Heng Ji

LLMs in medical diagnosis are alarmingly prone to jumping to conclusions, often answering before seeing all the evidence, but strategically delaying the question and evidence presentation can boost accuracy by up to 62.6%.

LRMs can often recover from injected errors in their reasoning steps, revealing a hidden "critique" ability that can be harnessed to improve performance without additional training.

Learn a critic for coding agents from human-in-the-loop interaction traces alone, sidestepping the need for dense, verifiable rewards.

Protein language models, despite borrowing NLP architectures, distribute information across layers very differently, but a simple early-exit strategy can exploit this to boost both accuracy and efficiency.

Forget prompt engineering and fine-tuning: this "Reasoning Inception" method injects targeted reasoning into LLM agents at test time to fix conversational errors on the fly.