Mitigating Hallucinations in Large Vision-Language Models without Performance Degradation

🕒 Published (v1): 2026-04-22 09:02 UTC · Source: Arxiv · Venue: ACL 2026 · link

Why this paper was selected

Hallucination-free fine-tuning without performance degradation; ACL 2026

Ask a follow-up

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

💬 Ask ChatGPT✦ Ask Claude

TL;DR

MPD is a training-free, dual-stage framework that mitigates object hallucinations in LVLMs by (1) disentangling hallucinatory components from faithful semantics via orthogonal subspace projection and (2) selectively editing only the weight vectors most cosine-aligned with those hallucinatory directions. It achieves a 23.4% reduction in hallucinations while retaining 97.4% of general generative capability, with zero inference overhead.

Problem

Existing representation-intervention methods reduce hallucinations but degrade general generation quality. The paper identifies two root causes: first, naive difference-based hallucination extraction (\(X^-_\ell - X^+_\ell\)) entangles hallucinatory components with shared general semantics, inadvertently removing useful representations; second, global parameter updates across entire layers perturb hundreds of millions of weights, inducing overfitting and distributional drift.

Method

Stage 1 — Semantic-aware hallucination extraction. Contrastive query pairs are constructed using an auxiliary LLM: a hallucination-inducing prompt \(x^-_i\) and a semantically equivalent faithful prompt \(x^+_i\) over the same image. Token-level hidden states at transformer layers \(\ell \in \{L_0, \dots, L\}\) are averaged into matrices \(X^+_\ell, X^-_\ell \in \mathbb{R}^{N \times D}\). SVD on the faithful matrix gives the grounded subspace projection \(P_\ell = U_\ell U_\ell^\top\). The hallucinatory component is extracted as the orthogonal residual:

\[\tilde{X}_\ell = (I - P_\ell) X^-_\ell\]

Proposition 1 (proven in the paper) shows this estimator has lower expected squared error than the naive difference \(X^-_\ell - X^+_\ell\) in recovering the disentangled hallucinatory signal.

Stage 2 — Selective parameter editing. For each weight row \(w^{(i)}_\ell\) in layer \(\ell\), the average cosine similarity to hallucination vectors \(\tilde{x}_{\ell,j}\) is computed:

\[s_i := \frac{1}{N} \sum_{j=1}^N \frac{\cos(w^{(i)}_\ell,\, \tilde{x}_{\ell,j})}{\|w^{(i)}_\ell\| \cdot \|\tilde{x}_{\ell,j}\|}\]

The top-\(K\) most aligned weight vectors \(I^\text{hall}_\ell\) are selected. A null-space projection matrix is constructed:

\[\tilde{Q}_\ell = I - \tilde{X}^\top_\ell (\tilde{X}_\ell \tilde{X}^\top_\ell)^{-1} \tilde{X}_\ell\]

Selected weights are updated as \(w^{(i)}_\ell \leftarrow \tilde{Q}_\ell w^{(i)}_\ell\), projecting them away from the hallucination subspace while leaving unrelated weights untouched.

Key Contributions

• Orthogonal projection-based hallucination disentanglement that prevents contamination with general semantics, proven to be a strictly more accurate estimator than naive difference-based extraction
• Cosine-similarity-guided selective weight editing that reduces parameter perturbation by 37% (MiniGPT-4) to 42% (mPLUG-Owl2) compared to Nullu
• A complete dual-stage framework with no inference-time overhead — editing is a one-time offline operation
• Theoretical justification (Proposition 1) grounding the orthogonal residual as a principled hallucination estimator under a Gaussian noise decomposition model

Results

• CHAIR (MSCOCO): MPD achieves best CHAIRS across all three models — 19.40 (MiniGPT-4), 14.00 (mPLUG-Owl2), 12.80 (LLaVA-1.5-7B) — versus Nullu's 21.40 / 15.60 / 15.20; BLEU scores remain competitive or improve
• POPE: Highest F1 in all sampling settings (random/popular/adversarial) on all three models; e.g., random-setting LLaVA-1.5-7B: 88.20 vs. Nullu 87.43 vs. original 86.03
• LLaVA-Bench: LLaVA-1.5-7B accuracy improves from 5.59 → 6.39 and detailedness from 4.72 → 5.52 (GPT-4V-evaluated, scale 0–10)
• MME: Best scores on Existence and Count subsets; competitive on Position and Color
• HallusionBench: Consistent gains over Nullu on fACC, qACC, easyA, hardA, and aACC across all three models
• Overall headline: 23.4% hallucination reduction, 97.4% generative capability preserved

Limitations

• Evaluated on only three relatively small LVLMs (MiniGPT-4 V2, LLaVA-1.5-7B, mPLUG-Owl2); generalization to larger, more capable models (e.g., LLaVA-1.6, InternVL, GPT-4V-scale) is untested
• Requires contrastive paired data generated via LURE; the quality and diversity of these pairs directly affect hallucination component extraction accuracy
• Layer range \(\{16\text{–}32\}\) is chosen empirically and requires per-model tuning; ablations show significant performance degradation when only later layers (e.g., 28–32) are edited
• Hyperparameters \(K\) (top-K weight selection), \(C\) (SVD rank), and \(N\) (sample size) add tuning burden
• Theoretical analysis assumes Gaussian residuals (\(\epsilon^- \sim \mathcal{N}(0, \sigma^2_- I)\)), which may not hold in practice for all model families

Relevance to Vision-Language Models

Object hallucination is one of the central reliability failures in LVLMs, and MPD directly advances the representation-intervention line of work (building on Nullu, ROME, and subspace-steering methods) with a principled fix for its known degradation problem. The orthogonal projection formulation provides mechanistic insight into how hallucinatory semantics are encoded separately from grounded semantics in hidden-state space, which informs our understanding of cross-modal alignment failures. The method's model-agnostic nature and zero inference overhead make it broadly applicable to deployed LVLMs. For researchers tracking VLMs, this is a concrete step toward reliable multimodal generation without the cost of annotation-based fine-tuning.