acceptodds
Under review as a conference paper at ICLR 2027

Executable Semantic Survival in Language-Model Relays

Abstract

Repeated language-model communication can substantially change a message while preserving what a receiver does with it. Surface similarity therefore does not determine operational meaning, and one-step retention may not predict long-horizon semantic survival. We formalize this distinction by defining two messages as equivalent when a frozen receiver selects the same action in the same task context, inducing a context-conditioned executable semantic quotient. We characterize when this quotient evolves as a closed Markov process via strong lumpability and bound the error of approximate closure over finite horizons. To study long-horizon preservation, we stop a trajectory at its first action change and represent the surviving dynamics with a killed operator. Its powers recover the full first-passage survival law, revealing how equal one-step retention can hide different futures because surviving mass may occupy different within-fiber states. We further analyze observable repair using only public messages and receiver outputs, and use two frozen receivers to form survival bounds under evaluator ambiguity. Controlled constructions verify the theory. Empirically, in open-model relays up to 35B parameters, natural messages with identical current actions can still diverge under matched future rewrites, while killed-operator predictions outperform repeated one-step retention, including for held-out receivers. Together, these results show that semantic stability is a trajectory property: a message can be executable now yet be only metastably executable under repeated transmission.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
Language-Model Relays, Executable Semantics, Semantic Metastability, Markov Lumpability, First-Passage Analysis

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.60 over 5 reviews, above 70% of the others.
Reviewer_56dJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper models repeated language-model rewriting as a stochastic transmission process. Semantics are operationalized through the action selected by a frozen receiver in a fixed task context: messages inducing the same action belong to the same executable fiber. Strong lumpability characterizes when the quotient process has closed Markov dynamics. A killed operator, which removes trajectories on their first exit from the initial fiber, is used to analyze first-exit times and survival probabilities. The paper also studies observable repair based on receiver outputs and uses two receivers to report intersection and union events. Experiments cover multiple tasks, rewriter scales, transmission conditions (conservative, mixing, and compression), and held-out trajectories and receivers.

Critical strengths
  1. The paper clearly separates surface change from receiver-defined executable behavior. Its point that identical current actions can have different future survival is important and is formalized rather than treated only as an intuition.
  2. The theoretical results are tied to explicit assumptions. Strong lumpability, killed-process survival, and repair domination are stated as mathematical properties with conditions, rather than as unconditional empirical claims.
  3. The paper distinguishes receiver agreement from correctness against task labels and explicitly preserves disagreement between two receivers through intersection/union reporting.
Critical weaknesses
  1. The central notion of semantic survival is defined by a fixed candidate set and the actions of particular receivers. The paper does recognize receiver dependence, but broad statements about semantic preservation should remain explicitly qualified as receiver-defined executable preservation. Candidate-set changes and unconstrained user interpretations are outside the current operational metric.
  2. The operator estimates use design trajectories and are evaluated on held-out trajectories, which is a sound design. Still, some estimated rows may be based on limited observations, and longer-horizon predictions compound estimation error. The paper would benefit from a concise summary of effective per-state sample sizes and uncertainty for the most important language-model estimates in the main text, with sparse-state behavior made easy to locate.
  3. Repair is evaluated with two receivers that trigger when their public actions disagree or either action is invalid, and a repairer that receives the current and initial messages. This is a specific feedback protocol. The results support claims under that protocol; they do not establish that arbitrary repair systems improve semantic survival. The paper should keep this scope prominent when summarizing practical implications.
  4. Because a receiver chooses among eight candidate actions in the main relay tasks, semantic changes outside the supplied candidate set are not measured. This is a limitation of the measurement setup rather than a flaw in the Markov analysis; an explicit sensitivity study over candidate-set construction would strengthen external validity.
Questions for authors
  1. Could the authors summarize the distribution of per-state design counts and uncertainty in the estimated transition rows used for the language-model survival predictions?
  2. How sensitive are the main results to the construction of the eight-candidate action set, including distractor difficulty and near-duplicate candidates?
  3. Which repair mediation outcomes are most common across the main task conditions, including cases where repair changes a pre-repair survivor into an exit?
4 comments
AuthorsOpenReview ↗

Response to Reviewer 56dJ

We sincerely thank the reviewer for the thoughtful and constructive assessment. We appreciate your recognition of the operational semantic definition and the separation between executable preservation and gold task correctness. Your suggestions led us to clarify the fixed-context setting and add analyses of sparse transition estimates, repair mediation, and candidate-set difficulty.


W1: Receiver-defined meaning and limits of the action set

Critical weakness 1:

The central notion of semantic survival is defined by a fixed candidate set and the actions of particular receivers. The paper does recognize receiver dependence, but broad statements about semantic preservation should remain explicitly qualified as receiver-defined executable preservation. Candidate-set changes and unconstrained user interpretations are outside the current operational metric.

A1 — Response:

We agree with the importance of this qualification. Executable semantic survival is defined with respect to a frozen receiver and a fixed task context. The receiver's initial selected action determines the basin, and the survival event requires every intermediate message to remain in that basin.

This operational definition makes changes in tool routing, banking intent, and referent selection directly measurable. Gold task correctness is reported as a separate quantity. For example, an initially incorrect receiver may consistently preserve its original choice, producing high receiver-defined survival while remaining incorrect relative to the task label. Retaining both measurements makes their interpretation clear.

We have strengthened the description of the eight-candidate-per-episode benchmark and the context dependence of the induced fibers. The sensitivity analysis with harder distractors, discussed in A4, further quantifies how changing the candidate action set affects survival. For richer action spaces, the same mathematical formulation requires an explicit executable outcome criterion suited to the task.

We appreciate this suggestion and have made the operational scope prominent in the formal definitions, experiment description, and discussion.

Revised manuscript: Section 3, Executable Semantic Transmission; Section 8, Executable relay tasks; Section 9, Scope and deployment cost; Appendix K, task construction; Appendix U, Correctness and semantic survival.

AuthorsOpenReview ↗

Response to Reviewer 56dJ

W2: Sparse states and uncertainty in estimated transition rows

Critical weakness 2:

The operator estimates use design trajectories and are evaluated on held-out trajectories, which is a sound design. Still, some estimated rows may be based on limited observations, and longer-horizon predictions compound estimation error. The paper would benefit from a concise summary of effective per-state sample sizes and uncertainty for the most important language-model estimates in the main text, with sparse-state behavior made easy to locate.

