Youwang Deng

Autopilot makes silent fabricated success structurally impossible, achieving a staggering reduction in fabrication rates from 33.7% to just 0.67% on challenging tasks.

Memory in LLMs is not a one-size-fits-all; trade-offs between behavioral consistency and factual accuracy reveal a hidden alignment tax that could redefine personalization strategies.

Bigger embedding models don't always win in agent memory retrieval: MiniLM-384 beats BGE-large on lexical queries, suggesting that encoder capacity isn't the only bottleneck.