Research map / Multi-agent systems and games / LLM agents

InEx: Hallucination Mitigation via Introspection and Cross-Modal Multi-Agent Collaboration

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies · 3 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 InEx, a training-free multi-agent framework that combines TVER-guided internal introspective decoding with iterative cross-modal verification by textual and visual self-reflection agents and an image editing agent to reduce MLLM hallucination.

Problem

Hallucination in multimodal LLMs (MLLMs) that are factually inaccurate relative to the input image, where existing methods need human supervision, retraining, or treat in-processing and post-processing mitigation separately.

Approach

Combines an entropy-based introspection signal (Text-to-Visual Entropy Ratio) that triggers visual-feature injection and logit fusion inside the decision MLLM with an external loop where a textual reflection agent checks answers against dense captions and a visual reflection agent checks consistency of an image edited from the answer with the original image, feeding feedback back to the decision a

Key results

• On POPE MSCOCO with LLaVA-1.5-7B, InEx reaches 88.73 average accuracy versus 79.83 for the base model and 84.14 for OPERA.
• On MME hallucination subset, InEx scores 673.3 for LLaVA-1.5-7B (vs 643.3 base) and 732.3 for GLM-4V-9B (vs 697.3 base).
• On LLaVA-Bench, InEx improves LLaVA-1.5-7B accuracy from 63.4 to 66.5.

Surprising findings

• Baseline ICD and VCD reduce accuracy on LLaVA-Bench and MMBench for LLaVA-1.5-7B, e.g., ICD drops MMBench by about 23 points.

Limitations

• No support for audio modality; the framework is limited to text and vision.
• Multi-agent loop with ensemble verification and image editing adds computational cost and latency.

Experimental setup

Datasets: MSCOCO A-OKVQA GQA

Benchmarks: POPE CHAIR HallusionBench VizWiz-VQA MME MMBench MM-Vet LLaVA-Bench

Metrics: accuracy F1-score CHAIR_S CHAIR_I recall MME score

Baselines: OPERA ICD VCD

Builds on: OPERA VCD ICD IC-Edit

Closest papers by meaning

• BALTO: Balanced Token-Level Policy Optimization for Hallucination Mitigation (NeurIPS 2026 · similarity 0.91)
• Multi-Agent Undercover Gaming: Hallucination Removal Through Counterfactual Test for Multimodal Reasoning (AAAI 2026 · similarity 0.90)
• Visual Multi-Agent System: Mitigating Hallucination Snowballing via Visual Flow (ICLR 2026 · similarity 0.90)
• Auditing Meta-Cognitive Hallucinations in Reasoning Large Language Models (NeurIPS 2025 · similarity 0.90)
• Token-Guard: Towards Token-Level Hallucination Control via Self-Checking Decoding (ICLR 2026 · similarity 0.90)
• Tracing the Cascade: A Topology-Aware Evaluation Framework for Scientific Agent Hallucinations (NeurIPS 2026 · similarity 0.89)