Mitigating Multimodal LLMs Hallucinations via Relevance Propagation at Inference Time

Technical Summary: Mitigating Multimodal LLMs Hallucinations via Relevance Propagation at Inference Time

Problem Statement
Multimodal Large Language Models (MLLMs) have demonstrated significant capabilities in integrating language with perceptual inputs such as images and audio. However, a critical limitation persists: these models frequently suffer from hallucinations, generating outputs inconsistent with or unsupported by the provided perceptual evidence. The authors posit that this phenomenon stems from an inherent imbalance in modality utilization during inference. Specifically, textual tokens tend to dominate the generation process, while visual or auditory tokens receive insufficient attention. Consequently, models often rely on strong textual language priors rather than grounded perceptual evidence, leading to plausible but factually incorrect descriptions of non-existent objects or events.

Methodology: Learning Inference-time Modality Enhancement (LIME)
To address this imbalance without modifying the underlying model parameters or requiring additional training data, the authors propose LIME, a training-free framework that operates at inference time. The method is grounded in an interpretability analysis using Layer-wise Relevance Propagation (LRP).

1. LRP Analysis: The authors employ LRP to decompose model predictions into additive contributions of individual input tokens. Their analysis reveals a systematic bias where textual tokens are assigned significantly higher relevance scores than modality tokens, even in tasks critically dependent on perceptual input.
2. Inference-Time Optimization: LIME intervenes directly in the model's key-value (KV) representations during the decoding process. Instead of fine-tuning weights, it iteratively optimizes additive updates ($\Delta K, \Delta V$) to the key and value tensors at each decoding step.
3. Objective Function: The optimization aims to shift the model's attention toward modality tokens while preserving the original model's linguistic priors. This is achieved by minimizing a composite loss function:
   • Relevance Objective ($\mathcal{L}_{rel}$): Inspired by Noise Contrastive Estimation (NCE) and contrastive learning, this term treats modality tokens as positive evidence and textual tokens as competing alternatives. It maximizes the relevance assigned to modality tokens relative to the total relevance of all tokens.
   • KL-Divergence Regularizer ($\mathcal{L}_{KL}$): This term constrains the updated distribution to remain close to the original model's distribution, preventing the model from deviating too far from its pre-trained behavior and ensuring decoding stability.
4. Execution: The updates are computed via gradient-based optimization (using Adam) for a small number of steps at each decoding step and are discarded afterward, ensuring the intervention remains local and does not accumulate over time.

Key Contributions
The paper outlines three primary contributions:

1. Interpretability Analysis: An LRP-based analysis demonstrating that hallucinations in MLLMs are directly associated with the under-utilization of modality-specific tokens during inference.
2. LIME Framework: A novel, training-free, relevance-guided inference-time framework that mitigates hallucinations by explicitly increasing the contribution of modality tokens relative to textual inputs through KV representation adjustments.
3. Empirical Validation: Demonstration of the method's effectiveness and generality across multiple state-of-the-art multimodal models (including LLaVA, Qwen-VL, SALMONN, and Qwen2-Audio) and benchmarks in both vision and audio domains.

Results
The authors evaluated LIME on several benchmarks, reporting consistent improvements in grounding and reductions in hallucinations:

• Vision Benchmarks: On the POPE benchmark (MSCOCO and A-OKVQA), LIME achieved the highest average accuracy and F1 scores among training-free baselines (e.g., OPERA, VCD, ICD, MemVR, V-ITI), particularly showing robustness on adversarial splits. On the CHAIR benchmark, LIME consistently reduced the rate of hallucinated object mentions (lower CHAIRS and CHAIRI scores) while maintaining competitive recall.
• Audio Benchmarks: On Audio Hallucination QA and AIR-Bench, LIME improved accuracy and F1 scores across speech, sound, and music domains, outperforming the Audio-Aware Decoding (AAD) baseline.
• Relevance Analysis: Quantitative analysis confirmed that LIME increases modality reliance (the proportion of relevance attributed to perceptual tokens) and improves spatial/temporal grounding (the concentration of relevance on ground-truth object regions). Qualitative visualizations show that relevance heatmaps become more localized and semantically aligned with the input evidence under LIME.
• Computational Overhead: The method introduces additional latency (approximately 9-10x slowdown) and modest memory overhead due to the per-step optimization, making it suitable for offline or high-accuracy settings rather than real-time low-latency applications.

Significance and Claims
The paper claims that directly controlling modality utilization at inference time is an effective strategy for enhancing the reliability of MLLMs. By leveraging LRP signals to rebalance the influence of textual and perceptual tokens, LIME offers a generalizable solution to hallucinations that does not require the computational cost of retraining or fine-tuning. The authors conclude that hallucinations are not solely a result of insufficient training data but also arise from suboptimal internal reasoning and information integration during decoding, which can be corrected through inference-time intervention. The method is presented as a practical approach for improving multimodal grounding while preserving the generative quality of the original model.