acceptodds
Under review as a conference paper at ICLR 2027

When the Latest State Does Not Stick: STATEFLOW for Reliable Decisions over Evolving Records

Abstract

The latest record is not always the current state. In evolving organizational records, the value that should govern a decision may need to be reconstructed from a sequence of updates involving source authority, hierarchical inheritance, entity migration, persistent overrides, and derived rules. Meanwhile, outdated but plausible values remain visible in the history and can still influence an LLM's answer. We study this failure mode, which we call the derived-state persistence gap. Using 1,200 mechanism-isolating benchmark instances, we apply matched interventions that remove stale semantics, explicitly provide the resolved state, or hide history at decision time. Across seven model configurations, these interventions improve macro exact set accuracy from 62.7% to 89.4-91.1%, showing that failures arise not only from stale information, but also from reconstructing and reasoning over the current state. We introduce STATEFLOW, which verifies whether each update is authoritative and applicable, executes valid state transitions externally, and presents the decision model with a canonical current state instead of the full history. STATEFLOW reaches 94.8% accuracy, improving over Full History by 32.1 points and over a deployable LLM State Rewrite baseline by 12.1 points. A format-matched reference-state ablation reaches the same 94.8% accuracy with identical final decision inputs, localizing the remaining system level gap to state construction. A natural history evaluation on public Jira issues shows the same benefit from explicit current-state representation and history isolation. Our results suggest that reliable reasoning over evolving records requires reconstructing the state that is valid now, rather than simply attending to what was written most recently.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Context Effects, Semantic Framing, Internal Representations

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 2 reviews, above 27% of the others.
Reviewer_KbXFOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates the challenges of maintaining accurate state information in LLMs, particularly when outdated but plausible information remains in the history and influences current decisions. The authors introduce a benchmark comprising 1,200 instances designed to isolate the effects of stale information through controlled experimental conditions. The proposed method, STATEFLOW, maintains valid state information externally and provides the model with the resolved current state. Experimental results demonstrate the effectiveness of the proposed approach, showing substantial improvements over the baselines.

Critical strengths
  • This research addresses an important challenge in understanding and reasoning over long historical contexts in LLMs. The motivation is clear and well justified.
  • The experimental setup is well designed, as it systematically distinguishes among stale-semantic interference, explicit state representation, and history isolation. The experimental findings provide a solid foundation for the proposed STATEFLOW approach.
  • The introduced benchmark dataset is also a valuable contribution and could serve as a useful resource for the research community.
Critical weaknesses
  • The generalizability of the proposed method has not been fully explored. While STATEFLOW demonstrates clear benefits in the experimental setting involving Jira tasks with diverse histories, it remains unclear whether the approach can be effectively applied to other domains requiring long-context understanding, such as long-context question answering or personal AI agents for everyday tasks.
  • The models evaluated in Table 1 appear to be somewhat outdated. It would be valuable to evaluate more recent state-of-the-art models on the proposed benchmark to assess whether the identified challenges persist in newer models.
Questions for authors

.

Reviewer_MPaAOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper addresses a problem in evolving records: when the records are modifies multiple times, and some of them are unauthorized, the LLMs should refuse to accept unauthorized modification and maintain a correct current state. The authors construct a synthetic benchmark and simulate different record variants: full history, neutral replacement, state write-back, and state-only. It also proposes a framework called StateFlow which shows better performance than other simple baselines.

Critical strengths
  1. The question studied is meaningful. The system should not trust the unauthorized record and maintain a trustful modifying history, so the motivation is sound and has practical value.
Critical weaknesses

1.The evaluation setting is far from reality. In introduction part, the paper claims to address the case that in real-world cases, we should distinguish the unauthorized modification and authorized modification. This is a reasonable motivation. However, in the experiment part, the author basically uses a synthetic dataset as the primary benchmark, which already uses the pre-defined ground truth (who is autoreactive and who isn't) to generate the dataset, which unfortunately miss the real-important point: the ambiguous and flexibly in natural language makes it difficult to identify who is really autoreactive.

2.The method is trivial and highly maps to the benchmark construction process. The method uses the programmatic way to reconstruct the current state, which is exactly how the benchmark is constructed. This justifies why the proposed StateFlow method has better performance but also makes it far from the realistic scenario and limits the method novelty.

Questions for authors

None

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.