Skip to content

Discovering Compositional Hallucinations in LVLMs

🕒 Published (v1): 2025-01-01 · Source: NeurIPS · Venue: NeurIPS 2025 · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

This paper identifies Simple Compositional Hallucination (SCHall): LVLMs that answer both a visual-centric and a text-centric question correctly in isolation still hallucinate when those questions are composed into a single query. The authors build SCBench to measure this phenomenon and propose VLR-distillation—a method combining novel Vision Language Register tokens with KL-divergence distillation from a text-only branch—to mitigate it.

Problem

Existing LVLM hallucination research targets isolated failures: either visual recognition errors (POPE, CHAIR) or textual factuality errors (TruthfulQA). A critical gap remains: when neither type of error occurs individually, composing the two sub-questions into one query causes unexpected hallucination. No benchmark or mitigation method addresses this failure mode, and it grows more prominent as model visual recognition improves (LLaVA: 19.1% → InternVL: 27.5% of residual errors).

Method

SCBench construction: Triplets of (simple visual question, simple textual question, compositional question) are built semi-automatically using GPT-3.5 over images from MMBench/MME/SEEDBench. Only compositional questions that most LVLMs answer incorrectly (but both sub-questions correctly) are retained after manual filtering.

Root-cause analysis: Two contributing factors are identified: 1. Visual abstraction failure: Masking irrelevant image regions (+7.6 pp for LLaVA) and text insertion both improve compositional accuracy, indicating the model cannot selectively target relevant visual content under implicit queries. 2. Language processing degradation: Attaching even a blank image to text-only ScienceQA questions drops accuracy by 2.1–5.78%; logit-lens analysis shows answer logits in compositional settings are roughly half those in text-only settings, with correct answers appearing later in the output sequence.

VLR-distillation (two training stages): - Stage 1 – VLR pretraining: N=4 randomly initialized Vision Language Register tokens are inserted between image and text inputs. A modified causal mask forces output tokens to attend to image tokens only through the VLRs (not directly), training VLRs as selective visual-to-linguistic bridges. - Stage 2 – Distillation learning: A text-only teacher branch replaces image tokens with caption tokens and computes output logits. The student (multimodal branch with VLRs) is trained with L = L_reg + L'_reg + L_KL, where L_KL is KL divergence between teacher and student distributions. Language model weights are tuned via LoRA; VLRs are frozen. This preserves the LM's text-processing capability under visual input.

Key Contributions

  • Definition and characterization of Simple Compositional Hallucination (SCHall) as a distinct, pervasive failure mode in LVLMs.
  • SCBench: A curated benchmark of visual/textual/compositional triplets spanning perception, science, commonsense, factual, language, scene, and math categories.
  • Empirical evidence that SCHall prevalence increases with model capability (stronger visual encoders shift errors toward composition).
  • VLR-distillation: First targeted baseline for SCHall, combining selective visual register tokens with text-teacher KL distillation.

Results

  • SCBench overall accuracy: VLR-distillation improves LLaVA1.5-7B by +8.05 pp (33.75→41.80), Qwen-VL-Chat by +4.65 pp (42.41→47.06), MiniGPT-4 by +12.07 pp (14.86→26.93); all existing methods (VCD, PAI, CODE, REVERIE, CCA) show negligible or negative gains on SCBench.
  • POPE (LLaVA1.5): Accuracy +4.17/+3.40/+2.22 pp on random/popular/adversarial splits; MiniGPT-4 gains +9.14/+8.60/+7.94 pp; recall improves by ~+9 pp across settings.
  • MME-hall (LLaVA1.5): Total score improves from 510.33→585.67 (+75 points).
  • General benchmarks (LLaVA1.5): MMBench +1.1 (64.3→65.4), ScienceQA +1.0 (66.8→67.8), HallusionBench +1.9 (47.6→49.5), MM-Vet +2.2 (31.1→33.3).
  • Ablation: VLRs alone yield larger gains than distillation alone; combined achieves best results across all POPE splits.

Limitations

  • SCBench is semi-automatically constructed and limited to question types derived from existing benchmarks; coverage of open-ended or multi-hop compositions is absent.
  • Root-cause analysis is preliminary: image masking with manual annotations yields only modest +8 pp improvement, suggesting additional unexplored factors.
  • VLR count (N=4) chosen heuristically; gains plateau quickly (N=2,4,8,16 perform similarly).
  • Broader ethical limitations: method inherits biases from underlying LVLMs (LLaVA, GPT-4o) used in benchmark construction.
  • Experiments limited to LLaVA1.5, Qwen-VL-Chat, and MiniGPT-4; no evaluation on large frontier models beyond GPT-4o in the preliminary analysis.

Relevance to Vision-Language Models

SCHall exposes a failure mode that becomes more severe as VLMs improve at individual tasks—making it especially relevant as the field moves toward stronger, more capable models. The finding that visual inputs degrade language processing even with uninformative images has direct implications for multimodal architecture design and training. VLR-distillation offers a lightweight LoRA-compatible plug-in for existing LVLMs that generalizes beyond the SCBench setting to standard hallucination and VQA benchmarks. This work complements compositional reasoning benchmarks (GQA, CLEVR) but uniquely focuses on single-step compositions that are individually trivial, pinpointing a gap in current evaluation and training pipelines.