Research map / Vision-language and multimodal / Multimodal LLMs

HalluciDoctor: Mitigating Hallucinatory Toxicity in Visual Instruction Data

CVPR 2024 · Vision-language and multimodal › Multimodal LLMs · approach: Preference tuning against hallucination · 48 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.

Detects and removes object, relation, and attribute hallucinations in LLaVA-158K via question-based cross-checking across MLLM experts (producing LLaVA+), then adds counterfactual instructions with a seesaw co-occurrence strategy (LLaVA++) to reduce hallucinations and improve MLLMs.

Problem

Machine-generated visual instruction data (e.g., LLaVA-158K generated by text-only GPT-4) contains hallucinations that transfer into MLLMs, and prior mitigation methods either add training cost, inference overhead, or only suppress outputs superficially without cleaning the data.

Approach

Decompose hallucination detection into answer-chunk extraction via scene graph parsing, question generation, and consistency cross-checking of BEM scores between description answers and answers from multiple MLLM experts; remove low-consistency chunks, then expand data with counterfactual instructions chosen by an enhancement-times-inhibition seesaw score over object co-occurrence statistics.

Key results

• LLaVA+ reduces CHAIR_obj on LLaVA-158K from 28.1 to 8.3 and CHAIR_rel from 36.0 to 20.7 in the data analysis.
• MiniGPT-4 fine-tuned on LLaVA++ reaches instance-level CHAIR_obj 5.7 and CHAIR_rel 5.4, versus 12.0 and 12.2 for w/ LLaVA.
• MME total for MiniGPT-4 rises from 1148.93 (w/ LLaVA) to 1207.18 with LLaVA+ and 1276.0 with LLaVA++ (perception 955.28 for LLaVA++ vs 859.64 for LLaVA).

Measured effects reported in the paper

Surprising findings

• Setting the consistency threshold above 0.5 causes sharp MME performance decline, as too many answer chunks are removed as hallucinations.
• Hallucination reduction in the instruction data was measured with pseudo-labels from GroundingDINO and BLIP rather than human annotation, so the reported gains depend on these proxies.

Limitations

• Relies on pseudo-labels from GroundingDINO and BLIP for the hallucination statistics, which may be inaccurate.
• Consistency threshold is sensitive; high thresholds remove too much semantic content and hurt performance.
• Fine-tuning experiments limited to 7B MiniGPT-4 and mPLUG-Owl with limited compute (RTX 3090).

Experimental setup

Datasets: LLaVA-Instruction-158K MiniGPT4-Instruction MSCOCO Visual Genome

Benchmarks: extended CHAIR MME POPE OwlEval NoCaps GQA AOK-VQA

Metrics: CHAIR_obj CHAIR_rel CHAIR_attri MME perception/cognition score accuracy F1

Baselines: Faithful Prompt LURE VIGC w/ LLaVA w/ LRV LRV-Instruction

Builds on: LLaVA CHAIR LRV-Instruction POPE

Closest papers by meaning

• Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.91)
• Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling (NeurIPS 2025 · similarity 0.91)
• Seeing Is Believing: Rich-Context Hallucination Detection for MLLMs via Backward Visual Grounding (AAAI 2026 · similarity 0.91)
• Detecting and Preventing Hallucinations in Large Vision Language Models (AAAI 2024 · similarity 0.91)
• Scalable Token-Level Hallucination Detection in Large Language Models (NeurIPS 2026 · similarity 0.91)
• Detecting and Mitigating Hallucination in Large Vision Language Models via Fine-Grained AI Feedback (AAAI 2025 · similarity 0.91)