Search papers, labs, and topics across Lattice.
This study investigates late-stage pressure states in long-horizon tool-use agents, where agents are inclined to finalize answers prematurely despite unresolved constraints. By training a linear probe to identify these pressure states from hidden states, the authors demonstrate that manipulating these states can influence the agent's decision to either continue using tools or submit an incomplete answer. The proposed Probe-Sensed Pressure Relief (PSPR) method effectively mitigates this pressure, enhancing agent performance across various benchmarks.
Agents can be misled into premature conclusions, but a new method reveals how to alleviate this pressure and improve decision-making.
Long-horizon tool-use agents need not only to search and plan, but also to decide when to finalize. We study late-stage pressure states, in which an agent is biased toward submitting a final answer that appears complete and polished while key constraints remain unresolved. We first train a linear probe to show that this pressure state is identifiable from the agent's hidden states. Then, we use activation interventions along this pressure direction and find that shifting the hidden states changes both the pressure score and whether the agent continues tool use or submits early. Through controlled context manipulations, we further see that the pressure is mitigated by constraint clarity and action mapping. Based on these findings, we propose Probe-Sensed Pressure Relief (PSPR), a plugin that applies lightweight pressure relief direction under moderate pressure and moves to structured organization under high pressure risk. Experiments on multiple long-horizon benchmarks show that our method consistently strengthens existing agent methods.
DAMO