: French Fries, Tater Tots, Hush The Sides listed on the menu are: Puppies, Coleslaw, and Chef’s Veg. French Fries | Tater Tots (3.75) Hush Puppies (5) (3.95) DPO: The menu lists the following sides: Coleslaw (2.75) French Fries / Tater Tots – $1.75 Beer-Battered Onion Rings Hush Puppies (s) – $1.75 Side Caesar Salad (5.75) Coleslaw – $1.50 Side House Salad (5.75) Chef’s Veg – $1.50 Chef’s Veg (3.50) Q: Look over here and read the Side Caesar Salad – $1.50 Side House Salad – $1.50 sides (photo captured by author). Millibilly Cheddar Biscuit – $1.50 Beer-Battered Onion Rings – $2.50
3
Image
LLaMA3-MM
Gemini 3 Flash
GPT-5.3 Instant
Analysis
No, this is a train station platform at night. A skyline shows a city's profile against the sky. This image features a brick building with overhanging lights, a dark platform, and railroad tracks, DPO: No, this looks like a residential area capturing a localized, industrial mood with houses and trees. There’s no skyline or rather than a distant urban horizon. skyscrapers visible.
I don’t see a skyline. The image appears to show a dark nighttime alley or walkway beside a building, with colorful artificial lights and railroad tracks extending into the distance under a deep blue sky.
LLaMA-3 MM and GPT correctly reject the premise (no skyline). Gemini provides a detailed but partially hallucinated description (e.g., inferred sunset, specific tree types).
SFT: You’re looking inside an elevator, where a small child is standing.
You are looking at a child standing inside an elevator with blue chevronpatterned walls and a wood-look floor. The child is wearing a dark jacket with a yellow lining and is carrying a red and blue backpack. The elevator's control panel and a informational sign are also visible.
A small child standing in an elevator corner, holding a scooter. The elevator has padded blue walls, wood-pattern flooring, and a control panel with buttons and notices.
LLaMA-3 MM provides a correct and minimal description. Gemini and GPT hallucinate objects (backpack, scooter) with weak physical plausibility.
SFT: A person is looking out from inside a vehicle, possibly a car or SUV.”
You’re viewing an impressionist painting of a car’s interior from the backseat. The artist uses thick, visible DPO: You’re looking at an impressionist brushstrokes to capture light and form, painting. A person is sitting inside a vehicle, showing the front passenger seat, a possibly a car or SUV. They’re in the seatbelt, and a suburban street through passenger seat and the driver’s seat is the window, transforming a routine empty. Outside, you can see a residential drive into a textured, artistic scene. area with houses and lawns.
You’re looking at a stylized, impressionistic view from inside a car. The interior frame, seat, and a passenger are visible. Through the window, a suburban street with houses and a parked car appears, all painted with loose, colorful brushstrokes.
All models correctly identify both interior and exterior context. Differences are mainly stylistic; no major hallucinations.
SFT: I’m seeing several cars on a road, along with traffic lights and other vehicles.
A nighttime driving scene viewed from inside a car. The dashboard is visible, and ahead there’s heavy traffic with many red brake lights glowing, suggesting congestion or a stoplight on a busy road.
Gemini hallucinates dashboard map details. GPT and Gemini both assume “brake lights,” which may simply be tail lights at night. LLaMA-3 MM remains more conservative but slightly vague.
SFT: I see a peaceful street scene with buildings and trees in the background. The sky is cloudy, and the streetlights are glowing.
Q: You’re seeing a skyline (photo of PxHere, licensed under CC0).
DPO: You’re looking inside an elevator, where a small child is standing with a toy.
Q: What am I looking at? (AI augmented)
Q: What am I looking at? (AI generated) The image shows the interior of a car at night, facing a line of traffic. The digital dashboard displays a map and vehicle DPO: I’m seeing a line of vehicles on a road, stats, while the infotainment screen including several cars and trucks. There are shows various app icons. Through the traffic lights on the road and buildings windshield, you can see several cars around. with their brake lights illuminated at a stoplight. Q: What are you seeing? (photo captured by author)
4
Image
LLaMA3-MM
Gemini 3 Flash
GPT-5.3 Instant
Analysis
you’re looking at a 3D-rendered room featuring a unique visual glitch: a realistic ceiling fan casting a pixelated, Minecraft-style shadow. The scene includes a TV displaying DPO: You’re looking at a living room with "AFTERMATH," large speakers, a cat a TV, furniture, and various items. On the tree, and a desk, blending realistic TV, a movie or show is playing with the title textures with digital anomalies. ‘Aftermath’.
You’re looking at a living room. A TV is centered on a stand, showing a program. Surrounding it are speakers, shelves, and furniture like a chair and table. A ceiling fan with lights is above, casting shadows on the ceiling.