Question 1:

Could the authors summarize the distribution of per-state design counts and uncertainty in the estimated transition rows used for the language-model survival predictions?

A2 — Response:

Thank you for highlighting sparse transition-state occupancy. We audited the effective source-transition exposures used to estimate the killed operators. The counting procedure includes a source transition only while its trajectory has survived through the source generation, matching the first-passage event.

Among design-observed rows of the three-variable empirical state representation, 5.64% have zero at-risk source exposures, and 23.38% have only 1–9 exposures. These counts show that the empirical transition matrix contains both well-supported and sparsely supported rows.

We also performed a trajectory-level design bootstrap with 2,000 resamples per condition/evaluator group. The mean width of the resulting horizon-grid prediction intervals is 0.111. Resampling complete trajectories respects the dependence among successive generations and provides a direct measure of finite-design uncertainty.

Alongside these diagnostics, we included an exposure-matched pooled hazard baseline and observable-state ablations. In the current manuscript, pooled hazard achieves MAE 0.0461 and the three-variable killed operator achieves 0.0382 across 720 held-out groups. The state ablation reports A0 = 0.0461, A3 = 0.0403, and A7 = 0.0382. These results suggest that observable state information improves aggregate prediction, while the sparse-row and bootstrap statistics explain why uncertainty needs to be reported together with point estimates.

The exact killed-operator survival identity continues to describe the specified microstate process. The fitted operator is a finite-data approximation, and its precision depends on available at-risk exposures. We have brought this statistical distinction into the appendix discussion and the main prediction summary.

We appreciate this constructive suggestion, which gives readers a clearer picture of how operator estimates are supported by the actual trajectory data.

Revised manuscript: Section 9, Observation 3 and Figure 4; Appendix N, prediction-error construction; Appendix P, State occupancy and uncertainty, Figure 11, and Microstate interpretation.

AuthorsOpenReview ↗

Response to Reviewer 56dJ

W3: Protocol-specific repair and induced exits

Critical weakness 3:

Repair is evaluated with two receivers that trigger when their public actions disagree or either action is invalid, and a repairer that receives the current and initial messages. This is a specific feedback protocol. The results support claims under that protocol; they do not establish that arbitrary repair systems improve semantic survival. The paper should keep this scope prominent when summarizing practical implications.

Question 3:

Which repair mediation outcomes are most common across the main task conditions, including cases where repair changes a pre-repair survivor into an exit?

A3 — Response:

We appreciate the request for a clearer characterization of the implemented repair protocol. Our mechanism evaluates each proposed rewrite using two frozen receivers. A disagreement or invalid output triggers intervention, and the repairer receives the current message together with the stored initial seed. The paired trajectory analysis measures the full downstream effect of this policy.

For immediate mediation, the proposed and repaired messages form four mutually exclusive outcomes: rescued exit, induced exit, preserved survival, and preserved exit. Across the original study's 3,287 triggered repair events, there are 1,313 rescues and 116 induced exits. The remaining 1,858 triggers preserve the current event status. The detailed mediation breakdown gives 1,071 preserved-survival and 787 preserved-exit cases.

In the 35B, 32-generation mixing panel, the task-level rescue/induced-exit counts are 22/0 (Attribute), 66/0 (Banking77), 39/0 (Intent), and 79/1 (Tool). These are immediate pre/post-repair comparisons, and they help explain why favorable immediate interventions are more frequent in that panel.

The first-passage policy effect is measured separately across the full subsequent relay. A rescued proposal can still be followed by later semantic exit, and a repaired message changes the input received by subsequent rewriters. Theorem 6.1 supplies an elementwise-domination condition for a guarantee at every horizon; the observed mediation outcomes and paired survival curves characterize the actual intervention implemented in our study.

The new direct seed-restoration comparison provides an additional practical control. Under matched 35B mixing trajectories, both LLM repair and seed restoration improve horizon-32 survival relative to unmodified transmission, with rates 0.8281 and 0.7653, respectively, compared with 0.6602. LLM repair has the higher observed survival by 0.0628, while both interventions improve over the unmodified policy.

We have made the protocol, immediate event categories, and downstream policy estimand clearer in the revised manuscript. The complete four-cell task-level breakdown is provided here as an additional response statistic; the paper presents aggregate mediation counts and per-condition trigger/net measurements.

Revised manuscript: Section 6, Theorem 6.1; Section 9, Observation 5; Appendix S, Immediate and downstream effects and Relation to domination; Appendix R, direct seed-restoration comparison, Figure 15 and Table 49.

AuthorsOpenReview ↗

Response to Reviewer 56dJ

W4: Sensitivity to distractor difficulty

Critical weakness 4:

Because a receiver chooses among eight candidate actions in the main relay tasks, semantic changes outside the supplied candidate set are not measured. This is a limitation of the measurement setup rather than a flaw in the Markov analysis; an explicit sensitivity study over candidate-set construction would strengthen external validity.

Question 2:

How sensitive are the main results to the construction of the eight-candidate action set, including distractor difficulty and near-duplicate candidates?

A4 — Response:

Thank you for proposing a candidate-set sensitivity analysis. We conducted a predeclared hard-distractor experiment for Banking77 and manual intent routing, with 64 episodes per task. Each episode retains eight candidate actions, while the harder condition uses ontology-near alternatives designed to create closer operational distinctions.

We evaluated the same unrepaired message trajectories under the original and harder candidate sets. Because the candidate context defines the executable partition, the initial receiver action is computed independently under each candidate set before constructing its first-passage survival event. Holding the rewrite trajectory fixed isolates the effect of the changed receiver context on the measurement.

At generation 32, the results are:

TaskOriginal candidatesHard distractors
Banking770.5620.344
Manual intent0.6090.578
Two-task average0.5860.461

The paired average change is −0.1250, with 95% interval [−0.1953, −0.0547]. Banking77 has a larger reduction (−0.2188); the manual-intent reduction is −0.0313, with an interval spanning zero. This task heterogeneity is informative because the candidate distinctions differ in how closely they overlap semantically.

The experiment confirms that executable survival depends on the receiver's declared action context. It also provides a concrete sensitivity measurement supporting the scope discussion.

We appreciate the suggestion, which strengthens the paper's treatment of external validity and clarifies how action-set geometry affects the observed survival event.

