Haoyue Liu

Sampling rollouts via GRPO actively degrades tool-selection training as policies concentrate, whereas exact policy optimization over enumerable combinatorial tool spaces eliminates vanishing gradients and boosts accuracy by up to 14.2 points.

Adaptive similarity margins in HN-CLIP boost retrieval accuracy by up to 4.3% while training 2.4x faster than leading methods.

Incorporating visual feedback into prompt optimization leads to significant performance improvements in vision-language tasks, revealing previously unrecognized error patterns.

TARA redefines prompt optimization by routing specific failures to tailored repair operators, achieving superior semantic accuracy without the need for generator retraining.

You can slash LLM prompt evaluation costs by 35-60% without sacrificing accuracy by intelligently selecting which examples to use.

Decomposing prompts into independently optimizable "factors" lets you zero in on failure points and slash prompt optimization costs by up to 87%.