Architecture linked to hallucinations in vision language models
Where Hallucinations Live: A Cross-Architecture Circuit in VQ-Tokenized Vision-Language Models
Computer Vision and Pattern RecognitionMachine Learning
Summary
Some vision-language AI models mistakenly 'see' objects that aren’t really there, which is called hallucination. The authors found that this problem is linked to a specific part of the model’s early architecture, especially in models using vector quantization to understand images. By identifying and modifying this part, they reduced hallucinations more effectively than other methods. This shows that hallucinations come from how the model is built and trained, not just from how it guesses answers.
What this means in practice
- •For vision ai engineers: Improve vision-language model design by targeting early-layer circuits to reduce object hallucination during image understanding tasks.
- •For ai safety teams: Identify architectural causes of hallucination in multimodal models to develop safer AI systems with more reliable grounded outputs.
Authors
Shamanthak Hegde, Xiangrui Liu, Maitreya Patel, Yezhou Yang
Abstract
Unified vision-language models (VLMs) that tokenize images through a vector-quantized (VQ) codebook routinely hallucinate objects on grounded yes/no benchmarks, yet existing decoding-time fixes treat this as generic miscalibration without an architectural account. Using activation patching across twenty-five models spanning eight LLM families, we identify an early-layer ($L_0$) attention routing circuit shared across VQ-tokenized VLMs and propose a three-gate diagnostic that distinguishes the models carrying it from those that do not. The diagnostic isolates ten positive models (five natural unified-VQ VLMs across three LLM families and five induced variants) and rejects the remaining fifteen. A single-variable architectural swap (LLaVA-1.6 CLIP+MLP $\rightarrow$ VQ+Linear) installs the circuit, while a matched-compute MLP control on identical data does not, isolating vector quantization as the source of the pathological signal; the routing pathway that carries it is one that the backbone already provides. Against tuned VCD and DoLA baselines, tuned DoLA wins on binary calibration, but \textbf{only $L_0$ ablation reduces object hallucination in open-ended generation} (CHAIR$_i$ reduces by $31\,\%$ relatively, whereas tuned DoLA and VCD leave it unchanged or worsen it). These results recast object hallucination in unified VQ VLMs as a property of architecture and pretraining, and yield a targeted intervention that mechanism-agnostic decoding cannot replicate.