Governed symbolic patches erase recurring LLM faults

A new technique fixes recurring errors in AI agents by rewriting knowledge graphs, not weights ANNEAL. ANNEAL is a neuro-symbolic agent that patches the symbolic structures encoding how tasks are executed, with governance guarantees and deterministic rollback, while leaving foundation-model weights untouched ANNEAL.

Its core mechanism, Failure-Driven Knowledge Acquisition, localizes the responsible operator, synthesizes a typed patch through constrained LLM generation, and validates the proposal via multi-dimensional scoring, symbolic guardrails, and canary testing before commit ANNEAL. Every accepted edit carries full provenance and deterministic rollback capability, anchoring repairs to an auditable trail rather than a brittle prompt or weight tweak ANNEAL.

Across four domains and 27 multi-seed runs, ANNEAL is the only evaluated system that commits persistent structural repairs, unlike strong baselines that show episodic recovery but still harbor recurring faults at holdout time ANNEAL. Strong baselines such as ReAct and Reflexion achieve high episodic recovery yet retain 72-100% holdout failure rates on recurring faults, whereas ANNEAL reduces these to 0% in the tested recurring-failure settings ANNEAL.

Ablation studies confirm that removing FDKA eliminates all structural repairs and can drop the success rate by up to 26.7 percentage points, underscoring the engine room of this approach ANNEAL. These results position the method as a complementary paradigm to weight-level and prompt-level adaptation rather than a simple add-on to existing self-evolving strategies ANNEAL.

This work signals a shift toward governance grounded in symbolic repair instead of brute force retraining, and it frames repairs as auditable edits to a process knowledge graph rather than opaque model changes ANNEAL. For product teams, that means you can fix systemic faults without retraining, potentially speeding deployment while keeping a verifiable edit history ANNEAL. The approach relies on a knowledge-graph layer and guardrails, which is a nontrivial but increasingly practical investment for teams aiming to ship safer, more reliable agents in production ANNEAL. Going forward, broader validation across domains and tighter integration with MLOps will determine how quickly these symbolic repairs migrate from experiments to everyday pipelines ANNEAL.