acceptodds
Under review as a conference paper at ICLR 2027

From Memory to Authorized Action: How State Representations Shape LLM Agent Decisions

Abstract

LLM agents must retain delayed user requests across interactions and execute them when the relevant conditions are met. However, retaining a request does not establish whether it is currently authorized for execution or should instead be deferred or closed. We introduce ReCASH-PM, an empirical framework for auditing the evidence that connects remembered requests to current action decisions in frozen language models. The framework separates state acquisition, request representation, and action selection to examine how authorization evidence is preserved and used. Across 512 human-evaluated dialogue prefixes from 128 dialogues, we evaluate two model families across three quantized configurations, comparing dialogue histories, summaries, and structured states alongside alternative decision rules. Our audit reveals that original state compilation improves aggregate decision accuracy by 10.55–24.02 percentage points over direct-history prompting while reducing authorized-action recall from 72.73–93.18% to 0–3.41%. Predicate checks distinguish whether acquired fields satisfy execution conditions from whether the model selects execution, while ablations with acquisition fixed reveal sensitivity to the state components presented. A revised interface separating active-request matching from closure evidence partially restores recall while increasing false execution. Higher aggregate decision accuracy therefore does not guarantee retention of authorized actions. ReCASH-PM jointly assesses authorized-action recall, false execution, erroneous closure, and state-acquisition cost, providing a practical basis for diagnosing how remembered evidence supports or suppresses action.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Large Language Model Agents, Agent Memory, Memory-Guided Decision Making, Authorization Tracking

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_8fqiOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces ReCASH-PM, a framework that separates state acquisition, request representation, and action selection to audit whether remembered requests warrant execution. Across 512 dialogue prefixes, two model families, and 13 conditions, structured state improves overall accuracy by 10.55–24.02 percentage points while reducing authorized-action recall to 0–3.41%. Field-based rules, alternative information views, and closure checks help diagnose these failures. Additional evaluations on ABCD, EAL, ToolSandbox, and PM-Bench examine action prediction and execution. The results show that higher accuracy can conceal missed authorized actions; the revised interface partially restores recall but increases false execution, with no consistent gains on complete execution tasks.

Critical strengths
  1. The paper introduces ReCASH-PM, which separately examines state acquisition from dialogue, request-state representation, and action selection. The actions are FIRE for immediate execution, DEFER for waiting, VERIFY for verification, and RETIRE for closure. The most notable finding is that structured state can improve four-class accuracy while almost entirely suppressing actions that should be executed.
  2. The primary experiments use human reference labels and paired bootstrap resampling with dialogue-level clusters.
  3. The authors explicitly acknowledge that Revised does not meet the prespecified criterion for joint improvement and that the actual execution experiments show no consistent gains. This transparency deserves recognition.
Critical weaknesses

1 The central mechanism remains unresolved Original and Revised jointly change the schema, writer instructions, and output budget, while fields-only and summary-only expose different information (§5.2). These comparisons therefore do not isolate representation effects. Moreover, acquired fields frequently rule out FIRE on authorized cases (§5.1); historical results show that Gemma marks authorization as unknown in all 88 FIRE cases (Table 25). Although the authors acknowledge the limits of causal attribution, the evidence still cannot distinguish state-acquisition failures from failures in using the resulting representation. 2 The scientific contribution beyond the closest prior work remains unclear EAL-Bench already separates authorization errors in memory from their propagation to actions and uses memory interventions to localize failures. StateMem studies evolving state with length- and cost-matched controls, while PIS implements intention lifecycle logic through explicit state and code. This paper adds a four-action taxonomy, occurrence-specific closure evidence, and comparisons on shared dialogue prefixes. However, it remains unclear what new diagnostic conclusions these additions enable. The principal findings substantially overlap with existing work on authorization fidelity and the tradeoff between missed and unauthorized actions.

Questions for authors
  1. How can the effects of errors in state acquisition be distinguished from the effects of state representation on decisions?
  2. If the executor still cannot correctly assess authorization when given accurate memory, how can the results support conclusions about the role of memory?
Reviewer_XdCnOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

This paper studies how different memory formats affect agent decisions. The authors compare full dialogue history, summaries, and structured states. They find that structured states can improve overall accuracy while causing agents to miss many actions they should take.

