Tag
A new Meta paper finds that agent memory systems work better by applying a stricter confidence cutoff only to user value and belief claims, rather than using a single global threshold. A targeted bar on values reduced unsupported saved facts from 6.2% to 4.0% while retaining about 13 points more of the source's key facts.
研究提出按语义类别条件化的记忆保留置信阈值，而非单一全局阈值，用于智能体长期记忆的写入边界控制，在Meta部署的冷启动记忆管线中将不受支持的保留率从6.2%降至4.0%。
Jevstiller is a tool that distills a larger AI model into a local model with a configurable disagreement bound, enabling faster inference while ensuring agreement with the original model on a specified percentage of requests.
This paper introduces JEV-as-a-Judge, a cost-effective evaluation method for LLMs that uses a decision-only judge with confidence thresholds to accept certain verdicts and escalate uncertain ones, achieving comparable accuracy to state-of-the-art models at significantly lower cost.
A SpaceX AI insider shares five calibration tips for improving Grok Bot agent performance, including confidence thresholds, task breakdown, negative examples, action gating, and edge case testing.