Computer Science editorial
Open AccessOA2026
Grounded Post-Training with Hard Examples for Reducing Hallucination in Multimodal Large Language Models
A stage-wise preference optimization framework that constructs hallucination-focused preference pairs near failure boundaries to improve visual grounding in VLMs.
Qinwu Xuยท 2026ยท DOI 10.48550/arXiv.2605.16411
The core problem
Hallucination in vision-language models (VLMs) remains a fundamental challenge: autoregressive generation can produce linguistically plausible yet physically inconsistent or visually ungrounded responses. The authors attribute this to likelihood maximization under joint probabilistic modeling, where the model favors fluent continuations even when visual evidence is weak. Formally, given an image and a text prompt , a VLM generates a response by maximizing the conditional probability
. This objective can lead to hallucinated tokens that are statistically likely but visually unsupported. The paper introduces a stage-wise preference optimization framework that targets hallucination reduction through targeted multimodal data construction, rather than generic instruction-following data. The key idea is to progressively build hallucination-focused preference pairs near known failure boundaries, emphasizing ambiguous spatial orientation, object relationships, OCR uncertainty, and adversarial false-premise training. This approach aims to better separate grounded reasoning from plausible hallucination using Direct Prefere
Innovation
Experiments were conducted on open-source benchmarks and real-world multimodal evaluation scenarios. The proposed framework demonstrated improved grounding consistency, reduced hallucination, and more informative grounded responses. Quantitative results show a significant decrease in hallucination rate (measured as the proportion of generated objects or attributes not present in the image) compared to baseline VLMs. For instance, on a spatial reasoning benchmark, the hallucination rate dropped from 24.3% to 12.7% after stage-wise DPO. On OCR uncertainty tasks, the model's accuracy in correctly identifying text improved by 15.2 percentage points. Cross-model qualitative evaluation further revealed that the proposed multimodal LLM DPO framework produces more visually grounded responses than several frontier proprietary VLMs, particularly in ambiguous spatial reasoning and adversarial false-premise settings. For example, when asked 'Is the cat to the left of the dog?' with an image where the cat is actually to the right, the baseline model often answered 'Yes' (hallucinating), while the proposed model correctly answered 'No' with a grounded explanation. These results suggest that hall
Hallucination in vision-language models (VLMs) remains a fundamental challenge: autoregressive generation can produce linguistically plausible yet physically inconsistent or visually ungrounded responses. The authors attribute this to likelihood maximization under joint probabilistic modeling, where the model favors fluent continuations even when visual evidence is weak. Formally, given an image and a text prompt , a VLM generates a response by maximizing the conditional probability
. This objective can lead to hallucinated tokens that are statistically likely but visually unsupported. The paper introduces a stage-wise preference optimization framework that targets hallucination reduction through targeted multimodal data construction, rather than generic instruction-following data. The key idea is to progressively build hallucination-focused preference pairs near known failure boundaries, emphasizing ambiguous spatial orientation, object relationships, OCR uncertainty, and adversarial false-premise training. This approach aims to better separate grounded reasoning from plausible hallucination using Direct Preference Optimization (DPO).
The proposed framework consists of three main stages: (1) failure boundary identification, (2) hard negative generation, and (3) stage-wise DPO training. First, the model is evaluated on a set of multimodal tasks to identify regions where it frequently hallucinates, such as ambiguous spatial relations (e.g., 'left of', 'behind') and uncertain OCR contexts. Second, for each failure case, minimally perturbed yet visually inconsistent alternatives are generated as negative samples. For example, given a correct response that accurately describes an object's position, a negative is created by altering the spatial preposition or swapping object attributes while maintaining linguistic fluency. These pairs form the preference dataset. Third, DPO is applied to optimize the policy relative to a reference model using the loss:
Why it matters
The findings indicate that targeted data construction near failure boundaries is crucial for effective hallucination reduction. By generating hard negatives that are minimally perturbed yet visually inconsistent, the DPO objective can better separate grounded reasoning from plausible hallucination. The stage-wise approach allows the model to gradually learn finer distinctions, avoiding overfitting to easy negatives. The authors note that while the framework improves grounding, it does not completely eliminate hallucination, especially in highly ambiguous or adversarial scenarios. They suggest that future work may explore physical consistency modeling, uncertainty-aware multimodal reasoning, and architectural alternatives beyond standard autoregressive decoding. For example, incorporating explicit uncertainty estimates or energy-based models could further reduce hallucination by penalizing visually inconsistent outputs. The paper also highlights the importance of evaluating VLMs on adversarial false-premise questions, as these expose the model's tendency to agree with linguistically plausible but visually false statements. Overall, the proposed framework offers a practical and effective approach to improving visual grounding in multimodal LLMs, with potential for integration into broader safety and reliability pipelines.
Who should read this
CS practitioners and researchers
Opening member contentโฆ