Look Clearly Before Answering: Mitigating Hallucinations in LVLMs via Saliency-Driven Perceptual Realignment

🕒 Published (v1): 2026-07-18 14:39 UTC · Source: Arxiv · link

Why this paper was selected

Saliency-driven hallucination mitigation in LVLMs; core reliability problem, buildable method

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

SDPR (Saliency-Driven Perceptual Realignment) is a training-free inference-time framework that mitigates hallucinations in Large Vision-Language Models (LVLMs) by diagnosing and correcting visual awareness degradation at three stages: perception (attention hijacking by sink tokens), memory (KV cache spatial distortion), and decoding (language-prior dominance). It achieves 6.93% improvement on POPE and 33.3% reduction in \(\text{CHAIR}_s\) over prior SOTA without any fine-tuning.

Problem

LVLMs hallucinate because visual grounding progressively erodes during inference. Two underexplored mechanisms are identified: (1) non-semantic "sink tokens" hijack attention in MHSA heads, suppressing salient visual cues (hallucinatory responses show up to \(1.79\times\) higher top-\(k\) attention concentration than faithful ones); (2) causal prefill attention asymmetrically contextualizes visual tokens in the KV cache — earlier visual tokens cannot access later visual or query tokens — inducing position-dependent spatial distortion (Spearman rank correlation between PCA scores of cached visual keys and token sequence positions is \(1.62\times\) higher in hallucinatory than faithful responses). Existing decoding-based or attention-intervention methods address only one of these failure modes.

Method

SDPR intervenes at three stages without any training:

Stage 1 — Saliency-Driven Attention Redistribution (SDAR): Select the \(k\) lowest-entropy MHSA heads (entropy \(E_h^l = -\sum_{i=1}^{N_v} AW_{h,i}^{v,l} \log AW_{h,i}^{v,l}\)) as putative sink-dominated heads. Aggregate their visual attention maps to locate a top-\(K\) candidate set \(\mathcal{C}^l\). For each selected head, replace each candidate's attention score with the average over its center-inclusive \(3\times3\) spatial neighborhood, then Min-Max scale and \(L_1\)-normalize. This weakens isolated non-semantic peaks while preserving coherent semantic activations — avoiding the semantic-agnostic failure of pure magnitude-based removal.

Stage 2 — Saliency-Driven Cache Alignment (SDCA): At a designated intermediate layer \(l_s\), compute a query-visual saliency map $\(S_q = \sigma\!\left(\bar{h}_q \,(H_v^{l_s})^\top\right), \quad \bar{h}_q = \frac{1}{N_q}\sum_{i=1}^{N_q} H_{q,i}^{l_s}\)$ where \(\sigma\) is Min-Max normalization. Broadcast \(S_q\) to recalibrate the cached visual keys \(K_v^{l_s}\), injecting explicit query conditioning to counteract position-dependent distortion from asymmetric prefill contextualization.

Stage 3 — Prior-Constrained Contrastive Decoding (PCD): Construct a language-prior reference from early-layer hidden states (isolating intrinsic language-prior predictions). Contrastively subtract this reference from the original logits to penalize predictions dominated by language priors over visual evidence. Adaptive decoding weights control the contrastive strength.

Key Contributions

• Systematic empirical characterization of two previously underexplored visual degradation mechanisms in LVLMs: attention-sink hijacking and asymmetric KV cache contextualization, with quantitative evidence linking both to hallucination rates.
• SDPR: a training-free, holistic framework that jointly addresses perception, memory, and decoding-stage degradation in a single inference pass.
• The SDAR component introduces spatial-context-guided redistribution that avoids the semantic-agnostic failure mode of naive magnitude-based attention removal.
• Demonstrated generalizability across diverse LVLM architectures (LLaVA-1.5, Qwen-VL, InternVL) with minimal runtime overhead.

Results

• POPE hallucination benchmark: +6.93% over prior SOTA methods.
• CHAIR\(_s\): 33.3% reduction relative to prior SOTA.
• Outperforms existing training-free methods (VCD, ONLY, and attention-intervention baselines) on both hallucination-specific and general-purpose benchmarks.
• Evaluation conducted on MME and other multimodal benchmarks across multiple LVLM architectures.
• Quantitative analysis: hallucinatory responses exhibit \(1.71\times\)–\(1.79\times\) higher Gini/top-\(k\) attention concentration; \(1.62\times\) higher Spearman sequence-position correlation in cached visual keys.

(Note: full per-benchmark tables were truncated in the provided text; only headline figures are reported above.)

Limitations

• All three components introduce hyperparameters (\(k\) sink-head count, \(K\) top candidate tokens, intermediate layer \(l_s\), contrastive weight \(\lambda\)) whose optimal values may vary across architectures and tasks.
• The SDAR spatial-averaging heuristic assumes visual tokens are arranged in a 2D grid, which may not hold for all visual tokenization schemes.
• PCD's language-prior reference is constructed from early-layer hidden states; the quality of this proxy depends on model depth and architecture, potentially limiting transferability to very shallow or non-standard LVLMs.
• Evaluation is restricted to existing LVLM families (LLaVA, Qwen-VL, InternVL); applicability to models with interleaved vision-language architectures or different prefill strategies is not addressed.
• Training-free design cannot correct for hallucinations rooted in fundamental pretraining data biases.

Relevance to Vision-Language Models

SDPR directly targets the inference dynamics of transformer-based LVLMs, making it relevant to any practitioner deploying models like LLaVA, Qwen-VL, or InternVL without the budget for RLHF or preference-tuning. The mechanistic analysis — linking attention sink patterns and KV cache spatial distortion to hallucination rates with concrete metrics (Gini, Spearman correlation) — provides a reproducible diagnostic toolkit for the VLM hallucination literature. The three-stage decomposition (perception → memory → decoding) offers a principled framework that complements existing decoding-only methods (VCD) and attention-only interventions, suggesting that comprehensive hallucination suppression requires addressing all three loci simultaneously. This work advances the training-free inference-time intervention line alongside ICD, OPERA, and VCD, while being the first to explicitly model KV cache distortion as a distinct and quantifiable hallucination source.