Revised manuscript: Section 8, Executable relay tasks; Section 9, Table 1 and Scope and deployment cost; Appendix K, task construction; Appendix R, Candidate-set difficulty sensitivity, Figure 16 and Table 50.

Reviewer_BxdSOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This paper basically studies what happens when a language model rewrites the same message many times and another language model reads each version to choose an action. It seems to examine how long that choice stays the same.

Critical strengths

There are no compelling strengths at the moment. If this paper had been framed as a study of agents and emergent misalignment, that may have been far more suitable to ICLR standards. That was not the goal of this paper, but is the closest thing that comes to my mind which may have been nice and relevant, and that may have outweighed the significant weaknesses (see below).

Critical weaknesses

1. AI overuse & Poor writing: This paper is poorly written. It reads like a haphazard collection of AI-generated [to be quite honest] derivations for a weakly motivated problem.

2. Unclear problem setting: There is no clear explanation in this paper as to why a useful system would repeatedly rewrite the same request. Why would a process of changing “Freeze my bank card” to “Block my card” to “Disable my payment card” continue for 16 or 32 rounds? Was this meant to be an implicit reference to communication between agents? The paper should explain what the task is, why it matters, and why it requires repeated rewrites or paraphrases. Without that connection, this does not appear to be a very relevant problem setting in context of existing literature.

3. Vague analytical goals: There are many proofs and experiments, but they do not collectively teach us much beyond the already-studied observation that repeated language model rewriting can distort information [1]. The paper neither makes a case for what its statements add to that understanding, nor why simpler analyses would not suffice.

[1] Mohamed et al., “LLM as a Broken Telephone: Iterative Generation Distorts Information,” ACL 2025.

Questions for authors

N/A

9 comments
AuthorsOpenReview ↗

Response to Reviewer BxdS

Dear Reviewer BxdS,

Thank you for returning to the scientific motivation and contribution questions. We understand why the original presentation could leave the role of repeated rewriting unclear: it introduced the mathematical formulation before giving a concrete workflow. Your bank-card example is particularly useful for explaining both the setting and the result. We have revised the Introduction, Related Work, and experimental description to make the operational problem explicit.

W2: Unclear problem setting

Reviewer's concern:

2. Unclear problem setting: There is no clear explanation in this paper as to why a useful system would repeatedly rewrite the same request. Why would a process of changing “Freeze my bank card” to “Block my card” to “Disable my payment card” continue for 16 or 32 rounds? Was this meant to be an implicit reference to communication between agents? The paper should explain what the task is, why it matters, and why it requires repeated rewrites or paraphrases. Without that connection, this does not appear to be a very relevant problem setting in context of existing literature.

A2 — Response:

We appreciate this question. Yes, agent-to-agent communication is an important motivating setting, together with instruction summarization, context compression, and reformatting for downstream tools. In a multi-stage workflow, one assistant may interpret a user's instruction, another component may summarize it for a coordinating agent, and a downstream router may restate it for tool selection. Each component performs a useful local operation while passing its current public version to the next component. This creates successive message transformations even when the underlying user intent stays the same. Multi-agent conversation frameworks such as AutoGen [2] provide a concrete architectural example of such handoffs.

The bank-card example illustrates precisely why we study executable preservation. Under a fixed receiver and task context, “Freeze my bank card,” “Block my card,” and “Disable my payment card” may all select the same freeze_card action. Their wording differs, but the immediate operational outcome agrees. The practical question is whether later transformations remain within this action class or eventually produce a message that selects another operation. This matters in tool-routing workflows because a changed action can lead to an incorrect tool invocation.

The 16- and 32-generation runs are stress tests of transmission depth. We isolate a single rewrite-and-receive channel and apply it repeatedly so that the number of transformations is an experimental variable. This reveals both immediate errors and delayed failures after several initially successful handoffs. At each generation, the rewriter receives the latest public message, while the same frozen receiver selects from the same eight candidates. The fixed context makes changes in the selected action directly measurable. In workflows where an intermediate message can be acted upon, the first changed action is consequential; a later return to the original wording or action leaves that earlier execution risk intact. We also record endpoint actions and returns separately.

This is a deliberately controlled model of one communication mechanism within a larger workflow. Its value is to reveal how an instruction's operational reliability changes as the number of transformations increases. We agree that this connection deserved a clearer explanation at the beginning of the paper, and your comment helped us improve it.

Revised manuscript: Section 1 (Introduction) now opens with a concrete multi-stage agent handoff and the bank-card example, illustrating why repeated message transformations matter. Section 8 (Experiments, Transmission conditions) describes the controlled 16- and 32-generation relays, with each rewrite using the latest public message and a fixed receiver and task context. Section 2 (Related Work) connects this setting to multi-agent communication frameworks [2].

References

[2] Qingyun Wu et al. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. Conference on Language Modeling (COLM), 2024. https://www.microsoft.com/en-us/research/publication/autogen-enabling-next-gen-llm-applications-via-multi-agent-conversation-framework/

AuthorsOpenReview ↗

Response to Reviewer BxdS

W3: Analytical goals, novelty, and simpler alternatives

Reviewer's concern:

3. Vague analytical goals: There are many proofs and experiments, but they do not collectively teach us much beyond the already-studied observation that repeated language model rewriting can distort information [1]. The paper neither makes a case for what its statements add to that understanding, nor why simpler analyses would not suffice.

A3 — Response:

Thank you for pointing us to Mohamed et al. [1]. We agree that it is important related work, and we have added a direct discussion in Section 2. Our work shares a similar motivation with Mohamed et al.: understanding how information changes during repeated language-model transformations. However, the two studies address distinct scientific questions. Mohamed et al. investigate textual and factual information distortion in iterative translation, including the effects of language choices, chain complexity, and prompting. We analyze the survival of actions selected by a frozen receiver and ask whether current executable equivalence determines future first-exit behavior:

If two messages select exactly the same action now, do they necessarily have the same chance of preserving that action in future relays?

This is the main analytical question of our paper. A frozen receiver defines which messages are currently action-equivalent. We then measure first-exit survival: the probability that every message through generation TT continues to induce the initial action. This outcome separates changes in surface form from changes in execution and records failures that occur before the final generation.

