acceptodds
Under review as a conference paper at ICLR 2027

HISTORY IS NOT STATE: CAUSAL DIAGNOSTICS FOR SUFFICIENT-STATE FAILURES IN LANGUAGE AGENTS

Abstract

Language agents are increasingly deployed in interactions where user goals, ev-idence, and environment states evolve over time. Final accuracy alone does not reveal where failures enter the history-to-state-to-decision chain. We pro-pose MarkovLens, a causal diagnostic framework that uses task-sufffcient states as intervention targets: canonical, non-leaking representations under which the target decision should be independent of the surface history, given the current query and decision context. We introduce four diagnostics: the sufffcient-state gap (SSG), Markov invariance violation (MIV), state update error (SUE), and obsolete-state attraction (OSA). On a 96-state balanced User Constraint Update grid, qwen2.5:7b improves from 0.833 raw-history accuracy to 1.000 gold-state accuracy (SSG = +0.167, 95% CI [0.130, 0.206]), while mistral:7b ex-hibits the opposite compact-state pattern (0.917 to 0.833, SSG = −0.083, 95% CI [−0.151, −0.023]). This opposite sign is not a contradiction: interface-equalized controls show that explicit-state rendering is itself part of the mechanism. Cross-task results extend the diagnostic beyond user constraints: on a 48-group Code Requirement Drift slice, qwen2.5:7b again shows a positive SSG (0.755 to 1.000, SSG = +0.245, 95% CI [0.135, 0.354]), while mistral:7b remains di-rectionally negative but statistically unstable. Tool/Environment Mutation shows that compact states can hurt while dialogue-style states can recover performance, and OSA calibration supports strong stale-state attraction for qwen2.5:7b but not equally for mistral:7b. MarkovLens is therefore a diagnostic microscope rather than a leaderboard or universal repair: it localizes whether models fail to infer the current state, fail to use an explicit state interface, or remain attracted to stale states.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
MarkovLens, Task-Sufficient State, Causal Diagnostics, Language Agents

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_LyubOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes a framework called MarkovLens for diagnosing failures in task state forLLM agents. It distinguishes interaction history from the current “task state”, and evaluate whether models can use the state to make the correct decisions. Experiments on Qwen and DeepSeek models show that providing correct state can substantially improve accuracy and make decisions more consistent.

Critical strengths
  1. It is insightful to differetite the distinction between correcting an error, following decision-relevant changes, and ignoring answer-preserving changes. This is beyond a simple task accuracy and is a more informative of evaluation.
  2. The experiments are carefully designed, e.g., using equivalent histories and matched counterfactual interventions to isolate the effect of state information.
  3. The results are supported by experiments with different inference settings.
Critical weaknesses
  1. The evaluation is mostly based on synthetic tasks with predefined, deterministic state transitions and answers. It is unclear how much the findings generalize to real-world agentic interactions, where the current state may be ambiguous or not observable. Although the paper includes semi-natural language experiments, these are still generated from the same underlying tasks.
  2. Related to the previos point, the framework assumes that the correct task state is readily available and can be provided to the model. However, the agent needs to extract and maintain this state from the interaction history. Table 10 shows that state extraction performance drops significantly under unseen values. Therefore, it is unclear how much the proposed framework can help diagnose real-world agents when the correct state is not readily available.
  3. The main matched-intervention experiments focus on user constraint update, while the other two task families do not have extensive evaluation.
Questions for authors
  1. Can MarkovLens be applied to real-world agent benchmarks where the current state is not predefined?
  2. Would the same repair patterns hold when the state is extracted by an LLM rather than provided as ground truth?
Reviewer_wzfhOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

I think the paper is largely AI-generated; it introduces MarkovLens, a diagnostic for whether language agents follow the current task state when a conversation contains outdated information. The paper is not well written and is only 7 pages of main content.

Critical strengths

I don't think there is any strength.

Critical weaknesses

The paper appears to be largely AI-generated, with only a trivial evaluation contribution.

Questions for authors

No questions I could ask.

Reviewer_FyL1OpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes MarkovLens, a controlled framework for studying failures to use task-relevant state in language agents. It generates replay-verified histories leading to equivalent final states and evaluates a sequence of interventions: replacing or augmenting history with an oracle state, corrupting a decision-relevant state field, and changing state fields while preserving the correct answer. Supporting analyses examine history-path sensitivity, obsolete-state attraction, and explicit state extraction.

Critical strengths

The intervention design goes beyond a raw-versus-summary accuracy comparison. Retaining the history during repair and holding the prompt layout fixed during state-content interventions are useful controls. The corruption and answer-preserving conditions provide more informative evidence than accuracy improvements alone.

Critical weaknesses

I believe the diagnostic specificity of the proposed framework is poorly established. The main results show that a model can use a supplied correct state, follow a supplied incorrect state, and often ignore edits that do not affect the answer. However, a model that treats the supplied state block as its primary decision input and largely bypasses the history could exhibit the same pattern. The experiments therefore do not sufficiently distinguish a history-to-state construction failure from difficulties in retrieving relevant information, resolving conflicting statements, or following the presentation and authority of the added state block. I appreciate that the paper explicitly avoids claiming identification of an internal mechanism. Nevertheless, a diagnostic framework should establish what its intervention signature distinguishes beyond responsiveness to a useful oracle representation. The interface and state-extraction controls are relevant, but they are not integrated into the full matched intervention sequence in a way that resolves this ambiguity. Consequently, the evidence for a distinctive diagnosis is substantially weaker than the evidence for improved performance with oracle assistance.

Also, the principal evidence for decision selectivity is not presented as a sufficiently comparable contrast, and its validation remains narrow. In particular, the analysis needs to separate answer relevance from differences in field identity, edit construction, eligible examples, and weighting. Reporting invalid outputs separately is helpful, but conditional stability should not obscure failures to return a usable answer. The BF16 Preserve result illustrates this distinction: there are no changes among 352 valid pairs, but 32/384 outputs are invalid. These concerns matter because the full intervention sequence is concentrated on the synthetic User Constraint Update task. The other task families mainly provide screening or interface evidence. The additional models and rewrites increase robustness within this setting, but do not establish that the diagnostic interpretation transfers across substantially different agent tasks.

Questions for authors

What observable outcome distinguishes a history-to-state construction failure from simply preferring the supplied state block over the history? Could you evaluate a semantically equivalent natural-language summary under matched placement, instructions, and authority using the complete repair–corrupt–preserve sequence? The existing interface controls do not appear to answer this question directly.

Can you report Corrupt and Preserve outcomes using a common weighting scheme and, where feasible, matched groups, field identities, and edit types? Please report gold-answer retention, alternative-answer changes, and invalid outputs separately, alongside both unconditional and valid-only estimates. Please also specify the exact calculation producing the 2.66% group-cluster estimate.

The primary Preserve study contains 1,600 histories across ten field paths, whereas the BF16 replication reports 384 outputs. Which fields and interventions were used in the latter, and how were they selected? Does the low answer-change rate hold across the same ten-field intervention distribution?

Does the full repair–corrupt–preserve signature hold in Code Requirement Drift and Tool/Environment Mutation? Given the substantial interface effects already reported, this would materially affect my assessment of whether the framework provides a transferable diagnosis.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.