Skip to content

On Epistemic Uncertainty of Visual Tokens for Object Hallucinations in Large Vision-Language Models

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

High epistemic uncertainty in visual tokens produced by the vision encoder is a key driver of object hallucination in LVLMs. The authors propose a training-free method that identifies uncertain tokens via PGD adversarial perturbations and suppresses them by masking their self-attention outputs in intermediate vision encoder layers. This approach reduces CHAIR hallucination scores by up to 38% across multiple architectures and composes additively with existing LLM-side mitigation methods.

Problem

Prior training-free hallucination mitigations target the language model (decoding strategies, LLM attention) and incur multiple expensive LLM forward passes or yield modest gains; the vision encoder as a source of hallucination is underexplored. No efficient method existed to identify which individual visual tokens are epistemically uncertain without thousands of MC-dropout forward passes.

Method

Uncertainty estimation via adversarial attack. Given image \(x\) and vision encoder \(F_V\), the method runs PGD (\(k{=}3\), \(I{=}200\) steps) to maximise MSE between clean and perturbed features. Hidden states are extracted from early layers \(S{=}\{1,\ldots,10\}\) for both original and attacked images. Layer-wise L2 deviation norms \(u^l = \|f^l_\text{attk} - f^l_\text{orig}\|_2\) are min-max normalised and averaged into uncertainty map \(U\), then binarised with threshold \(\sigma_{th}\) to produce binary mask \(M\) (0 = uncertain token). Theorem 3.2 proves that the expected squared deviation upper-bounds the differential entropy of each token's representation, and Lemma 3.1 justifies local Gaussianity under small perturbation — grounding the proxy in information theory.

Masking. Uncertain tokens (where \(M{=}0\)) are suppressed by element-wise multiplying the self-attention value output by \(M\) in intermediate VE layers (13–17 for LLaVA-1.5/Shikra; 9–16 for MiniGPT-4). The residual connection preserves global semantics while reducing reliance on high-uncertainty tokens. No LLM weights are modified.

Key Contributions

  • Theoretical proof (Lemma 3.1 + Theorem 3.2) that adversarial representation deviation upper-bounds visual token epistemic uncertainty, with empirical validation against MC dropout.
  • Statistical confirmation (Spearman \(\rho > 0.7\), \(p < 0.05\)) that uncertain visual tokens correlate with object hallucination severity on COCO/CHAIR.
  • Training-free, VE-only intervention: PGD proxy + mid-layer attention masking, ~5× faster than MC dropout.
  • Demonstrated composability with OPERA, VCD, PAI, and Devils (LLM-side methods), producing additive gains.

Results

  • LLaVA-1.5-7B, CHAIR (greedy): \(C_s\) 47.4→29.2 (↓38.4%), \(C_i\) 12.2→9.3 (↓23.8%); F1 preserved (77.9→78.2).
  • Shikra-7B, CHAIR (greedy): \(C_s\) 58.0→43.2 (↓25.5%), \(C_i\) 15.6→11.7 (↓25.0%).
  • MiniGPT-4, CHAIR (greedy): smaller gains (\(C_s\) ↓4.2%), attributed to Q-Former bottleneck between VE and LLM.
  • POPE: comparable or slightly improved across all models and splits; no degradation.
  • AMBER (LLaVA-1.5-7B, greedy): CHAIR ↓23.9%, hallucination ↓19.9%, AMBER score 88.2→91.2 (+3.4%).
  • Composing with PAI yields up to ↓28.2% CHAIR on AMBER; integration with all four baselines uniformly adds gains.
  • Ablation: early VE layers (1–10) for mask generation and mid-layers (13–17) for application are jointly optimal; attention-level hard masking outperforms input masking, output masking, MLP masking, and soft masking.

Limitations

  • Gains on MiniGPT-4 are substantially smaller because its Q-Former compresses visual tokens before the LLM, limiting the propagation of VE-level suppression.
  • Requires running a PGD attack (200 steps) at inference time per image, adding latency (~2.4 s vs. ~12.2 s for MC dropout in examples shown, but still non-trivial overhead).
  • \(\sigma_{th}\) must be tuned per baseline–method pair, reducing plug-and-play simplicity.
  • Theoretical analysis assumes small perturbation and \(C^2\)-smooth networks; validity under very large \(k\) or non-standard VE architectures is not analysed.
  • Evaluated only on object hallucination; effects on other hallucination types (attribute, relation) are not addressed.

Relevance to Vision-Language Models

This paper directly targets a core reliability failure mode of LVLMs — object hallucination — by reframing it as a vision encoder uncertainty problem rather than a language decoding problem, which is the dominant framing in the literature. It introduces a practical, training-free diagnostic (PGD-based epistemic uncertainty proxy) that reveals which image regions the VE is least confident about, and shows these regions causally drive hallucinations. The orthogonality to LLM-side methods makes it immediately composable with the existing ecosystem of OPERA, VCD, and attention-based fixes. For researchers tracking VLMs, this work opens a new axis of intervention — vision encoder intermediate representations — and provides a theoretically grounded connection between adversarial robustness and uncertainty quantification in multimodal encoders.