LATIDIA · Investigación

Olvidar no siempre es malo: un mecanismo de reparación de la memoria neuroinspirado para agentes de LLM envenenados

arXiv: 2610.13267v1Announce Type: new Abstract: Los agentes del modelo de lenguaje grande (LLM) son muy vulnerables a los ataques de envenenamiento de la memoria. Las defensas existentes se basan principalmente en módulos externos para el aislamiento de la memoria estática

La noticia

arXiv:2610.13267v1 Announce Type: new Abstract: Large language model (LLM) agents are highly vulnerable to memory-poisoning attacks. Existing defenses primarily rely on external modules for either static memory isolation or continuous online auditing, resulting in low memory efficiency and high computational overhead. To address these limitations, inspired by neuroscientific mechanisms that weak reactivation can destabilize memories and induce decay in the absence of reinforcement, we propose an agent-intrinsic training-free memory repair framework, DREAM (Dynamic Reactivation for Engram Attenuation in Memory), that enables selective functional forgetting of poisoned memories. Specifically, DREAM implements a three-stage pipeline: perturbation-induced implicit reactivation to reveal the structural fragility of malicious memories, adaptive anomaly diagnosis based on multi-dimensional