The theory identifies what a simpler analysis can miss. For instance, Theorem 5.2 constructs two transmission processes with identical one-step retention of 0.90.9. After eight steps, their survival probabilities are approximately 0.430 and 0.900; after 32 steps, they are 0.034 and 0.900. Both are consistent with the same first-step result. The difference comes from where the surviving probability mass moves within the current executable class. A single retention number discards that information. The killed operator retains it, with survival given by Pr⁡(τS>T)=βKT1\Pr(\tau_S>T)=\beta K^T\mathbf{1}. Strong lumpability specifies when an action-level Markov description is sufficient; the state-sufficiency criterion in Section 7 makes the missing within-class information precise.

The distinction is also testable in natural messages. In our common-future experiment, 160 message pairs have identical current action vectors, and both branches are continued under the same rewriting model with eight matched future decoding tapes. At horizon eight, the observed mean absolute survival gap is 0.144, compared with 0.072 under the specified paired-permutation null (p<10−4p<10^{-4}). This directly tests whether currently equivalent messages can follow different futures. We report the permutation assumption and finite-sample uncertainty alongside the result. In the revised aggregate-prediction experiment, the multi-state killed operator achieves held-out MAE 0.0382, compared with 0.0461 for a pooled stationary hazard fitted on matched design histories. This comparison evaluates whether retaining within-class transitions helps beyond an average retention rate.

These results explain the role of the theoretical tools. The Markov-chain results are established mathematics; our contribution is the executable first-passage problem they characterize, the conditions under which current action is sufficient, and the matched experimental tests of the resulting predictions. The repair and dual-receiver analyses extend the same operational definition to intervention and evaluator dependence.

We have revised the manuscript to organize these pieces around one clear progression: define what it means to preserve execution, identify when current action or one-step retention loses future information, test that distinction under matched continuations, and measure interventions that extend survival. We hope this makes the specific scientific contribution, and its relationship to [1], easier to assess. We appreciate the opportunity to clarify it.

Revised manuscript: Section 1 (Introduction) now foregrounds the operational question and realistic sources of message transformation; Section 2 (Related Work) directly compares with Mohamed et al. [1]; Sections 3–5 and 7 provide the executable quotient, closure, first-passage, and state-sufficiency results; Section 8 specifies the task and relay; Section 9, Observations 2–5, Figures 3–5, and Table 1 report surface/action separation, held-out prediction, matched common futures, and repair.

References

[1] Amr Mohamed, Mingmeng Geng, Michalis Vazirgiannis, and Guokan Shang. LLM as a Broken Telephone: Iterative Generation Distorts Information. Proceedings of ACL, 2025, pp. 7493–7509.

Reviewer_BxdSOpenReview ↗

Still Unanswered Questions Regarding Scientific Motivation

Dear Authors (cc. Area Chair),

Thanks, but the AutoGen citation does not justify your repeated rewriting setup. Agents communicating to solve a task does not explain why they would repeatedly rewrite the same underlying request for 16 or 32 rounds. Calling this a “stress test” doesn't answer why we need such a stress test in the first place.

For my third weakness, you basically said that you have a similar motivation to Mohamed et al., but that you study action preservation rather than textual distortion. But this still doesn't answer my concern, since you're studying that distinction in the same scenario I found unconvincing in the first place (this 16/32-round repetition). And then, rather than motivating that, you just gave more numbers and references that assumed the derivations were necessary in the first place.

I maintain my score of 1, respectfully.

Thanks,

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

This paper investigates the phenomenon of semantic drift in repeated language-model (LM) communication. The authors argue that surface-level similarity is a poor proxy for operational meaning. Instead, they propose formalizing semantic equivalence via a context-conditioned executable semantic quotient, where messages are grouped into "executable fibers" based on the discrete action they induce in a frozen receiver. The paper leverages Markov chain theory (specifically strong lumpability and killed operators) to model first-passage semantic exit. It further introduces an observable repair mechanism and dual-receiver bounds to handle evaluator ambiguity. Empirically, the theory is validated across model scales (2B to 35B) and diverse tasks, demonstrating that killed-operator predictions significantly outperform repeated one-step retention in forecasting long-horizon semantic survival.

Critical strengths

The theoretical foundations are rigorous, correctly applying concepts from Markov chain theory (lumpability, Dobrushin contraction, killed operators) to the LM relay setting. The empirical evaluation is extensive, featuring exact finite-system controls and large-scale neural relay experiments with proper design/held-out splits.

Framing LM message preservation as a first-passage time problem over executable fibers, rather than a sequence modeling or representation similarity problem, is a novel and insightful perspective. The "dual-receiver sandwich" for partial identification is a clever way to handle the lack of a ground-truth semantic oracle.

As autonomous multi-agent LM systems become more prevalent, understanding long-horizon semantic stability is critical. This work provides a much-needed theoretical and empirical foundation showing that one-step accuracy is fundamentally insufficient for predicting long-term agent behavior.

The use of exact finite-system sweeps to verify algebraic identities, combined with large-scale neural experiments across multiple model families (Qwen, Phi) and tasks (Banking77, tool routing), makes the claims highly robust.

Critical weaknesses

The empirical tasks rely on selecting from a small, fixed set of candidate actions (8 to 77 options). It is unclear how the "executable fiber" concept scales to open-ended generation, continuous control, or highly compositional tool-use sequences where the action space is vast or unbounded.

The theory assumes a fixed, frozen evaluator EcE_c . In real-world multi-agent systems, receivers may adapt, learn, or use dynamic contexts, which would cause the fiber definitions to shift over time, violating the stationary or time-indexed Markov assumptions.

While observable repair improves survival, the paper does not discuss the computational cost (e.g., additional forward passes, latency) of invoking the repairer model. In practical deployments, this trade-off is crucial.

The heavy formalism, while mathematically sound, occasionally obscures the practical intuition. For instance, the state augmentation discussion (Section 7) could benefit from a more concrete, step-by-step example of how ΔT(ψ)\Delta_T(\psi) is computed in practice.

Questions for authors

How would you define and measure an "executable fiber" in a task where the receiver's action is a free-form text generation or a complex API call with continuous parameters, rather than a classification among fixed candidates?

Could the intersection/union bounding approach (Theorem 6.2) be naturally extended to an ensemble of N>2N > 2 receivers to progressively tighten the partial identification interval? If so, what are the diminishing returns?

Could you add a brief discussion or a small ablation on the computational cost (e.g., number of extra LLM calls per trajectory) of the observable repair mechanism, and at what point the survival gains justify the latency overhead?

