Technical Summary: Capability Degradation and Preservation in Lifelong LLM Agent Adaptation

Problem Definition: Capability Erosion in Self-Evolving Agents

Recent frameworks have enabled Large Language Model (LLM) agents to autonomously refine their reasoning workflows, accumulate reusable skills, update underlying model parameters, and maintain persistent memory. While these systems promise a future of agents that continually expand their competence, this paper identifies a fundamental failure mode: capability erosion under self-evolution.

The authors demonstrate that self-evolution is often non-monotonic. When an agent adapts to a new task distribution ($Y$), the process of optimizing for $Y$ frequently degrades previously acquired capabilities on earlier task distributions ($X$). This phenomenon manifests structurally across three dimensions:

1. Retrospective Capability Decay: Performance on previously mastered tasks drops after fine-tuning or updating for new tasks.
2. Behavioral Policy Drift: Workflows become unnecessarily complex and bloated, introducing noise and execution steps that were not present in successful prior behaviors.
3. Generalization Erosion: Adaptation narrows the effective reasoning space, interfering with previously acquired skills or memory retrieval patterns.

The paper argues that this is not merely a failure of specific algorithms but a structural consequence of unconstrained continual self-modification. Whether the mutable repository is a workflow graph, a skill bank, model parameters, or a memory store, optimizing solely for the current task distribution introduces destructive interference, overwriting or degrading the structures that supported successful behavior on prior tasks.

Methodology: Capability-Preserving Evolution (CPE)

To address this, the authors propose Capability-Preserving Evolution (CPE), a general stabilization principle. CPE modifies the standard self-evolution objective to include an explicit retention regularizer.

Theoretical Framework

The authors model self-evolution as a sequential process where an agent updates a capability state $R_t$ to minimize loss on a new distribution $D_t$. Naïve evolution minimizes only the current loss $L_t(R)$. CPE instead solves:
$$R^{CPE}_t \in \arg \min_{R \in \mathcal{R}} L_t(R) + \lambda \Omega_t(R, R_{t-1})$$
where $\Omega_t$ measures the deviation from prior capability-supporting structures, and $\lambda$ controls the stability-plasticity trade-off.

Local analysis (Proposition 1 and 2) shows that naïve updates increase retained risk whenever the update direction overlaps with directions of high curvature (sensitivity) for prior tasks. CPE suppresses this by penalizing updates along these sensitive directions, effectively constraining the agent to move toward the new distribution without destroying the "capability support" required by old tasks.

Dimension-Specific Instantiations

The paper implements CPE across four distinct evolution channels, each with a tailored regularizer:

1. Workflow Evolution:
   • Repository: Executable workflow graphs/prompt programs.
   • Mechanism: The regularizer $\Omega_{wf}$ discourages deviations from "anchor behavioral signatures" extracted from a seed workflow. It prevents the accumulation of overly complex validation steps and structural detours that degrade simple-task performance.
2. Skill/Tool Evolution:
   • Repository: A bounded-capacity skill bank of procedural templates.
   • Mechanism: The regularizer $\Omega_{sk}$ prevents the eviction of useful skills due to capacity limits. It employs capability-preserving consolidation, merging semantically related skills and protecting high-utility historical entries from direct deletion when new skills are admitted.
3. Model Evolution:
   • Repository: Trainable model parameters (e.g., LoRA adapters).
   • Mechanism: The regularizer $\Omega_{md}$ uses an Elastic Weight Consolidation (EWC) style penalty. It estimates parameter importance (via Fisher information) for prior domains and penalizes movement along those directions, allowing adaptation only in low-importance directions.
4. Memory Evolution:
   • Repository: Persistent external memory banks.
   • Mechanism: The regularizer $\Omega_{mem}$ uses evidence-gated preservation. Memories that accumulate repeated self-verified support are treated as stable and protected from eviction or destructive rewriting, while low-evidence entries remain adaptable.

Key Results

The authors evaluate CPE against unconstrained ("Vanilla") self-evolution across four major benchmarks and multiple model backbones (including GPT-5.1, GPT-5 nano, GPT-4o mini, Qwen3-0.6B, and Llama3.2-3B).

• Workflow Evolution (EvoAgentX on $\tau^2$-Bench):
  • Vanilla evolution led to bloated workflows and a drop in simple-task performance (e.g., from 57.1% to 41.8% on Airline tasks with GPT-5.1).
  • CPE preserved a more compact execution structure, improving retained simple-task performance to 52.8% while maintaining or improving complex-task adaptation.
• Skill Evolution (MemSkill on MATH):
  • Vanilla evolution progressively overwrote earlier skills (e.g., Geometry) as the distribution shifted to Number Theory, causing continual degradation on Algebra and Geometry evaluation sets.
  • CPE maintained stable utilization of prior skills, resulting in higher retained accuracy across all sequential domains (e.g., Geometry accuracy improved from 58.9% to 59.3% with GPT-5 nano).
• Model Evolution (STaR on MedMCQA):
  • Vanilla self-training exhibited catastrophic forgetting on earlier medical domains (Anatomy, Biochemistry) as the model adapted to new ones.
  • CPE consistently improved retained accuracy across all domains and model sizes (e.g., Anatomy accuracy on Qwen3-0.6B rose from 29.4% to 30.5%).
• Memory Evolution (Dynamic Cheatsheet):
  • Unconstrained memory updates degraded old-task retention by an average of 2.3%.
  • CPE reduced this retention gap to 0.7%, stabilizing retrieval patterns for previously learned domains without hindering new-task adaptation.

Significance and Claims

The paper claims to be the first to systematically identify capability erosion as a structural failure mode across all major dimensions of self-evolving agents (workflow, skill, model, and memory).

• Unified Perspective: The authors provide a unified interference perspective, showing that despite heterogeneous implementations, all four dimensions suffer from the same mechanism: optimizing for new tasks overwrites structures supporting old tasks.
• Stabilization Principle: The paper posits that stable long-horizon self-evolution requires not just acquiring new capabilities but explicitly preserving previously learned ones. CPE is presented as a general design principle to achieve this.
• Modest Scope: The authors acknowledge that their current preservation mechanisms are lightweight and domain-specific. They frame the work as a verification of the erosion phenomenon and a demonstration that explicit regularization can mitigate it, rather than a complete solution for all future autonomous agent challenges. They note that real-world agents may face more complex, open-ended trajectories than the controlled sequential shifts tested here.

In conclusion, the paper argues that without explicit preservation constraints, self-evolving agents will inevitably drift away from reliable behavior, and that Capability-Preserving Evolution is a necessary component for building agents capable of genuine lifelong learning.