acceptodds
Withdrawn from ICLR 2027

Self-Cleaning and Captured Anyway: A Measured Primitive for Error in a Store an Agent Writes to Itself

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Agents increasingly write their own conclusions into a store they later retrieve from, closing a loop in which an error can be read back as evidence for itself. That contamination has been reported as a one-way process whose severity decays, and it is summarised by per-model averages. Against the store deployed systems actually use — one that only ever appends — the first description is wrong and the second has no referent. We take the loop to its infinite-tenure limit against an append-only store, prove that because writing never deletes the reachable state space has a hard upper edge at (n-1)/n, and measure the process with a single primitive carrying no fitted parameter: the copy function, the probability a model repeats what its store says. The outcome is then a choice between two edges rather than a decay: at a false-record fraction of 0.9 the interval between the two modes holds 3.6% of 220 runs where a uniform spread would put 20.6%, and is strictly empty on the first 15. The primitive orders every model's capture rate to within one seed and survives the move to real facts, its sign against the critical value predicting the direction of drift on 353 of 360 real-fact runs, and scale does not rescue the store. Of four interventions with criteria frozen first, timing dominates fraction at matched budget, a consistency gate helps the false majority, and provenance marking moves nothing unless the mark correlates with truth. Two consequences travel past this apparatus: a falling exact-match score on such a store measures loss of determinism, not of information, at a measured mismeasurement of +0.383; and the resampling unit is the seed, at a design effect of 3.75 under which one of our own rank-correlation claims falls from Spearman +0.98 at three seeds to between +0.31 and +0.80 at forty-four. An append-only store's fate is therefore readable in advance from one parameter-free measurement, at a seed budget the field currently underspends.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, memory contamination, self-referential retrieval, model collapse, catastrophic forgetting, evaluation methodology, pre-registration, statistical power, retrieval-augmented generation, agent memory

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_yScKOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper studies an agent that retrieves from a growing store and appends its own answers. It measures a copying response as a function of retrieved contamination, then uses this to model how errors persist. The experiments show sharply different outcomes across seeds and suggest that grounding early can help more than spending the same budget later.

Critical strengths
  1. The seed-level results are more informative than a single mean contamination score. I like the extended runs and the matched-budget timing experiments: they make the feedback problem tangible.

  2. The authors report failed predictions and interventions rather than only successful examples. That is useful evidence about where the proposed model stops working.

Critical weaknesses
  1. The sampling model does not match the stated process. Section 3 samples without replacement yet Definition 1 uses a binomial distribution; the exact distribution is hypergeometric. This matters early: with nine false records out of ten and eight draws, the chance of retrieving the true record is 0.8, not 1 - 0.9^8, approximately 0.57. I also could not find a derivation of the claimed critical threshold gamma = 1/k from the stated drift equation. Finally, the finite-store bound (n-1)/n approaches one as the store grows, so it does not by itself establish a nonzero gap from full contamination in the infinite-tenure limit. These points affect the paper's central explanation.

  2. Some interpretations go further than the experiments support. Keeping a true record in the store does not establish that the model can still use it reliably. Likewise, pass@8 exceeding pass@1 shows that repeated sampling can recover a correct answer; it does not on its own distinguish retained capability from other sources of stochastic success. I would narrow the claims about capture and preserved capability unless there is a more direct test.

Questions for authors
  1. Clarification: does the simulator sample with or without replacement? With the correct sampling law, which drift predictions and intervention conclusions remain unchanged, and what assumptions yield gamma = 1/k?

  2. Clarification: is the infinite-tenure claim only about never reaching exactly one at a finite time, or about a limiting contamination fraction bounded away from one?

  3. Please also reconcile Section 4.3's claim that source marking lowers capture and gamma with Table 2's statement that those two predictions failed.

  4. Why are there some anonymous citations in the reference, e.g.:

Anonymous. The capability convergence hypothesis: Capability from access structure, not scale. arXiv preprint, 2026a. author list and repository withheld for double-blind review; a concurrent manuscript of the same research programme, arXiv identifier withheld.

Anonymous. Position: Preregistration Protects the Criterion, Not the Comparison: One Programme Audited Against Its Own Laws Seven Times. concurrent manuscript; author list and repository withheld for double-blind review, 2026b.

Anonymous. The Infinite-Rerun Limit: Computing what a criterion can decide, before it is frozen: a resolution, a rule, and the flip rate a band buys. concurrent manuscript; author list and repository withheld for double-blind review, 2026c.

Reviewer_ddH7OpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

[The following description constitutes my best understanding of the paper.]

The paper investigates what happens when LLM (agents) use memory/ some other form of storage to store data that they then later read back as input. In particular, the paper investigates how well factually wrong data can propagate and influence an agent's future behavior. The paper then proves theoretical results such as bounds on how well wrong data can dillute later model answers. Independently, the paper runs some experiments to back up their theoretical results.

Critical strengths

None

Critical weaknesses

I see two major classes of weaknesses with this paper:

  1. The paper is unreadable. The writing style is extremely dense, uses a lot of technical jargon, a lot of uncommon words, and feels extremely unstructured. The content itself feels scrambled, with important parts, such as clearly stating the goal of the paper, missing entirely. After having read through the paper, it is not clear to me what the main or important results are. In general, the paper contains all the signs of AI slop text. In order to improve my opinion on this point, the paper would require a complete rewriting which seems very hard to be done over the course of the rebuttal period.

