Polished but Unresolved: Identifying Late-Stage Pressure States in Long-Horizon Tool-Use Agents

2609.00823 ↗ · 2026-09-01 · cs.AI, cs.CL

Haoyang Chen, Yi Liu, Jianzhi Shao, Xiaozhou Xu, Zhe Sun, Wei Hu

Accepted in the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026)

The finding

Headline claim

Long-horizon tool-use agents carry a linearly identifiable internal 'late-stage pressure' state that biases them toward polished submissions despite unresolved constraints, and a lightweight probe-and-steer plugin (PSPR) that reads and relieves this state improves constraint fulfillment on planning benchmarks.

The test

On DeepPlanning-Travel with Qwen3-14B (plus one replication at another scale/benchmark in appendix), the authors labeled action boundaries as pressure-driven commits, healthy commits, or productive continues using Gemini 3.1 Pro Preview to score constraint satisfaction and delivery polish, trained a linear probe on first-token hidden states (evaluated on 80 positives vs 80 negatives), built contrastive activation directions, and tested steering and context prompts on PresC-risk and ProdC node sets.

Effect size & scope

Probe separability is strong: AUROC 0.916 and PR-AUC 0.921 overall, 0.890 vs healthy commits and 0.929 vs productive continues. Steering effects are modest: subtracting the pressure direction moved CS from 0.21 to 0.25 and probe-measured pressure from 0.67 to 0.53, while adding it degraded CS to 0.17; residualizing out the generic commitment direction preserved the sign. Context prompts reduced mean pressure from 0.62 to 0.13 when both clarification and action mapping were added. All results are from one benchmark and mostly one 14B model.

Limitations

The authors concede gains are 'moderate' for lightweight interventions, and the full paper's PSPR benchmark tables, ablations, and uncertainty analysis fall outside the provided text. The pressure and polish labels depend on Gemini-as-a-judge, the behavioral evidence section is circular (labels define the phenomenon the probe is then claimed to discover), probe evaluation n is small (160), and generalization is shown for the probe but not demonstrated in the provided text for PSPR across models. Effects could partly reflect prompt/format artifacts at action boundaries despite the HealC/ProdC controls.

Counterpoint

A skeptic would note the 'internal state' may simply re-encode judge-scored surface features of the trajectory context rather than a causally meaningful pressure state; steering deltas of 0.04 CS are small and sign-only claims rest on few nodes; LLM-as-judge scoring of constraint satisfaction is itself error-prone; and success on one travel-planning benchmark with one open model family says little about whether production agents on heterogeneous stacks exhibit the same state.

Economy implication

Little or nothing yet for the AI economy in the next 12 to 24 months: this is a research-direction contribution on agent reliability diagnostics. Its economic relevance is only that if agentic reliability gains are to convert into enterprise ROI, failure modes like polished-but-wrong submissions are exactly the kind that require cheap internal monitoring, and this paper shows such monitoring may be lightweight rather than requiring a full evaluation pass.

Who this matters for

Technical Capability: Interpretability-minded practitioners get a concrete recipe: a linear probe with contrastive negative classes can detect premature-closure states, and constraint clarity plus action mapping in context measurably lower them.

Labor & Knowledge Work: If agent products are to substitute for multi-step knowledge work, the failure mode of confident but incomplete output is the main trust barrier buyers cite, and this work suggests it is detectable and steerable rather than fixed at pretraining.

Abstract

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.