In Section 7, providing a brief, concrete numerical example of how the horizon-specific fiber diameter ΔT(ψ)\Delta_T(\psi) is calculated from empirical trajectory data would greatly enhance the accessibility of the "event-specific state sufficiency" concept.

3 comments
AuthorsOpenReview ↗

Response to Reviewer WZMt

We sincerely thank the reviewer for the encouraging and constructive assessment of our work. We especially appreciate your recognition of the first-passage formulation and its connection to executable semantics. Your questions helped us clarify the applicable action spaces, time-varying contexts, evaluator ensembles, repair costs, and the practical interpretation of state sufficiency.


W1: Open-ended actions and continuous parameters

Critical weakness:

The empirical tasks rely on selecting from a small, fixed set of candidate actions (8 to 77 options). It is unclear how the "executable fiber" concept scales to open-ended generation, continuous control, or highly compositional tool-use sequences where the action space is vast or unbounded.

Question:

How would you define and measure an "executable fiber" in a task where the receiver's action is a free-form text generation or a complex API call with continuous parameters, rather than a classification among fixed candidates?

A1 — Response:

Thank you for asking how executable fibers extend beyond small discrete candidate sets. We first clarify the benchmark: each episode presents exactly eight candidate actions. Banking77 supplies a larger ontology of 77 intents, from which eight candidates are displayed in each episode.

The theoretical definition starts from a declared receiver-outcome map EcE_c in a task context cc. For a structured API call, one can apply a canonicalization procedure to the selected tool and validated arguments, then define an equivalence class using the resulting execution representation. For continuous parameters, the context can specify predefined tolerances, parameter regions, or a task-relevant outcome rule. Free-form text generation likewise requires a declared evaluation or execution event that makes the desired distinction observable.

These extensions preserve the underlying principle that equivalent messages induce the same operational outcome under a fixed criterion. Their empirical application introduces additional issues, particularly the size of the action space, validation of parameterized calls, and the sample complexity of transition estimation. Our current experiments establish the basic formulation in controlled eight-way action-selection tasks and make that operational scope explicit.

We have expanded the paper's discussion with concrete API-call examples and a clearer description of the benchmark setting. We appreciate the suggestion because it helps readers distinguish the general mathematical construction from the action space used in the experiments.

Revised manuscript: Section 3, Executable Semantic Transmission; Section 8, Executable relay tasks; Section 9, Scope and deployment cost; Appendix K, Executable Task Construction.


W2: Adaptive receivers and changing context

Critical weakness:

The theory assumes a fixed, frozen evaluator E_c. In real-world multi-agent systems, receivers may adapt, learn, or use dynamic contexts, which would cause the fiber definitions to shift over time, violating the stationary or time-indexed Markov assumptions.

A2 — Response:

We agree that changing receivers and contexts introduce an important extension. In the current experiments, a frozen receiver and a fixed candidate set define a stable executable basin for each episode. This enables a clear first-passage event: the receiver's selected action must remain equal to its initial action throughout transmission.

Our nonstationary analysis allows the transmission kernel to vary with generation. For a sequence of killed kernels on a fixed basin, survival is

S(T)=βK0K1⋯KT−11.S(T)=\beta K_0K_1\cdots K_{T-1}\mathbf 1.

An adaptive receiver creates a different type of variation because the map from messages to actions, and therefore the executable fibers, may change. A future extension could augment the microstate with receiver state and context, then specify a time-varying outcome event and its relationship to the original task. That formulation would permit the attribution of action changes to communication and evaluator adaptation to be studied separately.

We have clarified the difference between generation-dependent transmission and generation-dependent semantic definitions in the revised manuscript. This makes the exact scope of the current first-passage guarantees clear and identifies the additional definitions required for adaptive settings.

Revised manuscript: Section 9, Scope and deployment cost; Appendix I, Generation-Indexed Operators; Appendix H, Context specificity.

AuthorsOpenReview ↗

Response to Reviewer WZMt

W3: More than two receivers

Question:

Could the intersection/union bounding approach (Theorem 6.2) be naturally extended to an ensemble of N > 2 receivers to progressively tighten the partial identification interval? If so, what are the diminishing returns?

A3 — Response:

Thank you for this theoretical extension. The set-based intersection/union construction extends directly to an ensemble of mm frozen receivers with basins S1,…,SmS_1,\ldots,S_m. Define

U_m(T)=\Pr\!\left(\tau_{\cup_{r=1}^m S_r}>T\right).$$ These are lower and upper bounds for a target basin satisfying the corresponding set-inclusion assumption. The inclusion relations also reveal an important property of the basic construction: as receivers are added, the intersection can only shrink and the union can only grow, giving $$L_{m+1}(T)\leq L_m(T)\leq U_m(T)\leq U_{m+1}(T).$$ Thus the raw intersection/union interval becomes at least as conservative with ensemble size. Tighter partial identification would require additional assumptions about the target evaluator or the error relationships among receivers, such as calibrated reliability or dependence structure. Those assumptions could support alternative consensus-based bounds. We now state this monotonicity explicitly in Appendix H. The distinction is useful for interpreting evaluator ambiguity: additional receivers supply more perspectives, while the assumption-light outer envelope reports their combined disagreement. **Revised manuscript:** **Section 6**, Theorem 6.2; **Appendix H**, *More than two receivers*. --- ### W4: Repair inference cost and deployment trade-off **Critical weakness:** > While observable repair improves survival, the paper does not discuss the computational cost (e.g., additional forward passes, latency) of invoking the repairer model. In practical deployments, this trade-off is crucial. **Question:** > Could you add a brief discussion or a small ablation on the computational cost (e.g., number of extra LLM calls per trajectory) of the observable repair mechanism, and at what point the survival gains justify the latency overhead? **A4 — Response:** Thank you for encouraging us to quantify the inference overhead of observable repair. We added call-count measurements from the existing runs and a matched comparison with direct seed restoration. In the original **35B, 32-generation mixing** panel, **328 of 8,192 proposed messages** triggered a repair, a **4.0%** trigger rate corresponding to **1.28 triggered repair invocations per trajectory**. The receiver-feedback protocol also evaluates two frozen receivers after each proposal; those checks are part of the overall inference budget. In the additional **256-trajectory three-policy comparison**, the LLM-repair strategy required **1,062 extra model calls** (4.15 per trajectory) relative to unmodified transmission. The direct seed-restoration strategy required **634 additional calls** (2.48 per trajectory). Their horizon-32 intersection survival rates are **0.8281** and **0.7653**, respectively. LLM-generated repair is higher by **0.0628** in the observed average; direct restoration also improves over unmodified transmission (0.6602). These are descriptive survival comparisons. These results quantify the survival–cost trade-off. Direct restoration uses fewer calls and improves on unmodified transmission, while LLM-generated repair yields the highest observed survival. The application-specific cost of an erroneous intermediate action, serving latency, batching, and model configuration determine the practical operating point. We report additional calls because that measure is directly supported by the experimental records. We have incorporated this analysis into the scope and cost discussion and the new three-policy comparison. We appreciate the reviewer for linking theoretical intervention results to deployment considerations. **Revised manuscript:** **Section 9**, *Scope and deployment cost*; **Appendix R**, *Direct seed restoration versus model-generated repair*, **Figure 15** and **Table 49**; **Appendix S**, *Observable Repair Results*.
AuthorsOpenReview ↗