  2. The paper's content has significant issues. It is hard to make a general description of this, so let me list some examples:

    • The related work section does not contain any references to related works!
    • The paper misses a description of its experiment methodology. There is no description of datasets, models used, or experiment hyperparameters. For example, it is completely unclear to me what exactly is actually being stored in the agent memory.
    • The result section simply contains a lot of text that talks about some numbers for which it is extremely hard to understand where they are coming from, or what they mean. The section does not contain any tables or figures that would help visualize its results.
    • Many mathematical terms are never introduced. Listing some examples from the introduction: What is γeff\gamma_{\text{eff}}? What is nfalse,0n_{\text{false},0}? What is WW? Furthermore, there seem to be obvious errors in some formulas. For example, equation 3 contains an empty O(⋅)\mathcal{O}(\cdot) term.

For all these reasons, I recommend to reject the paper.

Questions for authors

Dear authors. It is completely acceptable if English is not your native language and you want to use AI tools for writing the paper. However, if you want to do this, I highly recommend writing the paper first in your native language, and then letting AI translate it into English. AI tools are much better at translating existing text, than coming up with the entire text for a research paper themselves!

Reviewer_RC7hOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper studies an append-only memory loop in which an LLM retrieves records, answers a question, and writes its answer back into the store. The authors measure the probability of a false answer as a function of the retrieved false fraction, and use this function in an urn simulator. Experiments find separated low- and high-contamination outcomes, a benefit from earlier grounding at a fixed budget, and substantial sensitivity to the retrieval rule. The paper also examines transfer to factual questions and the effect of treating shared topic seeds as the unit of statistical replication.

Critical strengths

The paper studies a valid problem in agentic systems.

Critical weaknesses

I have serious concerns about the overall writing quality and clarity of this manuscript. The paper contains numerous expressions that are grammatically awkward, semantically ambiguous, or unnecessarily convoluted, to the point that I find it difficult to understand the substantive contributions.

Even the title, "Self-Cleaning and Captured Anyway: A Measured Primitive for Error in a Store an Agent Writes to Itself," is difficult to interpret. In particular, what does "a store an agent writes to itself" mean? I assume the authors are referring to a memory store that an agent updates with its own outputs, but the actual phrasing is highly unnatural and obscures a relatively simple concept. Similarly, neither "Self-Cleaning and Captured Anyway" nor "A Measured Primitive for Error" clearly communicates the research problem or contribution.

Unfortunately, this problem extends throughout the manuscript. For example:

  1. In the abstract, the authors write, "the first description is wrong and the second has no referent." The reader must reconstruct which descriptions are being referenced, and the intended meaning of "has no referent" is unclear. The same paragraph describes the outcome as "a choice between two edges rather than a decay," without adequately explaining what these terms mean.

  2. The abstract concludes with "at a seed budget the field currently underspends." This is an extremely awkward expression. If the authors mean that existing research uses insufficient random seeds, why not state this directly?

  3. In the introduction, the authors write, "Three axes of our own, taken to destruction against pre-registrations we wrote first." I cannot understand what "taken to destruction" is intended to mean in this context. The introduction also contains a subsection beginning "What stands. Four explanations offered along the way fail their tests," which reads more like a collection of rhetorical fragments than a coherent scientific explanation.

  4. In the related work section, the authors state, "We ask one question of each neighbouring result." Although one can guess the intended meaning, this is an unnecessarily obscure way of describing a comparison with prior literature. Similarly, Section 5 begins with "This paper takes three things from the framework it descends from and deliberately takes no fourth." Such language adds complexity without conveying additional technical information.

These examples are not isolated grammatical mistakes. Throughout the manuscript, the authors repeatedly employ unusual metaphors, abrupt transitions, and unnecessarily elaborate constructions in place of direct scientific explanations. Even relatively basic concepts are expressed in ways that make them substantially harder to understand than necessary.

I also find the presentation of the mathematical framework particularly difficult to follow. For instance, the mean-field equation appears abruptly in the introduction, before the stochastic model and its assumptions have been clearly established. Consequently, it is difficult to understand the motivation for the equation, its precise interpretation, or how it supports the subsequent theoretical claims. At this stage, I cannot confidently evaluate the mathematical contributions because the exposition does not make the underlying model and arguments sufficiently accessible.

The authors explicitly acknowledge using generative AI to draft and rewrite substantial portions of the manuscript. While the use of AI-assisted writing is not inherently problematic, the final text appears to have received insufficient editorial and technical scrutiny. I cannot determine how much human revision took place, but the resulting manuscript contains pervasive problems with clarity, terminology, and logical flow that should have been addressed before submission.

Overall, I believe the manuscript requires substantial rewriting rather than minor language polishing. The authors should explain the research question, model, assumptions, and contributions in straightforward technical language, define the central terminology before using it, and remove expressions that obscure rather than clarify the underlying ideas. In its current form, the paper is exceptionally difficult to understand and evaluate.

Questions for authors

See the weakness section.

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

Related papers

Discussion (0)

Sign in to comment.