Research map / Vision-language and multimodal / Multimodal LLMs

COPO: Causal-Oriented Policy Optimization for Hallucinations of MLLMs

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 COPO, which adds a token-level causal completeness reward (sufficiency + necessity) into the GRPO advantage to reduce spurious background correlations and hallucinations in MLLMs.

Problem

MLLMs trained with outcome-only GRPO rewards attend to task-irrelevant background regions, creating spurious background-answer correlations that cause hallucinations.

Approach

Builds a structural causal model for MLLM generation and defines a causal completeness reward combining a sufficiency score (gain in answer correctness when a reasoning token is kept, averaged over sampled continuations) and a necessity score (drop in correctness when the token is masked); this reward is added to the GRPO advantage for reasoning tokens.

Key results

• On CHAIR, COPO reaches CHAIR_S 35.9 on InstructBLIP vs 37.6 for the best prior method (CSR), and CHAIR_S 18.8 on Qwen-VL vs 19.9 (GCPO).
• POPE F1 improves to 88.0 on LLaVA-1.5 and 88.3 on Qwen-VL, each about 1 to 1.5 points above the best prior method.
• The GPT-4o assisted evaluation gives an accuracy score of 8.71 vs 7.42 for DeCo.

Surprising findings

• Gradient saliency of the MLLM over background regions is higher than the text-only LLM regardless of whether the answer is correct.

Limitations

• Evaluated only on models up to 7B-scale; not tested at 72B or above 100B.
• Causal reward requires multiple sampled continuations per token, adding cost (H samples and masked generations per token).

Experimental setup

Datasets: MSCOCO V* dataset ArxivQA ThinkLite-VL

Benchmarks: CHAIR POPE MME GPT-4 assisted caption evaluation

Metrics: CHAIR_S CHAIR_I POPE F1 MME score GPT-4o accuracy/correctness/detailedness

Baselines: Beam Search Nucleus DoLa OPERA VCD DeCo VTI HA-DPO CLIP (OHD) POVID CSR GCPO

Builds on: GRPO DPO OPERA VCD DoLa

Closest papers by meaning

• Mitigating Hallucinations in Large Language Models via Causal Reasoning (AAAI 2026 · similarity 0.90)
• BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation (NeurIPS 2026 · similarity 0.90)
• Ground What You See: Hallucination-Resistant MLLMs via Caption Feedback, Diversity-Aware Sampling, and Conflict Regularization (AAAI 2026 · similarity 0.88)
• IRIS: Implicit Reward-Guided Internal Sifting for Mitigating Multimodal Hallucination (ICML 2026 · similarity 0.88)
• Mitigating Modality Prior-Induced Hallucinations in Multimodal Large Language Models via Deciphering Attention Causality (ICLR 2025 · similarity 0.88)
• Mitigating Hallucinations in Large Vision-Language Models via DPO: On-Policy Data Hold the Key (CVPR 2025 · similarity 0.88)