Organizations: King Fahd University of Petroleum & Minerals, Saudi Arabia · SDAIA-KFUPM Joint Research Center for Artificial Intelligence, Saudi Arabia
Large vision-language models can produce fluent responses containing objects, attributes, or relations unsupported by the image. We study an implemented alignment pipeline that normalizes released fine-grained hallucination supervision, validates chosen-rejected preference pairs with vision-language judges, repairs rejected chosen responses, and trains LLaVA adapters with standard or severity-aware direct preference optimization. The study uses 16,143 normalized feedback records and 8,386 released preference pairs, comparing direct training with zero-shot and two-shot judged severity-margin variants. The zero-shot variant reduces Object HalBench CHAIRs from 53.00 to 36.00 and CHAIRi from 15.72 to 10.41. Across aligned models, AMBER-Hal decreases from 35.90 to 25.10-26.90, accompanied by reduced visual coverage. In contrast, the base model retains the highest valid POPE adversarial F1, 84.18 versus 83.87-83.91 after alignment. These results support benchmark-dependent reductions in generative hallucination rather than uniform superiority of severity-aware alignment. We distinguish implemented pipeline components from the preference paths used in the reported experiments and discuss the limits of attributing gains to judging, repair, or the training objective individually.
Figures & tables
Fig. 1: Overview of the implemented fine-grained hallucination-mitigation pipeline.
Path
Source
Accepted
Rejected
Repaired
Approved repairs
Verified
Train-ready
Zero-shot
8,386
7,818
568
473
244
8,062
8,062
Two-shot
8,386
7,132
1,254
1,063
746
7,878
7,874
TABLE I: Judged preference construction for zero-shot and two-shot paths.
Model
Accuracy
Precision
Recall
F1
LLaVA-1.5-7B
85.13
89.92
79.13
84.18
Standard DPO
85.00
90.63
78.07
83.88
Direct HSA-DPO
85.03
90.70
78.07
83.91
Judged margin HSA-DPO-M (zero-shot)
84.93
90.25
78.33
83.87
TABLE II: POPE adversarial results (%; higher is better).
Fig. 2: POPE F1 and generative hallucination scores. Two-shot POPE F1 is unavailable.
Object HalBench
AMBER
Model
CHAIRs
CHAIRi
CHAIR
Hal
Cog
Cover
Base LLaVA
53.00
15.72
7.70
35.90
4.30
51.70
Standard DPO
37.67
11.98
5.20
25.20
2.50
49.90
Direct HSA-DPO
38.00
11.74
5.50
25.10
2.50
50.00
Zero-shot HSA-DPO-M
36.00
10.41
5.30
25.50
2.50
49.60
Two-shot HSA-DPO-M
40.33
12.81
5.60
26.90
2.70
50.40
TABLE III: Local generative evaluation: lower is better except for AMBER-Cover. Bold marks numerical best values.
TABLE V: Derived preference-retention rates (%). Repair approval uses repaired candidates as its denominator.
Model
POPE F1
CHAIRs
CHAIRi
AMBER-CHAIR
Hal
Cog
Cover
Standard DPO
-0.30
-15.33
-3.74
-2.50
-10.70
-1.80
-1.80
Direct HSA-DPO
-0.27
-15.00
-3.98
-2.20
-10.80
-1.80
-1.70
Zero-shot HSA-DPO-M
-0.31
-17.00
-5.31
-2.40
-10.40
-1.80
-2.10
Two-shot HSA-DPO-M
—
-12.67
-2.91
-2.10
-9.00
-1.60
-1.30
Appendix
TABLE VI: Signed metric changes relative to base LLaVA in reported score units. A negative change favors hallucination metrics, but is unfavorable for F1 and Cover. Missing two-shot F1 is not imputed.
Vision-Language Models (VLMs) have shown strong performance in visual understanding, yet they still suffer from hallucinations, generating content that is not grounded in the image. Preference alignment is a promising approach to improve visual faithfulness, but its success depends heavily on how preference pairs are constructed. Existing methods exhibit two key limitations; (a) intervention-based methods often introduce significant deviation from the policy distribution, and (b) sampling-based methods often underuse visual information during the construction. In this paper, we propose ViPSy (Vision-driven Preference Synthesis), a framework for constructing preference data that are both policy-aligned and visually grounded. Our framework consists of two stages; in the first stage, ViPSy derives a visual cue from recurring object-level content across semantically aligned image variants, so preference construction can rely on visual information rather than language priors. In the second stage, ViPSy conditions the policy's own rollouts on this cue, allowing candidates to be guided by visually grounded content while staying close to the policy's response distribution. The resulting candidates remain close to the policy's response distribution while better leveraging visual information from the image. Experiments show that the resulting VLM, preference-aligned with ViPSy-constructed preference pairs, achieves a new state-of-the-art in hallucination mitigation. Compared with the previous state-of-the-art method, it reduces hallucination rates on AMBER and Object HalBench by 35.7% and 24.5%, respectively. The resulting model further improves on general visual grounding benchmarks, e.g., MMStar, MMVP, and CV-Bench, while also yielding gains in semantic segmentation and ImageNet linear probing, underscoring the effectiveness of our framework in enhancing the model's visual capabilities.
Vision-Language Models (VLMs) increasingly power high-stakes applications, from medical imaging to autonomous systems, yet they routinely hallucinate, confidently describing content not present in the input. We investigate the root causes of these failure modes with a mechanistic analysis focusing on the decoder-based VLMs. We trace these failure modes to a geometric over-alignment: to bridge the modality gap required by attention mechanisms, decoder-based VLMs over-align visual embeddings with the text manifold, injecting a statistical linguistic bias that systematically overshadows fine-grained visual evidence. While prior work either aggressively closes this gap or suppresses hallucinations through expensive black-box decoding strategies, none addresses the underlying geometric cause. We provide the first quantitative characterization of this over-alignment, demonstrating that linguistic bias concentrates in the top principal components of a universal, dataset-agnostic text subspace. Building on this insight, we propose two complementary remedies: a training-free inference strategy and a bias-aware fine-tuning paradigm, both of which explicitly project out this subspace from visual representations. Our methods significantly reduce hallucinations across POPE, CHAIR, and AMBER benchmarks, and improve CLAIR scores on long-form captioning tasks, with the training-free variant adding no computational overhead over the base model.
Harshvardhan Saini, Samyak Jha, Yiming Tang +1
Indian Institute of Technology Dhanbad · National University of Singapore
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.