Response to Reviewer WZMt

W5: Numerical example for state diameter

Critical weakness:

The heavy formalism, while mathematically sound, occasionally obscures the practical intuition. For instance, the state augmentation discussion (Section 7) could benefit from a more concrete, step-by-step example of how Δ_T(ψ) is computed in practice.

Question:

In Section 7, providing a brief, concrete numerical example of how the horizon-specific fiber diameter Δ_T(ψ) is calculated from empirical trajectory data would greatly enhance the accessibility of the "event-specific state sufficiency" concept.

A5 — Response:

We appreciate this suggestion. We added a concise three-state numerical example directly after Corollary 7.2 to illustrate the horizon-specific state diameter.

Consider messages u,v,wu,v,w in the same executable fiber with killed transition matrix

K=(0.900\000.9\001).K=\begin{pmatrix}0.9&0&0\\\0&0&0.9\\\0&0&1\end{pmatrix}.

Both uu and vv have one-step survival probability 0.90.9. From uu, surviving mass remains in a state exposed to 10% leakage at each generation, yielding h4(u)=0.94=0.6561h_4(u)=0.9^4=0.6561. From vv, surviving mass reaches the stable state ww, yielding h4(v)=0.9h_4(v)=0.9. Hence an observation ψ1\psi_1 that records only current action and one-step survival assigns the same representation to uu and vv, while

Δ4(ψ1)≥0.9−0.6561=0.2439.\Delta_4(\psi_1)\geq0.9-0.6561=0.2439.

The minimax result in Theorem 7.1 then gives a worst-case absolute prediction error of at least 0.121950.12195 for any predictor restricted to that observation. This calculation illustrates the role of within-fiber destination structure using only a small matrix.

The empirical connection appears in the common-future experiment: Appendix Q, Table 46 contains action-equivalent natural-message pairs with observed horizon-8 survival of 8/8 versus 0/8 under matched tapes. The accompanying sampling-noise analysis characterizes uncertainty from those eight continuations, so the empirical values are interpreted as observed sample contrasts.

Together, the small finite-state example and the natural-message evidence show how the same current executable action can coexist with different future survival. We thank the reviewer for suggesting this clearer explanation.

Revised manuscript: Section 7, numerical example after Corollary 7.2; Appendix G, state-sufficiency proofs; Appendix Q, Table 46; Appendix R, common-future sampling-noise analysis.

Reviewer_89tnOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates long-horizon semantic preservation in repeated language-model communication. The main contributions of this work include defining action-based semantic equivalence, developing a Markov-chain and killed-operator framework to predict semantic survival, and introducing observable repair mechanisms. Experiments across multiple language models demonstrate improved survival prediction and repair effectiveness.

Critical strengths
  1. The paper provides a formulation of long-horizon semantic preservation, distinguishing executable meaning from surface similarity and one-step retention.

  2. The experiments cover multiple models and tasks, with matched-future comparisons and held-out receivers providing useful evidence for the proposed framework.

  3. Evaluator uncertainty is reported.

  4. Appendix provides extensive additional discussion and results.

Critical weaknesses
  1. The theoretical novelty is limited, as several results build on established Markov-chain theory. The bound in (38), appears to overlook the path-dependent nature of first-passage survival.

  2. The experimental baselines are not sufficiently convincing. In particular, the comparison with repeated one-step retention does not isolate the benefit of within-fiber state modeling, as the proposed methods use additional temporal information. Stronger baselines are needed to support the claimed improvements.

  3. Some related work is missing and not discussed in the paper.

  4. There are no controls. The matched-pair differences and the common-future gap may be sampling noise. Further experiments are needed.

Questions for authors
  • A pooled stationary hazard baseline using the same temporal information should be included as the killed-operator method.

  • The authors need to clarify the derivation of (38). The bound appears to apply a result on marginal distributions to first-passage survival, which depends on the entire trajectory rather than only the final state.

  • Could the authors compare the proposed repair mechanism with a simpler baseline that directly restores the original message when a disagreement is detected? This would help isolate the benefit of the repairer itself.

3 comments
AuthorsOpenReview ↗

Response to Reviewer 89tn

We thank the reviewer for the detailed and constructive comments. Your review prompted a correction to the path-dependent argument, a stronger baseline comparison, a sampling-noise control for the common-future experiment, and a direct seed-restoration intervention. We have incorporated these results into the revised manuscript and clarified how the theoretical and empirical contributions fit together.


W1: Theoretical novelty and Equation (38)

Critical weakness 1:

The theoretical novelty is limited, as several results build on established Markov-chain theory. The bound in (38), appears to overlook the path-dependent nature of first-passage survival.

Question:

The authors need to clarify the derivation of (38). The bound appears to apply a result on marginal distributions to first-passage survival, which depends on the entire trajectory rather than only the final state.

A1 — Response:

We appreciate both observations. The path-dependent issue in Eq. (38) is substantive, and we have corrected it using a coupling of complete receiver-action histories. For initial microstates x,x′x,x' in the same executable fiber, the revised bound is

∣hT(x)−hT(x′)∣≤1−∏t<T(1−δt)≤∑t<Tδt.|h_T(x)-h_T(x')|\leq 1-\prod_{t<T}(1-\delta_t)\leq \sum_{t<T}\delta_t.

