LIVE: The Simple Language Trick That Slashes AI Visual Hallucinations by 34 Points
A new AI paradigm lets language dynamically control vision encoders without retraining.
Traditional vision foundation models are static feature extractors, forcing downstream models to handle task adaptation. The authors of LIVE flip this paradigm by using language as high-level guidance to dynamically reshape vision embeddings during inference. This approach allows the encoder to focus on contextually relevant aspects of an image, producing controllable and generalizable representations without any task-specific fine-tuning.
Empirically, LIVE delivers striking gains: a 34-point improvement on the MMVP visual hallucination benchmark, and superior performance on visual question answering compared to rival models with far more parameters. The system also adapts to unseen instructions and tasks, hinting at a future where AI vision systems are directed purely by natural language prompts rather than costly retraining. Published at ICLR 2026, LIVE represents a fundamental shift toward adaptive, instruction-driven visual perception.
- LIVE reduces visual hallucinations by 34% on MMVP benchmark compared to static encoders.
- Outperforms vision-language models with orders of magnitude more parameters on VQA tasks.
- Zero-shot generalization to unseen instructions and tasks, eliminating task-specific retraining.
Why It Matters
LIVE enables controllable, hallucination-free vision AI that adapts to any task via language, saving massive retraining costs.