CDSE Paper Splits Agent Skills Into Four Components, Gains 7.6 Points
ruslansv · x · 2026-10-20
The arXiv paper "From Trajectories to Differentiated Skills: Cognitive Differentiation for Self-Improving Agents" (2610.23599) introduces CDSE, which lets LLM agent skills self-evolve without fine-tuning the base model.
Core idea: existing skill-evolution methods treat knowledge as one artifact, but real skills mix control logic, heuristics, tool guidance, and executable procedures. Inspired by cognitive science, CDSE splits a unified skill into Meta / Experience / Tool / Script components, extracts evidence-grounded semantic gradients from successful and failed trajectories, consolidates them by functional role, and applies component-specific updates with cross-component coordination.
Key numbers:
- Analysis of 500 top-ranked skill packages shows 96.8% span at least two functional roles
- Highest scores across all 36 settings (6 benchmarks, 3 agent models, 2 harnesses), beating the Base agent by 7.6 points and the strongest competing method in each setting by 2.6 points
- Maps cleanly onto SKILL.md-style progressive disclosure
More from coding & agent
- Coinbase launches perps trading for AI agents: Claude, ChatGPT and Grok can now go long US indices — MurrLincoln · 2026-10-20
- Google Cloud rethinks Kubernetes for agents: suspend/resume as core primitive — rseroter · 2026-10-20
- This MCP server is under 700 bytes: one time tool, zero dependencies, instant startup — kaphwor · 2026-10-20
- Anthropic Dev Shares the Prompts Behind His Opus 5.5 Code-Rendered Films — dfeinition · 2026-10-20
- Meta's ProPerAgent lifts memory-driven proactive assistants from 15% to 38% — dair_ai · 2026-10-20
- Claude Motion animations now send straight to Descript for full video production — descript · 2026-10-20