Research map / Vision-language and multimodal / Multimodal LLMs

HulluEdit: Single-Pass Evidence-Consistent Subspace Editing for Mitigating Hallucinations in Large Vision-Language Models

CVPR 2026 · Vision-language and multimodal › Multimodal LLMs · approach: Decoding-time hallucination mitigation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes HulluEdit, a single-pass, reference-free hidden-state subspace editing method that suppresses object hallucinations in LVLMs by projecting onto an online weighted-SVD visual subspace and an orthogonal anti-prior subspace.

Problem

Object hallucination in large vision-language models, where language priors override visual evidence; existing methods need reference models or multiple passes, or use static subspace edits that suppress genuine visual evidence.

Approach

Decomposes each hidden state at an edit layer into orthogonal visual-evidence, anti-prior (built from a non-visual text cache within the visual complement) and residual components, and applies closed-form, certificate-gated shrinkage to the non-visual parts, computed online without extra forward passes.

Key results

• On LLaVA-1.5-7B MSCOCO captioning, HulluEdit reports CHAIR_i 4.18 and CHAIR_s 13.00 versus Nullu's 5.30 and 15.20 and greedy's 7.08 and 20.40.
• On POPE Adversarial with LLaVA-1.5-7B, accuracy rises to 82.5 from 77.6 for greedy decoding.
• On MME, Existence improves from 181.67 to 195.00 over LLaVA-1.5, while Count drops from 118.33 to 105.00.

Surprising findings

• MME Count performance declines (118.33 to 105.00) under HulluEdit, which the authors attribute to numeric information in the residual subspace.
• DeCo performance is inconsistent on LLaVA-1.5-13B POPE (accuracy 82.2 vs greedy 87.6 on Random), showing large drops for some baselines.

Limitations

• Theoretical guarantees are weak: the claimed monotone VCR/PCR improvement and Lipschitz bound follow from the paper's own closed form but rely on assumptions and some inconsistencies (e.g., notation HalluEdit vs HulluEdit, U^T R=0 claim not guaranteed by construction).
• Gains on MME Count and CHAIR_s on MiniGPT-4 are mixed; reported results use only a few backbones and hyperparameters tuned on held-out sets.
• Several reported results (AMBER, Hallu-Bench, MMVet, Qwen2.5-VL) lack full methodological detail and comparisons with the main baselines.

Experimental setup

Datasets: MSCOCO

Benchmarks: POPE CHAIR MME AMBER Hallu-Bench MMVet

Metrics: Accuracy F1 CHAIR_i CHAIR_s BLEU throughput (tokens/s)

Baselines: Greedy VCD ICD VAF DeCo DoLa OPERA Nullu Woodpecker LURE HALC Beam Search

Builds on: Nullu DeCo VCD DoLa

Closest papers by meaning

• Hallucination-aware Intermediate Representation Editing in Large Vision-Lanugage Models (ICLR 2026 · similarity 0.93)
• Mitigating Object Hallucination in Large Vision-Language Models via False Discovery Controlled Visual Data Splitting (NeurIPS 2026 · similarity 0.93)
• Mitigating Object Hallucinations in Large Vision-Language Models through Visual Contrastive Decoding (CVPR 2024 · similarity 0.93)
• Mitigating Object Hallucination in Large Vision-Language Models via Image-Grounded Guidance (ICML 2025 · similarity 0.93)
• Interpreting and Editing Vision-Language Representations to Mitigate Hallucinations (ICLR 2025 · similarity 0.93)
• Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · similarity 0.92)