# Learning Loop Research Basis Use this reference only when reviewing or changing Learning Capture, Loop Discovery, intervention-ledger, or Skill-evolution behavior. It records the external research basis behind those contracts. It is not an evidence source for scoring a project and never substitutes for opened local instructions, code, tests, delivery state, or reviewed Task Episodes. This file owns the primary-source rationale and the known stability conditions for evolving a durable asset from experience. It does not own pattern signatures, recurrence thresholds, or coverage codes; use [Learning Loop Detection Patterns](learning-loop-patterns.md). It does not own owner selection; use [Loop Discovery](loop-discovery.md). ## Load When - A Learning Capture, Loop Discovery, or intervention-ledger contract change needs a primary-source rationale. - A proposal wants automatic Skill generation, asset merging, or unattended self-revision, and the review needs the published stability conditions. - A reviewer must keep `opportunity` and `readiness` claims separate from an `effectiveness` claim and needs the comparison designs that support each. ## Improvement axes [Self-Improvements in Modern Agentic Systems](#references) models an agent as a foundation model coupled to an operational scaffold of prompts, memory, tools, and control logic, and formalizes self-improvement as a self-induced update operator that commits changes to one of those targets. The two axes are orthogonal: | Axis | Update target | Repository position | | --- | --- | --- | | Foundation model improvement | Parameters, driven by self-generated demonstrations, self-evaluation, or environment exploration | Out of scope. This repository never updates model weights. | | Scaffolding improvement | Prompts, memory, tools, workflows, or the full scaffold | In scope. Every durable owner (Memory, Rule, Skill, Hook, Gate, Workflow, Agent, Eval) is a scaffold component. | Keep the boundary explicit in reader-facing output. Repository-side evolution changes reviewable scaffold artifacts; it does not change model capability, and a scaffold gain must never be reported as a capability gain. ## Distillation shape Skill-artifact work converges on the same progression, which is why the repository's stage vocabulary (`capture`, `generalize`, `codify`, `route`, `exercise`, `evaluate`, `maintain`) can carry these findings without a new model. | Stage | Research position | Repository owner | | --- | --- | --- | | capture | Raw trajectory stores are redundant and noise-heavy; distill instead. Successful runs become strategic patterns, failed runs become concise lessons (SkillRL). | Normalized Task Episodes and learning signals. | | generalize | Recurring lessons are induced across many traces in parallel, not from one trace (Trace2Skill). | Two comparable episodes as the default recurrence threshold. | | codify | Lessons consolidate hierarchically into one unified, conflict-free skill directory (Trace2Skill); failures materialize into structured skill folders (EvoSkill). | `recommendedOwner` selection. Consolidation is reported, not performed. | | route | Retrieval separates general heuristics from task-specific ones (SkillRL); guidance splits into task-level and step-level granularity (D2Skill). | Smallest-durable-owner selection in Loop Discovery. | | exercise | An artifact only counts once the agent actually runs under it. | `routed-but-not-applied` and `asset-updated-not-reexercised`. | | evaluate | A held-out or paired comparison decides whether the artifact is kept. | Intervention ledger. | | maintain | Prune by measured utility, not by age (D2Skill). | `stale-or-conflicting-asset` freshness review. | ## Stability conditions Every published result that reliably improves over its starting point adds an acceptance gate. Loosely controlled self-revision is reported as the failure mode, not the baseline. Treat the following as review questions for any evolution proposal. | Condition | Source | Current repository state | | --- | --- | --- | | Accept an edit only when a held-out score strictly improves | SkillOpt | Absent. `stopOrRevertCondition` records a revert trigger; there is no held-out acceptance gate. | | Retain only artifacts that stay on a validation Pareto frontier | EvoSkill | Absent. | | Keep the verifier isolated from ground-truth test content while it still returns actionable feedback | CoEvoSkills | Partial. Read-only specialist passes are separated from Lead reconciliation, but the verifier does not co-evolve. | | Compare paired baseline and artifact-injected runs under one policy, and use the gap as the update signal | D2Skill | Absent. `comparisonWindow.taskMix` compares observation windows, not controlled pairs. | | Bound each edit and buffer rejected edits so the optimizer stops re-proposing them | SkillOpt | Absent. Related patterns (`unvalidated-intervention`, `correction-not-promoted`) detect the symptom only. | | Assign credit below episode granularity before blaming a step | GiGPO | Absent. Findings attribute to a primary check at Task Episode level. | | Generate held-out tasks rather than assuming a fixed suite exists | AgentEvolver | Absent. The `eval` owner may be recommended but is not produced. | The paired-comparison gap explains a current honest limit: with only window comparison available, `improving` and `unchanged` ledger results project `later-validation` to `Exercised`, not `Outcome-supported`. Closing that gap requires a controlled comparison, not a scoring change. ## Transfer Evolved skill artifacts retain value across model scales, across model families, across execution harnesses, and on out-of-distribution tasks, with the underlying model frozen and without test-time retrieval. Human-authored skills can also underperform because of human-machine cognitive misalignment rather than missing content. Both findings support treating a repository-local artifact as a durable asset rather than model-specific tuning, and they justify reviewing a human-written asset against observed agent behavior instead of assuming authorship implies fitness. Neither finding licenses a claim about a specific asset in a specific repository. ## Interpretation boundary These sources explain the shape of the contracts. They do not freeze terminology, provider features, thresholds, or numeric scores. - Do not cite a published gain as evidence about this repository. Reported numbers come from bounded benchmark suites under a fixed harness. - Do not raise `later-validation` above `Exercised` without a comparable held-out or paired result recorded in the intervention ledger. - Do not auto-merge, auto-rewrite, or auto-delete a named durable asset from a single observation window. Consolidation findings stay advisory until an acceptance gate exists. - Do not convert an absent eval suite into a low score. Without a comparison set the honest state is `Unobserved`. - Weight-space methods stay out of scope even when cited here; they are recorded to mark the boundary, not to widen it. ## References Scaffolding improvement, skill artifacts: - Ni et al., *Trace2Skill: Distill Trajectory-Local Lessons into Transferable Agent Skills*, 2026. - Alzubi et al., *EvoSkill: Automated Skill Discovery for Multi-Agent Systems*, 2026. - Yang et al., *SkillOpt: Executive Strategy for Self-Evolving Agent Skills*, 2026. - Zhang et al., *CoEvoSkills: Self-Evolving Agent Skills via Co-Evolutionary Verification*, 2026. Skill banks coupled to policy training: - Xia et al., *SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning*, 2026. - Tu et al., *Dynamic Dual-Granularity Skill Bank for Agentic RL*, 2026. Credit assignment, task generation, and supervision, recorded as the out-of-scope boundary: - Feng et al., *Group-in-Group Policy Optimization for LLM Agent Training*, NeurIPS 2025. - Zhai et al., *AgentEvolver: Towards Efficient Self-Evolving Agent System*, 2025. - Zhao et al., *Self-Distilled Reasoner: On-Policy Self-Distillation for Large Language Models*, 2026. Umbrella framing: - Ren et al., *Self-Improvements in Modern Agentic Systems: A Survey*, 2026.