Gemini provides the most complete scene understanding including OCR text. LLaMA-3 MM is accurate but less detailed. GPT misses OCR signals.
SFT: You’re looking at a room that has several items such as furniture, decorations, and a door. DPO: You are looking at a watercolor painting of a room in front. The room has various items such as clothes, bottles, and a calendar. There’s also a shadow in the room that looks like the silhouette of a person.
You are looking at a watercolor painting of a room interior. It features a prominent dark, smoky wash on the ceiling above a cluster of gold and purple balloons. The scene includes a wooden door, an over-the-door organizer filled with tools, and a calendar hanging on the wall.
You’re looking at a stylized, watercolorlike interior scene. A door with a curtain is on the right, a refrigerator with magnets or photos on the left, and balloons hang near the ceiling. A dark cloud-like shape appears overhead, giving the scene a slightly surreal or moody atmosphere.
Gemini provides the most accurate and detailed description. GPT introduces hallucinated objects (refrigerator). LLaMA-3 MM is correct but less descriptive.
You are looking at an impressionist oil painting of a home interior. Thick, textured brushstrokes depict a living DPO: You’re looking at a painting of a area with a small wooden table, a larger room. The room has wooden flooring and dining table with chairs, and patterned curtains on the walls. The room also has curtains. The dark, reflective flooring some furniture, such as a table and a few contrasts with the soft, muted light chairs. filtering through the room.
You’re looking at a painterly interior of a living room or bedroom. Light filters through patterned curtains on the left, illuminating a small table and objects. On the right, a darker adjoining space with more furniture recedes, creating contrast between bright and shadowed areas.
GPT provides the most balanced and accurate description. Gemini slightly over-interprets objects. LLaMA-3 MM is the least detailed.
SFT: You’re looking at a living room with a TV, speakers, and furniture. On the TV, a movie or show is playing that shows a bridge.”
Q: What am I looking at? (AI generated)
Q: What am I looking at? (AI generated) SFT: You’re looking at a room with wooden flooring, curtains, and various furniture.
Q: Tell me what I’m looking at right now? (AI generated)
5
Image
LLaMA3-MM
Gemini 3 Flash
GPT-5.3 Instant
Analysis
SFT: You are looking at a bedroom with two You are looking at a bedroom people, perhaps changing clothes or illuminated with purple light. Two organizing. people are present: one is sitting on a bed with white linens, and another DPO: You are in a bedroom where two stands nearby wearing yellow shorts and people are clothed. One person is watching a dark shirt. A television is mounted on TV, and the other is perhaps looking for the wall, and a ceiling fan is visible something on the bed. above.
A bedroom scene with two people: one All exhibit some errors in standing near a wall-mounted TV adjusting spatial reasoning and or looking at it, and another bending over physical implausity. the bed. The bed is unmade, and a ceiling fan is above.
SFT: You’re looking at a fence on which vines or plants are growing. In the background, there’s an open area and a clear blue sky above.
You are looking at an outdoor scene on a bright, sunny day. In the foreground, there is a chain-link fence overgrown with dry, leafless vines. Beyond the fence is a flat, open field of dry grass or DPO: You’re looking at a chain-link fence dirt under a clear sky with a prominent on which vines or plants are growing. sun. Behind the fence, there’s an open area, and a bird is sitting on top of the fence. In the background, there’s a clear blue sky with the sun shining.
An outdoor scene with a chain-link fence in the foreground, dry grass and sparse vegetation beyond it, and a bright sun in a lightly clouded sky overhead.
SFT: HOLIDAY TOY DRIVE Experience the joy of giving.
HOLIDAY TOY DRIVE Experience the joy of giving.
Q: What am I seeing? (AI generated)
Q: What am I looking at? (photo captured by author)
HOLIDAY TOY DRIVE Experience the joy of giving.
DPO: HOLIDAY TOY DRIVE Experience the joy of giving. Buy toys to support military families through Operation Homefront.
Please purchase toys to support military families through Operation Homefront.
LLaMA-3 MM hallucinates a bird. GPT and Gemini remain grounded and accurate.
LLaMA-3 MM and Gemini provide accurate and faithful OCR. GPT adds There’s also a circular logo at the top right extra contextual content not that appears to include: present in the image. U.S. Marine Corps Reserve – Toys for Tots (based on the recognizable train logo)
Q: read the text (brochure of Operation Homefront)
Note: (1) the underlined words are hallucinations, (2) Some models tend to generate excessively long responses; therefore, prompts explicitly constrain outputs to approximately 50-60 words or fewer, (3) due to stochastic generation, evaluation noise, image resolution variations, and other nondeterministic factors, exact responses may not be fully reproducible across runs.
6