Title: Event-Specific State-Action Binding Predicts Action Control in Language Agents

Abstract:

While autonomous language agents are increasingly characterized by their exposure of traces, memories, plans, and constraints, current evaluation methodologies seldom verify whether these internal state variables are genuinely linked to the final actions taken. To address this gap, we present "causal state binding," an evaluation framework centered on interventions that assesses whether actions shift in response to specific, decisive state events while remaining stable against irrelevant cues. Our primary metric utilizes a hidden-target finite-action benchmark, where intervention targets are designated by the scorer prior to generation and kept hidden from the model’s visible prompt.

We analyzed 57,816 scored records distributed across seven corpus-level units. The results indicate that structured-agent conditions outperformed both high-randomness controls and targeted component removals in terms of responsiveness regarding reason, memory, veto, and self-continuity. Open-weight validation involving Qwen2.5 (7B, 14B, and 32B) and Mistral-7B demonstrated that mechanisms such as action priors, field-less prompts, or scrambled decisive contexts failed to replicate the signature of structured control. Furthermore, diagnostic finite-action probes revealed that only the minimal decisive-field readout successfully recovered the prescribed action patterns, whereas controls relying solely on surface features, action priors, or scrambled fields did not.

In a practical application across 300 SWE-bench Lite issue records and six API models, integrating an oracle-free causal state-binding composite into a full non-CSB baseline improved the constraint-clean issue-to-file hit@3 AUC from 0.873 to 0.935. It is important to note that this validation focuses exclusively on issue-to-file localization, rather than patch application or overall issue resolution. Collectively, these findings advocate for a new measurement principle in agent evaluation: action control is best predicted by event-specific state-action binding, rather than by output entropy, action-prior matching, or rationale format in isolation.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC