Imitating the Truth: Attention-aware Truth-Guided Enhancement for Hallucination Mitigation in Large Vision-Language Models

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

ICLR 2026; attention-aware truth-guided enhancement for LVLM hallucination

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

AGE (Attention-aware Truth-Guided Enhancement) is a training-free decoding-time framework that mitigates hallucinations in LVLMs by guiding models to imitate the stage-specific attention patterns of real (non-hallucinated) tokens. The core insight is that real and hallucinated tokens exhibit systematic, model-specific divergences in visual and textual attention across distinct layer stages, and correcting these divergences reduces hallucination. AGE achieves state-of-the-art results on COCO captioning, POPE, and MME without retraining.

Problem

Existing hallucination mitigation methods for LVLMs either require auxiliary modules or apply coarse, layer-uniform attention adjustments (e.g., VCD, OPERA). These approaches ignore the fine-grained, stage-dependent dynamics that differentiate real tokens from hallucinatory ones, leading to suboptimal and sometimes counterproductive corrections.

Method

AGE performs a preparatory attention discrepancy analysis: given \(N=100\) COCO images known to elicit hallucinations, it labels each generated token as real or hallucinated by comparing predicted objects against ground-truth annotations. For each layer \(l\), it computes per-modality average attention sums over real vs. hallucinated token sets and aggregates the difference across samples: $\(\text{Diff}^l_{\text{image}} = \frac{1}{N}\sum_{i=1}^{N}\left(\bar{s}^{(l,i)}_{\text{real,vision}} - \bar{s}^{(l,i)}_{\text{hall,vision}}\right)\)$ This reveals model- and stage-specific gaps partitioned into Early (layers 0–16), Middle (16–26), and Late (26–31) stages.

Two interventions are applied during inference:

1. Imitating Image Attention (universal, Late stage): A directional correction vector \(\delta \in \mathbb{R}^n\) is pre-computed from \(M=10\) hallucination-eliciting COCO samples as a weighted average of per-sample real-minus-hallucinated attention vectors at the final decoder layer: $\(\delta = \frac{1}{M}\sum_{i=1}^{M} w_i \cdot \left(a^i_{\text{real,vision}} - a^i_{\text{hall,vision}}\right)\)$ At inference, \(\hat{a}^l_{\text{vision}} = a^l_{\text{vision}} + \lambda_v \cdot \delta\) (with \(\lambda_v = 100\)) is injected into Late-stage layers (30, 31).

2. Imitating Text Attention (LLaVA-specific, Middle stage): Since text attention vectors vary in dimension, a self-multiplicative amplification is used instead of a fixed directional vector: \(\hat{a}^l_{\text{text}} = a^l_{\text{text}} + \lambda_t \cdot a^l_{\text{text}}\) (with \(\lambda_t = 3\), applied at layer 20).

For MiniGPT-4 and mPLUG-Owl2, where real responses consistently dominate visually across stages, only the image attention intervention is applied.

Key Contributions

• Token-level, layer-wise analysis identifying that hallucinations arise from failure to replicate stage-specific, model-specific attention dynamics of real tokens—not uniform lack of visual grounding.
• A directional visual attention correction vector \(\delta\) derived from only 10 calibration images, avoiding dependence on large auxiliary datasets.
• Stage-specific intervention design that adapts to each model's unique attention profile (image-only vs. image+text dual intervention).
• Training-free, architecture-agnostic application that preserves fluency while reducing hallucination.

Results

• COCO Image Captioning (CHAIR, \(N=500\) val images, max 64 tokens):
• Average \(\text{CHAIR}_S\): AGE = 17.15 vs. previous SOTA VISTA = 20.10 (−2.95 pp)
• Average \(\text{CHAIR}_I\): AGE = 6.35 vs. VISTA = 6.45 (−0.10 pp)
• Average BLEU: AGE = 15.56 vs. baseline greedy = 15.21 (+0.35 pp)
• On MiniGPT-4 alone: \(\text{CHAIR}_S\) 15.62 vs. VISTA 19.80; \(\text{CHAIR}_I\) 6.00 vs. 6.00
• POPE (MiniGPT-4, 3,000 binary questions):
• Average accuracy improvement over baseline: +17.09%
• Over OPERA: +20.09% accuracy, +1.25% F1
• MME (Existence, Count, Position, Color): state-of-the-art hallucination accuracy reported (full table truncated in provided text).
• Effective with as few as \(M=10\) calibration images; scales to 13B model variants.

Limitations

• The \(\delta\) vector and stage boundaries require per-model calibration runs on a small COCO subset; not fully zero-shot for new architectures.
• Text attention intervention design (self-multiplicative) is direction-agnostic and motivated empirically rather than derived from a closed-form discrepancy signal, potentially sub-optimal.
• Analysis confined to three 7B-class models (LLaVA-1.5, MiniGPT-4, mPLUG-Owl2); generalization to instruction-tuned models with different architectures (e.g., Flamingo-style cross-attention) is unvalidated.
• Hyperparameters (\(\lambda_v = 100\), \(\lambda_t = 3\), layer indices) are manually tuned per model; sensitivity analysis is in the appendix but not fully reported in the main text.
• BLEU as a fluency proxy is a weak signal; human evaluation of output quality is absent.

Relevance to Vision-Language Models

AGE provides interpretable, mechanistic evidence that hallucinations in LVLMs are not monolithic failures but stage-localized attention breakdowns, which directly challenges the design assumption of prior coarse decoding methods like VCD and OPERA. For VLM researchers, the layer-wise real-vs-hallucinated attention analysis framework is a reusable diagnostic tool applicable to any transformer-based LVLM. The result that \(M=10\) calibration images suffice to compute an effective correction vector implies the signal is concentrated and robust, informing future work on efficient alignment. This work sits within the growing line of inference-time intervention methods for VLMs and offers a principled alternative to contrastive decoding that does not require image distortion or token-level reranking.