Critical strengths

The paper points out an important problem that higher accuracy does not always mean better agent behavior. An agent can make fewer wrong actions simply by avoiding actions, but this also means missing many correct ones.

Critical weaknesses

The main contribution of ReCASH-PM is unclear to me. The paper introduces an auditing framework, but most experiments compare memory formats and decision rules. What does this framework provide beyond a direct analysis of model errors?

The paper tries to separate memory extraction, representation, and decision-making. However, the state-based rules also depend on the extracted state. If the state is wrong, comparing the rules with the final decision cannot tell us where the error comes from. The paper lacks a clear check of state correctness, which makes its ability to identify error causes questionable.

I question whether the main result is really about memory. Direct executes too many actions, while Original almost stops executing. The higher accuracy may simply come from acting less, rather than understanding authorization better. The paper does not rule out this explanation, so it is unclear what the results tell us specifically about memory representation.

Some experiments provide states that already contain the correct answer. For example, the full state rule reaches 100% accuracy because the supplied fields encode the reference labels. This shows that the rule works with correct inputs, not that the framework can handle real dialogue.

The paper uses many benchmarks with different tasks, inputs, and metrics. However, it does not clearly explain why the results differ or which findings hold across settings. More experiments do not necessarily provide stronger evidence for the same claim.

The revised state recovers some correct actions but also increases false execution, with no consistent improvement in full-task performance. The paper identifies a tradeoff but does not show how to address it.

Questions for authors
  • Could the authors clarify the main contribution of ReCASH-PM? In particular, how does the proposed auditing framework differ from existing evaluation methods and error analysis?
  • Could the authors explain how they separate the effects of state representation and extraction instructions? Since some experiments change both, I wonder how much each factor contributes to the results.
  • For the missed FIRE decisions, could the authors provide more analysis of errors caused by incorrect state extraction versus errors in action selection? I think this would help clarify where the main failures occur.
  • I wonder whether the differences across benchmarks are related to task types, authorization conditions, or model behavior. Could the authors discuss whether there are consistent patterns across these settings?
  • Could the authors clarify whether the accuracy gains mainly reflect a more conservative decision strategy or a better understanding of authorization? I think this distinction is important for interpreting the effect of memory representation.
  • Given that the revised state does not consistently improve end-to-end execution, could the authors discuss how the findings might help improve real agent systems?
Reviewer_62sGOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper investigates how memory representations influence authorization decisions in LLM agents. It introduces ReCASH-PM, an empirical auditing framework that separates state acquisition, request representation, and action selection. Experiments on 512 human-annotated dialogue prefixes reveal that structured state compilation can improve aggregate decision accuracy while substantially reducing authorized-action recall, and that a revised representation introduces a trade-off between recall and false execution. Additional evaluations examine closure correction and agent execution, but do not establish consistent end-to-end improvements.

Critical strengths
  1. Relevant research question. The paper addresses an important distinction between retaining user requests and determining whether they remain authorized for execution. Its observation that improvements in aggregate accuracy can conceal substantial losses in authorized-action recall is meaningful.

  2. Extensive empirical investigation. The authors evaluate multiple memory representations, decision rules, and agent benchmarks. Reporting authorized-action recall, false execution, erroneous closure, and acquisition costs provides a more informative evaluation than accuracy alone.

Critical weaknesses
  1. Insufficient methodological isolation. Although ReCASH-PM aims to distinguish state acquisition, representation, and action selection, the main comparisons do not fully isolate these factors. In particular, the Original-to-Revised comparison simultaneously changes the state schema, acquisition instructions, and generation budget. While the paper includes fixed-acquisition ablations, these do not establish the semantic correctness of the acquired fields or conclusively identify the causes of authorization failures. Consequently, the proposed framework remains largely descriptive rather than providing a convincing explanation of where and why the failures originate.

  2. Limited novelty, generalizability, and presentation quality. The contribution is primarily an empirical organization of existing state-tracking and authorization mechanisms, with limited technical novelty beyond the auditing setup. The main evaluation relies on only 128 dialogues and two relatively small model families, making it unclear whether the observed failure patterns generalize to stronger agents. Moreover, the extremely low success rates on ToolSandbox and the limited PM-Bench trajectories provide weak evidence about broader agent behavior. The presentation further reduces the paper's accessibility: several figures have excessively small labels, cluttered visual elements, and inefficient layouts.

Questions for authors

As stated in "Critical Weaknesses"

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.