Agent Self-Improvement with Frozen Weights: The Added Value and Limits of Failure Traces
cs.LG, cs.AI, cs.CL
Submitted: 2026-11-08
Updated: 2026-11-11
Code: https://github.com/hopit-ai/harness-mistakes-artifacts
Terminology
Sources
- Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models
- SkillOpt: Executive Strategy for Self-Evolving Agent Skills
- GEPA: Reflective Prompt Evolution Can Outperform Reinforcement Learning
- Harness 2: Recursive Agent Harnessing for an Open World
- Self-Evolving Agents Can Repair Failures but Struggle to Generalize from Them
- VibeMemBench: Evaluating Memory Systems for Coding Agents on Real Repository Coding Tasks
- Beyond Endpoint Performance: Process-Level Evaluation of Self-Evolving Agents
- Skill infinity: Recursive Self-Improvement for Agent Skill Optimization
- Beneath the Surface: What We Overlook When We Talk About Agent Evolution
- Can Coding Agents Sustain Continual Evolution without Forgetting?
- SelfImprove-Bench: Evaluating Generalization in Self-Improving Agents
- The Memory Trap: Language Agents Overgeneralize from Benign Interactions
- EvoGroup: A Self-Evolving Harness for Group Agents in Shared Contexts
- Learning the Task or Learning the Verifier? Rethinking Criteria Exposure in Embodied Self-Improvement
- DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression
- Terminal-RSI: Recursive Self-Improving Terminal Agents via Recursive Self-Rewrite and Policy-Refine
- $\tau^2$-Bench: Evaluating Conversational Agents in a Dual-Control Environment
- Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents
- Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses
Related papers
- A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
- FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment
- OptiMind: Improving Small Language Models for Optimization Modeling through Formulation Guidance and Multi-Turn Learning
- Staying in the Sweet Spot: Boost Reasoning Evolution via Responsive Rollout Difficulty Manipulation
- Examining the Impact of Data and Training Methods on Emergent and Subliminal Misalignment
- Rethinking Black-Box On-Policy Distillation: A Rubric-based Analysis