Here δt\delta_t bounds the discrepancy between next-action distributions for states within the same current fiber. Maximal coupling preserves agreement of the next action with conditional probability at least 1−δt1-\delta_t. Coupling the entire action history and tracking the absorbing exit indicator then bounds the difference between first-passage survival probabilities, including trajectories that leave and later reenter their initial basin. This correction leaves the independent killed-chain identity (Theorem 5.1) and the state-sufficiency criterion (Theorem 7.1) intact.

On novelty, we agree that strong lumpability and killed Markov chains are classical tools. Their role here is to make a concrete language-model question mathematically testable: two messages can induce the same action today yet have different probabilities of preserving that action under subsequent rewrites. The receiver-induced quotient fixes the present operational event; lumpability tests whether current action determines future transitions; killed dynamics define uninterrupted survival; and the state-diameter criterion quantifies information missing from a coarse observation.

The model-based experiment directly tests the same distinction. In the common-future design, natural message pairs have identical current receiver-action vectors and receive matched future decoding tapes from the same rewriter. Observed survival separation can therefore be studied with current executable behavior held constant.

We have revised the Introduction and theory discussion to show this progression from the question to the relevant results. We thank the reviewer for encouraging a clearer delineation of the mathematical tools and the problem-specific contribution.

Revised manuscript: Appendix G, corrected Eq. (38); Sections 3–7, formal framework; Section 2, related work; Section 9, Observation 4.

AuthorsOpenReview ↗

Response to Reviewer 89tn

W2: Matched-information prediction baseline

Critical weakness 2:

The experimental baselines are not sufficiently convincing. In particular, the comparison with repeated one-step retention does not isolate the benefit of within-fiber state modeling, as the proposed methods use additional temporal information. Stronger baselines are needed to support the claimed improvements.

Question:

A pooled stationary hazard baseline using the same temporal information should be included as the killed-operator method.

A2 — Response:

Thank you for recommending a stronger baseline with matched temporal information. We have added pooled stationary hazard, estimated from the same at-risk design transitions as the multi-state killed operator. We also included state ablations and used a consistent first-passage counting convention for all predictors. Each method is evaluated on the same held-out trajectories and predeclared horizons.

Across 720 held-out groups, the reported MAEs are:

PredictorHeld-out MAE
First-step retention0.1356
Pooled stationary hazard0.0461
Multi-state killed operator0.0382
Generation-indexed hazard0.0496

The killed operator improves over pooled hazard by 0.0079 MAE (17.1% relative reduction), with paired 95% interval [0.0008, 0.0175]. This directly evaluates the effect of within-fiber destination structure under a matched-information comparison.

For the state ablation, A0 (one pooled surviving state) gives 0.0461, A3 (repair status) gives 0.0403, and A7 (all three observable variables) gives 0.0382. The added state information improves on pooling, while A7 achieves the lowest MAE among these three representations, with an additional improvement of 0.0021 over A3. The other state combinations are displayed in Figure 11.

The results also refine the interpretation of the theoretical operator. The exact formula Pr⁡β(τS>T)=βKT1\Pr_\beta(\tau_S>T)=\beta K^T\mathbf 1 describes survival under the specified microstate kernel. The empirical study asks how accurately a compact, estimated state representation approximates this behavior. The pooled comparison now tests the practical value of retaining state transitions beyond an overall survival rate.

We have updated the main figures, Table 1, and the appendix tables to present the matched comparison consistently. We appreciate the suggestion, which strengthened the empirical test of our central mechanism.

Revised manuscript: Section 9, Observation 3, Figure 4, Table 1; Appendix M–N, estimation and statistics; Appendix P, Survival-Prediction Estimates, Figure 11 and 19 operator tables; Appendix Q, held-out receiver analysis; Abstract and Section 10.


W3: Missing related work

Critical weakness 3:

Some related work is missing and not discussed in the paper.

A3 — Response:

Thank you for highlighting the related-work gap. We expanded the discussion to include Mohamed et al., LLM as a Broken Telephone: Iterative Generation Distorts Information (ACL 2025), together with relevant research on multi-agent message handoffs.

Our work shares a similar motivation with Mohamed et al.: understanding information preservation during repeated language-model transformations. However, the two studies address distinct scientific questions. Mohamed et al. examine textual and factual information distortion during iterative generation. We study receiver-defined executable action preservation, especially whether two messages with identical current actions can have different future first-exit behavior. Our formulation measures the first action change while allowing extensive rewriting within an executable fiber.

The question gains importance in multi-stage systems where a request is repeatedly paraphrased, normalized, summarized, or compressed before a downstream component selects an action. The Introduction now provides a concrete handoff example, while Section 8 explains how the controlled 16- and 32-generation relays test accumulated transmission risk under a fixed receiver and task context.

The revised text also connects the empirical common-future intervention to this motivation: two messages that produce the same current action may lead to different future actions under matched subsequent rewriting. The result adds a future-oriented, action-based measurement to the broader literature on iterative language transmission.

Revised manuscript: Section 1 (Introduction), opening multi-stage handoff example; Section 2 (Related Work), Cultural transmission with language models (shared motivation and distinct research questions); Section 8 (Experiments), Transmission conditions (controlled relay horizons and fixed context); Section 9, Observation 4.

AuthorsOpenReview ↗

Response to Reviewer 89tn

W4: Sampling noise and empirical controls

Critical weakness 4:

There are no controls. The matched-pair differences and the common-future gap may be sampling noise. Further experiments are needed.

A4 — Response:

We appreciate the request for a statistical control on the natural-language common-future comparison. The original experiment uses 160 action-equivalent message pairs, each continued under eight matched decoding tapes. With eight observations per branch, empirical absolute survival gaps contain finite-sample variation. We have therefore added an explicit calibration of that variation.

We performed 20,000 within-tape branch-label permutations. Under a null model in which the two branches are exchangeable within a matched tape, each permutation yields a mean absolute gap across the message pairs. This provides a reference distribution for the empirical gap statistic.

At horizon 8, the results are:

StatisticValue
Observed mean absolute gap0.1438
Mean gap under the permutation null0.0718
Excess over null mean0.0720
Permutation p-value0.00005

The observed absolute separation is roughly twice the null mean. A message-pair bootstrap of the signed seed-minus-paraphrase contrast gives 0.0563, with 95% interval [0.0125, 0.1016]. The bootstrap unit is a complete message pair, preserving the dependence among its eight tapes.

