Research map / Vision-language and multimodal / Multimodal LLMs

When Semantics Mislead Vision: Mitigating Large Multimodal Models Hallucinations in Scene Text Spotting and Understanding

NeurIPS 2025 · Vision-language and multimodal › Multimodal LLMs · approach: Preference tuning against hallucination

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.

Identifies semantic hallucination in LMM scene text spotting and proposes a training-free method (ZoomText plus Grounded Layer Correction) that fuses hidden states from the most text-grounded LLM layer, plus a 1,740-sample TextHalu-Bench.

Problem

LMMs often give semantically plausible but visually incorrect answers when scene text is non-semantic or visually ambiguous, relying on language priors rather than visual grounding.

Approach

Layers in the LLM that allocate more attention to ground-truth text regions hallucinate less; ZoomText localizes text regions coarse-to-fine from query-to-image attention and cross-layer attention shift without external detectors, and Grounded Layer Correction fuses hidden states from the most grounded layer into the final-layer representation at decoding.

Key results

• Mini-Monkey + Ours improves TextHalu-Bench from 46.5 to 50.6 (+4.1) and ST-VQA from 66.7 to 70.6.
• Qwen2.5-VL + Ours improves TextHalu-Bench from 48.3 to 53.8 (+5.5) and SEED-Bench Text from 66.7 to 70.2.
• On TextHalu-Bench, GPT-4o reaches 45.3 F1 versus 96.8 for humans, while LLaVA-NeXT gains only +0.6 from the method.

Measured effects reported in the paper

Surprising findings

• Naive replacement of final-layer hidden states with grounded-layer states performs poorly across benchmarks, while fusion with weight 0.1 works best.
• Selective replacement and higher fusion weights improve some spotting tasks but substantially degrade general scene text understanding.

Limitations

• Adds inference overhead during prefilling for attention map computation and token selection.
• Effectiveness depends on the base model's OCR ability; gains are marginal for LLaVA-NeXT.

Experimental setup

Datasets: ICDAR 2013 ICDAR 2015 ICDAR 2019 CCPD MSRA-TD500 RoadText MPSC CTW1500 TotalText ST-VQA TextVQA OCRBench

Benchmarks: TextHalu-Bench ST-VQA TextVQA GOT OCRVQA SEED-Bench AI2D RealWorldQA MathVista POPE MME MME-VideoOCR

Metrics: F1 accuracy character-level F1

Baselines: Mini-Monkey Qwen2.5-VL LLaVA-NeXT GPT-4o Gemini1.5-Pro InternVL2.5 Ovis1.6 LLaVA-OV Adversarial training Chain-of-Thought prompting

Builds on: LLaVA Mini-Monkey Qwen2.5-VL

Closest papers by meaning

• Seeing Far and Clearly: Mitigating Hallucinations in MLLMs with Attention Causal Decoding (CVPR 2025 · similarity 0.92)
• Hallucination at a Glance: Controlled Visual Edits and Fine-Grained Multimodal Learning (NeurIPS 2025 · similarity 0.92)
• Causally-Grounded Dual-Path Attention Intervention for Object Hallucination Mitigation in LVLMs (AAAI 2026 · similarity 0.92)
• ConVis: Contrastive Decoding with Hallucination Visualization for Mitigating Hallucinations in Multimodal Large Language Models (AAAI 2025 · similarity 0.92)
• Analyzing and Mitigating Object Hallucination: A Training Bias Perspective (AAAI 2026 · similarity 0.92)
• Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens (CVPR 2025 · similarity 0.91)