Mitigating Multimodal Hallucination via Phase-wise Self-reward
| Authors | Yu Zhang et al. |
| Year | 2026 |
| HF Upvotes | 3 |
| arXiv | 2604.17982 |
| Download | |
| HF Page | View on Hugging Face |
Abstract
Large Vision-Language Models (LVLMs) still struggle with vision hallucination, where generated responses are inconsistent with the visual input. Existing methods either rely on large-scale annotated data for fine-tuning, which incurs massive computational overhead, or employ static post-hoc strategies that overlook the dynamic nature of hallucination emergence. To address these, we introduce a new self-rewarding framework, enabling dynamic hallucination mitigation at inference time without external supervision. On the empirical side, we reveal that visual hallucination exhibits phase-wise dynamic patterns, peaking at the onset of each semantic phase. Drawing on these insights, we propose PSRD (Phase-wise \textbf{Self-Reward Decoding) for online hallucination correction guided by phase-wise self-reward signals. To reduce the cost of repeated self-evaluation during decoding, we distill the hallucination guidance signal from LVLMs into a lightweight reward model. The reward model subsequently provides on-the-fly guidance for targeted intervention during the decoding process, enabling precise hallucination suppression. The proposed PSRD significantly reduces the hallucination rate of LLaVA-1.5-7B by 50.0% and consistently outperforms existing post-hoc methods across five hallucination evaluation benchmarks for four LVLMs. Further analysis confirms that PSRD effectively mitigates hallucination propagation and achieves a highly controllable trade-off between strong performance and inference efficiency.
Engineering Breakdown
Plain English
This paper addresses a critical problem in large vision-language models (LVLMs) where they generate text that contradicts what they actually see in images—a phenomenon called vision hallucination. Rather than requiring expensive fine-tuning on massive annotated datasets, the authors discovered that hallucination follows predictable patterns that intensify at specific points during text generation (the start of each semantic phase). They propose PSRD (Phase-wise Self-Reward Decoding), a method that corrects hallucinations on-the-fly during inference by using internal self-reward signals, without needing external supervision or retraining.
Core Technical Contribution
The key novelty is identifying that vision hallucination is not a uniform phenomenon but exhibits phase-wise dynamics—it peaks when the model begins generating a new semantic concept. This insight enables a fundamentally different approach: instead of static post-hoc filtering or expensive supervised fine-tuning, the authors introduce a self-rewarding mechanism that operates during decoding to dynamically detect and correct hallucinations as they emerge. PSRD uses phase-aware signals to guide the model's generation in real-time, making hallucination mitigation a problem of inference-time control rather than model retraining. This is architecturally novel because it treats hallucination correction as an online optimization problem with learned reward signals.
How It Works
The system operates during the decoding phase of text generation from a vision-language model. As the model generates tokens sequentially, PSRD monitors the semantic phases—natural breakpoints where the model transitions from discussing one visual concept to another. At the onset of each phase, the self-reward mechanism computes a score indicating hallucination likelihood based on alignment between the generated text and visual features extracted from the input image. When hallucination risk peaks (typically at phase boundaries), the decoding strategy adjusts—for example, by increasing sampling diversity, constraining token probabilities to vision-grounded vocabularies, or backtracking to earlier less-risky branches. The self-reward function is learned from data without requiring external annotators, using contrastive signals from the model's own reasoning.
Production Impact
For engineers deploying LVLMs in production, this eliminates the need to maintain separate fine-tuned model variants for hallucination reduction, which typically costs 2-4x the compute of inference. PSRD operates entirely at inference time, meaning existing deployed models can be augmented with the phase-wise self-reward module without retraining. The approach reduces memory footprint compared to fine-tuning approaches since you're not storing multiple checkpoint versions. Latency impact is modest—additional overhead comes from computing self-reward scores and phase-boundary detection, but this is cheaper than decoding longer incorrect sequences that would later require filtering. Integration is straightforward: wrap your existing LVLM decoder with the PSRD mechanism that observes semantic phases and adjusts generation. Trade-offs include slightly increased per-token compute and the need to learn phase boundaries for your specific task domain.
Limitations and When Not to Use This
The paper assumes hallucination follows consistent phase-wise patterns, which may not hold for all domains or image types—highly abstract images or unconventional layouts might not have clear semantic phases. Self-reward learning still requires some training data with hallucination labels, though significantly less than supervised fine-tuning; the paper doesn't fully clarify how much labeled data is needed or how to bootstrap in entirely new domains. The approach is tailored to text generation after image understanding; extending it to other multimodal tasks (video, audio-visual, 3D) is unexplored. Evaluation is likely limited to existing vision-language benchmarks, so real-world performance on adversarial inputs, out-of-distribution images, or edge cases where hallucination patterns differ remains unclear. The phase-wise assumption may also fail in streaming or interactive settings where semantic boundaries are ambiguous.
Research Context
This work builds on the growing understanding that hallucination in LVMs is a structural problem requiring inference-time intervention, moving beyond early approaches that treated it as a training-time issue. It extends self-rewarding concepts (similar to those in recent RLHF and direct-preference-optimization work) to the multimodal setting, specifically addressing vision grounding. The paper likely improves on existing benchmarks like MMVP (Multimodal Vision Perception), POPE (Poll-based Object Probing Evaluation), or CHAIR (Caption Hallucination Assessment), which measure how much generated descriptions contradict visual input. It opens a research direction toward learnable, adaptive inference-time constraints for multimodal models—moving toward a future where hallucination mitigation is a deployed component rather than a data-hungry retraining process. This complements orthogonal lines of work in mixture-of-experts routing and dynamic computation for efficient inference.
:::tip Subscribe Get weekly breakdowns of papers like this in AI Letters - the newsletter for engineers building production AI systems. :::