AGENTIC-SYSTEMSJuly 16, 20261 min read

Set-shifting Behavioral Test for Harnessed Agents

Researchers have introduced a new evaluation benchmark based on the cognitive psychology concept of "set-shifting" to measure how well LLM agents adapt when the reliability of their tools silently changes. The study reveals that many open-weight LLMs struggle to recognize these hidden shifts and adapt their tool-routing strategies.

Computer Science > Artificial Intelligence

Title:Set-shifting Behavioral Test for Harnessed Agents

View PDF HTML (experimental)Abstract:What happens to an LLM agent's tool choice when the reliable tool silently changes within an ongoing session? We borrow set-shifting from cognitive psychology to study how well agents adapt to hidden reliability shifts. Our benchmark mounts tool-skill libraries with redundancies, where many tools solve the same task but differ in hidden reliability. In our evaluation framework, a branched schedule shifts the reliable tool group at hidden boundaries and pairs every shift with a no-shift control. We find that agents, by default, settle on a small recurring routine within a few turns of each boundary, with call shares concentrating on a few discrete values after each reliability shift. We score the set-shifting accuracy for each agent trajectory: the joint probability of routing to the target tool group in every post-shift window. We test open-weight LLMs in an open-source agentic harness and find qualitatively distinct failure modes across the same set of routines. We also find that set framing, how the toolset presents the alternatives as competing or complementary, shifts the routing dynamics.

Current browse context:

Bibliographic and Citation Tools

Code, Data and Media Associated with this Article

Demos

Recommenders and Search Tools

arXivLabs: experimental projects with community collaborators

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.

Source: arXiv cs.AI Recent

More in this category

agentic-systems

Multi-Agent Collaborative Reasoning with Tool-Augmented Evidence for Urban Region Profiling

Researchers have introduced UrbanAgent, a novel agentic framework that revolutionizes urban region profiling. By leveraging multi-agent collaborative reasoning and tool-augmented external knowledge retrieval, UrbanAgent outperforms traditional models in estimating population, GDP, and carbon emissions.

agentic-systems

AI advice suppresses people's willingness to say "I don't know", even when the advice is wrong and accuracy is incentivized

A new study reveals that access to AI assistants nearly eliminates people's willingness to say "I don't know," even when the AI's advice is incorrect. Despite a significant drop in accuracy, users' confidence nearly doubled, suggesting AI alters our metacognitive threshold for decision-making.

agentic-systems

CayleyR: Solving the TopSpin puzzle via cycle intersection

Researchers have introduced cayleyR, an R package designed to solve complex permutation puzzles like TopSpin. By combining Cayley graph theory, bidirectional search, and Vulkan GPU acceleration, this tool opens new avenues for combinatorial optimization.

agentic-systems

Probabilistic Extension of Neuro-Symbolic AGI Robots based on Belnap's Typed Intensional FOL

Researchers have expanded the cognitive power of neuro-symbolic AGI robots by integrating probability computation based on Belnap's logic and Nilsson's probability structure, enabling real-time decision-making under uncertainty.

agentic-systems

OriginBlame: Record- and Token-Level Data Provenance for AI Training Datasets

Researchers have introduced OriginBlame, a record- and token-level data provenance system for AI training datasets. It precisely identifies data to be removed upon author request, eliminating catastrophic over-deletion and improving machine unlearning efficiency by 42%.

agentic-systems

AI-Native Insurance for Agentic AI: Pricing, Underwriting, and End-to-End Automation

The rise of Agentic AI, capable of autonomous decision-making and environmental interaction, introduces unprecedented risk and liability challenges. A new research paper proposes an AI-native mathematical framework for underwriting, pricing, and end-to-end automation of insurance contracts tailored for autonomous AI deployments.