Research map / Vision-language and multimodal / Multimodal LLMs

Analyzing and Mitigating Object Hallucination: A Training Bias Perspective

AAAI 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.

Introduces POPEv2, a benchmark of counterfactual (object-masked) MSCOCO training images showing LVLMs hallucinate on seen images, traces the bias to the LM head via probing, and proposes Obliviate, which unlearns hallucinated sub-sentences by fine-tuning only the LM head.

Problem

LVLMs still hallucinate objects, even on images seen during training; it is unclear whether they use their training data to acquire generalizable visual understanding or learn training biases.

Approach

Treat the discrepancy between model outputs and ground-truth training labels as a proxy for training bias, and unlearn the hallucinated sub-sentences via gradient ascent combined with an autoregressive loss, updating only the LM head (about 2% of parameters).

Key results

• On POPEv2, Obliviate raises LLaVA-1.5-7B TNR from 49.00 to 84.20 and F1 from 77.44 to 82.91 while updating about 2% of parameters.
• Obliviate lowers Object HalBench mention hallucination for LLaVA-1.5-7B from 25.1 to 18.3, outperforming VCD, VDD, OPERA, HA-DPO and POVID.
• Full-parameter SFT and LoRA degrade LLaVA-NeXT-7B POPEv2 F1 (80.44 and 77.59) versus 81.91 base, while Obliviate reaches 83.38 with 28 GB VRAM and 10 GPU hours.

Surprising findings

• Full-parameter SFT and LoRA with the unlearning objective degrade performance, which the authors attribute to disrupting generation patterns across all layers.
• Scaling model size does not reliably help: Qwen2-VL-72B (82.12 F1) underperforms Qwen2-VL-2B (91.04 F1) on POPEv2.

Limitations

• Relies on a fixed hallucination-detection pipeline and object extraction from captions, which may miss hallucinations or mislabel objects.
• POPEv2 is small (500 original and 500 counterfactual images), and the gains come from a single object-masking construction.
• The unlearning strength alpha is sensitive; at 0.05 the model shows repetitive outputs and degraded generation.

Experimental setup

Datasets: MSCOCO 2017 train LLaVA-150K POPE MSCOCO 2017 val

Benchmarks: POPEv2 POPE Object HalBench MMHal-Bench LLaVA-Bench MME perception ROPE

Metrics: F1 score TNR PBO accuracy precision recall response hallucination rate mention hallucination rate informativeness CHAIR

Baselines: VCD VDD OPERA HA-DPO POVID SFT LoRA

Builds on: POPE machine unlearning DPO VCD OPERA

Closest papers by meaning

• Multi-Object Hallucination in Vision Language Models (NeurIPS 2024 · similarity 0.95)
• Analyzing and Mitigating Object Hallucination in Large Vision-Language Models (ICLR 2024 · similarity 0.94)
• Devils in Middle Layers of Large Vision-Language Models: Interpreting, Detecting and Mitigating Object Hallucinations via Attention Lens (CVPR 2025 · similarity 0.94)
• Same Attention, Different Truths: Put Logit-Lens over Visual Attention to Detect and Mitigate LVLM Object Hallucination (CVPR 2026 · similarity 0.94)
• Causal-HalBench: Uncovering LVLMs Object Hallucinations Through Causal Intervention (AAAI 2026 · similarity 0.93)
• One Token, Two Fates: A Unified Framework via Vision Token Manipulation Against MLLMs Hallucination (CVPR 2026 · similarity 0.93)