Research map / Vision-language and multimodal / Multimodal LLMs
OPERA: Alleviating Hallucination in Multi-Modal Large Language Models via Over-Trust Penalty and Retrospection-Allocation
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 OPERA, a training-free beam-search decoding method that penalizes attention-based "over-trust" knowledge aggregation patterns and rolls back decoding to reallocate tokens, reducing hallucination in MLLMs.
Problem
Hallucination in multi-modal large language models during long text generation, where subsequent tokens over-rely on summary tokens and neglect image tokens.
Approach
Detects knowledge aggregation (columnar) patterns in the self-attention map over a local window of generated tokens, adds a penalty to beam-search logits to suppress them, and retrospects to the summary token to reselect tokens when the pattern persists. No extra data, models, or training are needed.
Key results
- On Shikra with 512 max new tokens, OPERA reduces CHAIR_S from 50.4 (beam search) to 36.2, and from 55.8 (DoLa) to 36.2, about 35% relative improvement over DoLa.
- GPT-4V correctness on MSCOCO improves from 5.52 to 6.26 for InstructBLIP and 5.25 to 6.29 for Shikra over beam search.
- POPE average F1 is marginally higher than beam search on all four models (e.g., LLaVA-1.5 85.4 vs 84.9), and MME on LLaVA-1.5 rises to 1515.4 from 1504.3 for beam search.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add over-trust penalty and retrospection-allocation on top of beam search (OPERA) | InstructBLIP, MiniGPT-4, LLaVA-1.5, Shikra · CHAIR 512 tokens and 64 tokens | CHAIR_S and CHAIR_I: e.g. Shikra CHAIR_S 50.4 -> 36.2 (512 tokens) | Helped |
| Add over-trust penalty only (no retrospection) (ablation) | Beam search on Shikra and LLaVA-1.5 · CHAIR 512 tokens | CHAIR_S: Shikra 50.4 -> 41.8; LLaVA-1.5 48.8 -> 45.2 | Helped |
| Add retrospection-reallocation only (no penalty) (ablation) | Beam search on Shikra and LLaVA-1.5 · CHAIR 512 tokens | CHAIR_S: Shikra 50.4 -> 46.8; LLaVA-1.5 48.8 -> 46.4 | Helped |
| Add over-trust penalty and retrospection-allocation on beam search | POPE F1 across four MLLMs · POPE random/popular/adversarial average | F1: marginal gains, e.g. LLaVA-1.5 84.9 -> 85.4 | Helped |
| Add OPERA decoding on LLaVA-1.5 7B | MME and MMBench · MME perception and MMBench | MME score / MMBench accuracy: MME 1504.3 -> 1515.4; MMBench 64.4 -> 64.4 | No clear effect |
| Add OPERA decoding on LLaVA-1.5 7B | PPL and GPT-4 text quality · 1000 MSCOCO images | PPL_1 (lower better): Beam 11.11 vs OPERA 11.67; Greedy 12.72 | Mixed |
| Increase candidate count N_can from 2 to 10 (ablation) | OPERA hyperparameter sweep · CHAIR 512 tokens | CHAIR_S: InstructBLIP best at N_can=2 (43.8); Shikra best at N_can=8 (33.3) | Mixed |
| Increase rollback threshold r from 15 to 25 (ablation) | OPERA on InstructBLIP · CHAIR 512 tokens | CHAIR_S: r=15 46.4 vs r=25 41.8 for InstructBLIP; other models best at r=15 | Mixed |
Surprising findings
- The over-trust penalty contributes more to the gains than the retrospection-reallocation strategy in ablations.
- Gains are marginal on short answers such as POPE yes/no responses, since the aggregation pattern manifests mainly in long sequences.
Limitations
- Cannot address all hallucination types, e.g., those from strong language biases or weak visual perception.
- Marginal gains on short answers (under about 10 tokens) because the aggregation pattern has hysteresis.
- Hyperparameters (sigma, r, N_can) are model-dependent and tuned on evaluation settings.
Experimental setup
Datasets: MSCOCO 2014 val Visual Genome
Benchmarks: CHAIR POPE MME MMBench GPT-4 hallucination evaluation on Visual Genome
Metrics: CHAIR_S CHAIR_I F1 SPI WPI HSPI HWPI HSR HWR PPL
Baselines: Greedy decoding Nucleus sampling Beam Search DoLa
Builds on: beam search DoLa anchor token
Closest papers by meaning
- MLLM can see? Dynamic Correction Decoding for Hallucination Mitigation (ICLR 2025 · similarity 0.93)
- MoLE:Decoding by Mixture of Layer Experts Alleviates Hallucination in Large Vision-Language Models (AAAI 2025 · similarity 0.92)
- Self-Introspective Decoding: Alleviating Hallucinations for Large Vision-Language Models (ICLR 2025 · similarity 0.91)
- Grounding Language with Vision: A Conditional Mutual Information Calibrated Decoding Strategy for Reducing Hallucinations in LVLMs (NeurIPS 2025 · similarity 0.91)
- Reducing Hallucinations in Large Vision-Language Models via Latent Space Steering (ICLR 2025 · similarity 0.91)
- Self-Correcting Decoding with Generative Feedback for Mitigating Hallucinations in Large Vision-Language Models (ICLR 2025 · similarity 0.91)
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.