ANNEAL: Adapting LLM Agents via Governed Symbolic Patch Learning

Researchers introduce ANNEAL, a neuro-symbolic AI system that fixes recurring failures in LLM-based agents by directly repairing symbolic knowledge structures rather than adjusting prompts or weights. The system uses constrained generation and multi-dimensional validation to make persistent, auditable repairs, achieving zero failure rates on recurring faults where baseline approaches like ReAct and Reflexion retain 72-100% failure rates.

ANNEAL addresses a fundamental limitation in current large language model agents: while they can recover from individual errors, they fail repeatedly on the same underlying fault when the process knowledge governing task execution remains unchanged. The research demonstrates that prompt-level and weight-level adaptations, the dominant approaches in self-evolving AI systems, treat symptoms rather than root causes. By introducing Failure-Driven Knowledge Acquisition (FDKA), the system localizes which operator or process rule caused a failure, generates a typed patch through constrained LLM generation, then validates it through symbolic guardrails and canary testing before deployment.

The governance layer represents a critical advance for real-world AI deployment. Every structural repair maintains full provenance and deterministic rollback capability, addressing safety concerns that have limited autonomous agent systems in production environments. This neuro-symbolic approach bridges two traditionally separate paradigms: the flexibility of neural networks with the interpretability and verifiability of symbolic systems.

For the AI industry, these results suggest that persistent fault elimination requires moving beyond training and prompting techniques to directly editing the operational knowledge encoded in agent systems. The evaluation across 27 multi-seed runs and four domains provides robust evidence that ANNEAL's structural repairs address genuine recurring failures rather than noise. The 26.7 percentage point performance drop when FDKA is removed confirms the mechanism's necessity.

Developers building autonomous systems should consider whether their current adaptation methods address recurring failures systematically. The research opens questions about scalability: how ANNEAL performs on complex, real-world task distributions and whether the governance overhead impacts deployment speed in time-sensitive applications.

• →ANNEAL directly repairs symbolic process knowledge rather than adjusting prompts or model weights, achieving zero failure rates on recurring faults versus 72-100% for baseline systems.
• →The Failure-Driven Knowledge Acquisition mechanism localizes faults, synthesizes typed patches, and validates changes through multi-dimensional scoring before committing edits.
• →All repairs carry full provenance tracking and deterministic rollback capability, enabling safe deployment of autonomous AI agents in production environments.
• →Ablation studies confirm FDKA is essential; removing it eliminates all structural repairs and reduces success rates by up to 26.7 percentage points.
• →Neuro-symbolic approaches combining neural flexibility with symbolic interpretability offer a complementary paradigm to weight-level and prompt-level adaptation for persistent fault elimination.