Research map / Vision-language and multimodal / Multimodal LLMs
Generate, but Verify: Reducing Hallucination in Vision-Language Models with Retrospective Resampling
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 REVERSE, a VLM that is trained with a 1.3M semi-synthetic hallucination-tagged dataset (confidence tokens) and uses retrospective resampling at inference to detect and correct hallucinations during decoding.
Problem
Vision-language models generate visual hallucinations (nonexistent objects, attributes, or concepts), and existing mitigation either adjusts decoding without correction or relies on multi-model post-hoc verification that tends to reject rather than refine outputs.
Approach
A single VLM is fine-tuned with special confidence tokens (<SPAN>, </CN>, </UN>) and a masked loss so it self-tags phrases as grounded or ungrounded; at inference, when P(</UN>) exceeds a threshold it backtracks to the last confident span and resamples (rejection sampling with rising temperature, plus query rewriting with hints) to correct the phrase.
Key results
- On CHAIR-MSCOCO with LLaVA-v1.5 7B, REVERSE (tau=0.0003) reduces CHAIR_i from 15.4 to 6.1 and CHAIR_s from 50.0 to 13.6 (tau=0.003 gives 10.3 / 37.0).
- On HaloQuest with LLaVA-v1.5 7B, REVERSE (tau=0.0003) raises average accuracy from 22.6 to 32.3; on Qwen2.5-VL 3B, tau=0.01 raises it from 33.5 to 45.1.
- On MMHal-Bench with LLaVA-v1.5 7B, REVERSE (tau=0.0003) improves score from 2.11 to 3.28 and lowers hallucination rate from 0.54 to 0.30.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use REVERSE with τ=0.003 instead of the LLaVA-v1.5 7B baseline | LLaVA-v1.5 7B · CHAIR-MSCOCO | CHAIR_i: 15.4 -> 10.3 | Helped |
| Use REVERSE with τ=0.003 instead of the LLaVA-MORE 8B baseline | LLaVA-MORE 8B · CHAIR-MSCOCO | CHAIR_i: 14.4 -> 12.2 | Helped |
| Use REVERSE with τ=0.01 instead of the Qwen2.5-VL FT 3B baseline | Qwen2.5-VL FT 3B · HaloQuest | Avg. Acc.: 33.5 -> 45.1 | Helped |
| Add hallucination-aware training (ablation) | LLaVA-v1.5-7B · AMBER-G | CHAIR: 7.8 -> 7.2 | Helped |
| Add hallucination-aware training (ablation) | LLaVA-v1.5-7B · AMBER-G | Cover: 51.0 -> 53.2 | Helped |
| Add rejection sampling after hallucination-aware training (ablation) | LLaVA-v1.5-7B · AMBER-G | CHAIR: 7.2 -> 6.0 | Helped |
| Add rejection sampling after hallucination-aware training (ablation) | LLaVA-v1.5-7B · AMBER-G | Cover: 53.2 -> 51.0 | Hurt |
| Add query rewriting after rejection sampling (ablation) | LLaVA-v1.5-7B · AMBER-G | Cover: 51.0 -> 52.2 | Helped |
| Increase maximum self-correction rounds from 0 to 50 (ablation) | REVERSE-v1.5-7B · AMBER-G test set | CHAIR: 7.8 -> 6.0 | Helped |
| Lower τ from 0.003 to 0.0003 (ablation) | REVERSE on LLaVA-v1.5 7B · HaloQuest visually challenging subset | VC Acc.: 31.5 -> 18.7 | Hurt |
| Lower τ from 0.003 to 0.0003 (ablation) | REVERSE on LLaVA-v1.5 7B · HaloQuest insufficient context subset | IC Acc.: 26.9 -> 58.8 | Helped |
| Apply REVERSE to discriminative VQA tasks | LLaVA-v1.5 7B · POPE | All accuracy: 85.9 -> 85.9 | No clear effect |
Surprising findings
- Hallucination-aware training alone already improves coverage and reduces hallucination, which the authors attribute to a DPO-like contrast between confident and unconfident phrases.
- On general and discriminative benchmarks (AMBER-D, POPE, MME-Hall, HallusionBench), REVERSE gives no substantial gains and sometimes drops (e.g., HallusionBench 46.94 to 45.44, MM-Vet 31.10 to 28.40).
Limitations
- Synthetic dataset relies on GPT-4o-mini negatives and lacks coverage of edge cases like false premises and insufficient context, with possible label noise and bias from MS-COCO.
- Retrospective resampling increases generated tokens (about 1.75x to 3.05x) and runtime up to roughly 3x baseline.
- Thresholds are set per model and no gains on binary discriminative tasks; visually challenging question performance drops at lower thresholds.
Experimental setup
Datasets: LLaVA-v1.5-665k MS-COCO OpenImages
Benchmarks: CHAIR-MSCOCO AMBER-G AMBER-D POPE MME-Hall MMHal-Bench HaloQuest HallusionBench GQA MM-Vet
Metrics: CHAIR_i CHAIR_s Cover Hall Cog accuracy F1 score hallucination rate CLAIR
Baselines: VCD OPERA DoLA AGLA MEMVR EOS HALVA HA-DPO Woodpecker LLaVA-RLHF HACL
Builds on: LLaVA-v1.5 LoRA Woodpecker DPO
Closest papers by meaning
- Self-Correcting Decoding with Generative Feedback for Mitigating Hallucinations in Large Vision-Language Models (ICLR 2025 · similarity 0.95)
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · similarity 0.93)
- Causally-Grounded Dual-Path Attention Intervention for Object Hallucination Mitigation in LVLMs (AAAI 2026 · similarity 0.93)
- Prefill-Time Intervention for Mitigating Hallucination in Large Vision-Language Models (CVPR 2026 · similarity 0.93)
- Hallucination-aware Intermediate Representation Editing in Large Vision-Lanugage Models (ICLR 2026 · similarity 0.93)
- Finding the Correct Visual Evidence Without Forgetting: Mitigating Hallucination in LVLMs via Inter-Layer Visual Attention Discrepancy (ICML 2026 · similarity 0.93)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.