Zhiheng Liu

Terminal-use agents are still far from achieving reliable general-purpose performance, with top models only scoring 65.8% on a new benchmark that spans diverse real-world tasks.

Ditching the vision encoder actually *improves* multimodal understanding at scale, proving that pixel embeddings alone can achieve state-of-the-art results in unified multimodal models.

LALMs struggle more with *hearing* the evidence than *reasoning* about it, and EvA's evidence-first fusion architecture proves it.