analiza aiqbench
This research identifies "memory-induced tool-drift," a critical failure mode where LLM agents allow stored personality preferences to silently corrupt tool-call parameters in unrelated contexts. The study demonstrates that this vulnerability persists across multiple memory frameworks and affects real-world tools, with biased memories causing significantly higher parameter degradation than neutral baselines. Consequently, the findings highlight an urgent need for architectural defenses that jointly address memory management and tool calling rather than relying solely on prompt-based filters.
Memory-Induced Tool-Drift in LLM Agents
arXiv:2605.24941v2 Announce Type: replace-cross Abstract: LLM agents increasingly combine long-term memory for personalization with tool calling for taking actions. We study a failure of this combination: personality-driven preferences stored in memory (e.g., cost-consciousness, impatience, risk tolerance) silently shifting tool-call parameters in contexts where they do not apply. We call this memory-induced tool-drift and stress-test it with MEMDRIFT, an iteratively refined benchmark of 105 scenarios spanning five bias dimensions and seven professional domains. Measured as excess parameter degradation over a neutral-memory baseline, biased memories in MEMDRIFT induce drift in all eight models we evaluate, raising degradation rates to 0.38--0.83 versus 0.05--0.12 for neutral memories. The drift persists under three widely used memory frameworks. A vulnerability audit of 6,062 tools across 288 MCP servers demonstrates that real-world tools are also susceptible to this failure mode. Mechanistically, we find that biased memories act as implicit steering vectors along the same direction as explicit behavioral instructions and draw attention away from task-relevant context. Neither prompt-based defenses nor relevance filters offer a complete solution, motivating defenses at the intersection of memory management and tool calling.
Abstract:LLM agents increasingly combine long-term memory for personalization with tool calling for taking actions. We study a failure of this combination: personality-driven preferences stored in memory (e.g., cost-consciousness, impatience, risk tolerance) silently shifting tool-call parameters in contexts where they do not apply. We call this memory-induced tool-drift and stress-test it with MEMDRIFT, an iteratively refined benchmark of 105 scenarios spanning five bias dimensions and seven professional domains. Measured as excess parameter degradation over a neutral-memory baseline, biased memories in MEMDRIFT induce drift in all eight models we evaluate, raising degradation rates to 0.38--0.83 versus 0.05--0.12 for neutral memories. The drift persists under three widely used memory frameworks. A vulnerability audit of 6,062 tools across 288 MCP servers demonstrates that real-world tools are also susceptible to this failure mode. Mechanistically, we find that biased memories act as implicit steering vectors along the same direction as explicit behavioral instructions and draw attention away from task-relevant context. Neither prompt-based defenses nor relevance filters offer a complete solution, motivating defenses at the intersection of memory management and tool calling.
Cryptography and Security (cs.CR); Machine Learning (cs.LG)
arXiv:2605.24941 [cs.CR]
(or arXiv:2605.24941v2 [cs.CR] for this version)
https://doi.org/10.48550/arXiv.2605.24941 Focus to learn more arXiv-issued DOI via DataCite
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Text integral sub licența deschisă a publicației · 271 cuvinte