ANNEAL: Adapting LLM Agents via Governed Symbolic Patch Learning

2026-05-19 · Source: cs.MA updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

Anneal is a neuro-symbolic agent designed to address recurring failures in LLM-based agents by directly repairing symbolic process knowledge graphs, rather than merely updating prompts, memory, or model weights. Its core mechanism, Failure-Driven Knowledge Acquisition (FDKA), localizes the responsible operator, synthesizes a typed patch via constrained LLM generation, and validates it through multi-dimensional scoring, symbolic guardrails, and canary testing before committing. Every accepted edit includes full provenance and deterministic rollback capabilities. Across four domains and 27 multi-seed runs, Anneal consistently achieved persistent structural repairs, reducing holdout failure rates on recurring faults to 0%. In contrast, strong baselines like ReAct and Reflexion, despite high episodic recovery, retained 72–100% holdout failure rates. Ablation studies confirmed FDKA's critical role, with its removal leading to a success rate drop of up to 26.7 percentage points and the elimination of all structural repairs.

Key takeaway

For AI Architects and Machine Learning Engineers deploying LLM agents in dynamic environments, Anneal offers a critical capability for persistent fault elimination. If your agents repeatedly encounter the same operational failures, consider integrating a neuro-symbolic approach like Anneal to directly repair underlying process knowledge. This ensures durable fixes with auditable provenance and rollback, moving beyond temporary episodic recovery to truly robust, self-evolving systems.

Key insights

Anneal enables LLM agents to perform governed, persistent symbolic repairs to process knowledge graphs, eliminating recurring failures.

Principles

• Persistent fault elimination requires structural knowledge repair.
• Multi-layered governance ensures safe and auditable agent adaptation.
• High episodic recovery does not imply persistent fault elimination.

Method

Anneal's FDKA pipeline localizes failures, synthesizes typed symbolic patches via constrained LLM generation, and validates them through scoring, guardrails, and canary testing before committing with provenance and rollback.

In practice

• Use governed symbolic edits for durable fault fixes.
• Implement multi-dimensional scoring for patch validation.
• Incorporate canary testing for empirical safety checks.

Topics

• ANNEAL
• LLM Agents
• Neuro-Symbolic AI
• Symbolic Patch Learning
• Failure-Driven Knowledge Acquisition

Code references

• sbhakim/anneal-agents

Best for: Research Scientist, AI Architect, Machine Learning Engineer, AI Scientist, AI Engineer, MLOps Engineer

Related on AIssential

• Shattering the Autoregressive Curse: Dynamic Epistemic Entropy Orchestrated Erasable Reinforcement Learning for LLMs — $ ext{E}^3 ext{RL}$ allows LLMs to self-correct reasoning errors by dynamically excising localized logical defects using intrinsic uncertainty, improving long-sequence performance.
• Neuro-Symbolic Agents for Regulated Process Automation: Challenges and Research Agenda — Integrating symbolic structures into LLM agent architecture enables compliance-by-construction, preventing control-flow violations in regulated processes.
• AEL: Agent Evolving Learning for Open-Ended Environments — AEL enables LLM agents to learn from experience by dynamically selecting memory retrieval policies and integrating reflective causal insights.
• Evaluating Agentic Configuration Repair for Computer Networks — Agentic LLM architectures, augmented with verification and context tools, significantly improve network misconfiguration repair.
• Temporal Predictive AI Agents: MILKYWAY — External harnesses enable LLMs to learn continuously and make robust future predictions without internal parameter updates.
• I Built an AI That Modernizes Legacy COBOL — And Then Made It Prove Its Own Work — Separating correctness (compiler theory) from intent (LLM semantics) is key to reliable legacy code modernization.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.MA updates on arXiv.org.