Zongchao Chen

Autonomous evolution of agent harnesses can yield significant performance improvements, yet struggles in specific task environments reveal critical limitations.

Turns out, telling LLMs *not* to use the answer when generating reverse chain-of-thought reasoning can actually make them *more* reliant on it—but a skeleton-guided approach breaks the cycle.

A 3B parameter model now rivals models 10x its size in reasoning, alignment, and agentic tasks, challenging the assumption that bigger is always better.

Achieve up to 55.7% token reduction and 4.1% accuracy improvement in language reasoning by selectively compressing reasoning traces, proving that less can be more.