Blog
Research & perspectives
ResearchAugust 2026
Meta-distillation: learning from feedback and implicit rewardMeta-distillation turns an agent's own work into plain-text lessons, injected at inference, with no weight updates. With no labels at all, the loop lifted expert-rubric scores for Opus 4.6 on legal work and for Gemini 3 Flash on live financial research.Read →
PerspectiveJuly 2026
The correction loop is the bottleneckThe path from an expert's correction to a shipped change in agent behavior (a spreadsheet of traces, a Jira ticket, a barely tested prompt edit) is the real constraint on enterprise agent deployment. Why that loop exists, why it's the bottleneck, and what replaces it.Read →
ResearchJune 2026
Meta-distillation: learning from a teacherWhen an expert reference exists, a stronger model or a human, meta-distillation contrasts a student's runs against it and distills the difference into plain-text lessons. The lessons took Mistral Large 3 from 28.2% to 45.7% on banking customer support and a Mistral legal student from 0.35 to 0.45, with no weight updates and no teacher at runtime.Read →