EverMind Releases Three Papers on Full-Stack Self-Evolving AI Architecture

量子位 · wechat · 2026-08-08

Incubated by Shanda Group, the EverMind team recently published three papers systematically presenting a full-stack technical solution for Self-Evolving AI, spanning the skill layer, harness layer, and model weight layer.

The team believes that combining long-term memory with continuous learning to achieve self-evolution is the core technical route for next-gen AI. The three papers tackle three major hurdles of self-evolution:

Harness Layer (HarnessBank): A framework enabling Agents to autonomously diagnose failures and rewrite their runtime logic. By separating "proposing changes" from "attributing changes" and introducing a Harness Gene Bank, it fundamentally solves the overfitting problem in self-improvement. Experiments show consistent performance gains of 5.1% to 15.4% across 7 benchmarks, even with frozen model weights.
Skill Layer (SkillCorpus): To address the fragmentation of open-source skill files, the team built a curated library of 100,000 high-quality skills. Paired with a proprietary retrieval and selection stack, it delivers stable capability enhancements for Agents in benchmarks.
Weight Layer (DASH): Targeting the distortion of supervision signals in traditional On-Policy Self-Distillation (OPSD) during long-sequence reasoning, the DASH method uses adaptive propagation gates. It dynamically adjusts the propagation range of supervision signals based on local divergence, significantly improving the model's learning efficiency.

Original post →

EverMind Releases Three Papers on Full-Stack Self-Evolving AI Architecture

More from coding & agent