We state the statistical assumptions explicitly. The permutation inference uses within-tape exchangeability under the null. The fraction of pairs with unequal observed eight-tape survival rates (35.6%) is a descriptive statistic; the permutation null itself yields nonzero empirical differences for some pairs. The experiment supports a difference in future behavior beyond the specified sampling-null reference, with uncertainty appropriately attached to message-specific rates.

The revised results strengthen the empirical link to the state-sufficiency theory while keeping the finite-sample interpretation precise. Thank you for prompting this important additional control.

Revised manuscript: Section 9, Observation 4 and Table 1; Appendix Q, within-fiber branch construction and Table 46; Appendix R, Common-future sampling noise, Figure 14 and Table 48.


W5: Direct restoration instead of an LLM repairer

Question:

Could the authors compare the proposed repair mechanism with a simpler baseline that directly restores the original message when a disagreement is detected? This would help isolate the benefit of the repairer itself.

A5 — Response:

Thank you for proposing direct seed restoration. This provides a simple and informative comparison for isolating the practical benefits of the receiver-feedback intervention.

We conducted a matched 35B mixing, 32-generation experiment with 64 episodes per task across Attribute, Banking77, Intent, and Tool. The three policies were: unmodified transmission, LLM repair after the original disagreement/invalid-output trigger, and direct restoration of the initial public seed under the same trigger. Both interventions take effect during the relay; their outputs become the next generation's rewriting input.

The final intersection-survival rates, averaged equally across the four tasks, are:

PolicySurvival at generation 32
Unmodified transmission0.6602
LLM-generated repair0.8281
Direct seed restoration0.7653

Both interventions improve survival over unmodified transmission. LLM-generated repair achieves the highest observed survival (0.8281), exceeding direct seed restoration by 0.0628. Direct restoration also improves survival by 0.1051 relative to unmodified transmission. These are observed point-estimate differences; the comparison does not establish a statistical significance claim.

The computational comparison provides additional context. Relative to unmodified transmission, LLM repair uses 1,062 additional model calls, compared with 634 for seed restoration. Direct restoration removes the generative repair step while retaining receiver-based detection and downstream verification.

The result emphasizes the practical role of observable receiver feedback together with access to the original message. Theorem 6.1 characterizes sufficient conditions for favorable survival effects at the operator level; this experiment evaluates two concrete intervention policies under the same trigger.

We appreciate this suggestion, which adds a useful and computationally transparent repair alternative.

Revised manuscript: Section 9, Observation 5 and Table 1; Appendix R, Direct seed restoration versus model-generated repair, Figure 15, Table 49; Appendix S, Observable Repair Results.

Reviewer_Xt64OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies whether repeated language-model rewrites preserve the action taken by a fixed receiver. It defines action-based semantic equivalence, models first semantic exit using killed Markov operators, and evaluates survival prediction, paired continuations, and observable repair. I recommend weak acceptance. The distinction between current executability and future stability is useful, and the paired continuation experiments provide meaningful empirical support. My reservations concern a specific survival bound and the attribution of prediction gains, rather than the basic first-passage formulation or the observed phenomenon.

Critical strengths
  1. The operational definition makes semantic preservation concrete and testable. Distinguishing uninterrupted survival from endpoint agreement avoids overlooking intermediate errors that could already have caused an incorrect action. The paper also appropriately separates preservation of the initial interpretation from task correctness.

  2. The paired continuation experiment directly examines the central question. Different messages with identical receiver actions are exposed to matched future decoding seeds, showing that current action alone need not determine future stability. The additional tasks, model scales, and held-out receivers strengthen the evidence beyond a single relay configuration.

Critical weaknesses

The connection between approximate closure and first-passage survival needs correction. The current argument bounds the distribution at a particular time, whereas survival depends on the entire preceding trajectory. A message can leave its original semantic basin and later return, making its final state appear stable even though survival has already failed. The authors should revise the argument to track whether an exit has occurred. This appears to be a localized issue and does not invalidate the basic killed-chain survival formulation.

The prediction results do not yet isolate the contribution of within-basin destination structure. The stationary operator uses transitions throughout the design trajectories, while the repeated-one-step baseline uses only initial retention. Appendix I also identifies the time-product estimator with the empirical design survival curve. A pooled-hazard baseline using all design transitions and explicit state ablations would make the explanation of the gains more convincing.

Questions for authors
  1. Can you provide a corrected version of Equation 38 using an absorbing augmentation and identify which subsequent conclusions depend on this bound?

  2. How much predictive improvement remains over a single-state hazard model fitted to all design transitions, and which state variables contribute that improvement?

4 comments
AuthorsOpenReview ↗

Response to Reviewer Xt64

We sincerely thank the reviewer for the careful and constructive review. We appreciate your recognition of the central distinction between current executability and future stability, especially the evidence from paired continuations. Your suggestions led us to strengthen the first-passage argument and introduce a matched-information prediction baseline. We address both points below and identify the corresponding manuscript changes.


W1: First-passage bound and absorbing event

Critical weakness:

The connection between approximate closure and first-passage survival needs correction. The current argument bounds the distribution at a particular time, whereas survival depends on the entire preceding trajectory. A message can leave its original semantic basin and later return, making its final state appear stable even though survival has already failed. The authors should revise the argument to track whether an exit has occurred. This appears to be a localized issue and does not invalidate the basic killed-chain survival formulation.

Question 1:

Can you provide a corrected version of Equation 38 using an absorbing augmentation and identify which subsequent conclusions depend on this bound?

A1 — Response:

Thank you for identifying the path-dependence issue. We agree that first-passage survival depends on the full action history. Theorem 4.2 concerns marginal distributions at a particular generation, while survival requires every action up to the horizon to remain in the initial basin. A return to the original action after an earlier exit still counts as an exited trajectory.

We have replaced Eq. (38) with a path-level coupling bound. Let δt\delta_t be the maximal within-fiber total-variation discrepancy between next-generation action distributions at generation tt. For initial microstates x,x′x,x' belonging to the same executable fiber, the revised result is

∣hT(x)−hT(x′)∣≤1−∏t=0T−1(1−δt)≤∑t=0T−1δt.|h_T(x)-h_T(x')|\leq 1-\prod_{t=0}^{T-1}(1-\delta_t)\leq \sum_{t=0}^{T-1}\delta_t.

The proof couples the complete receiver-action histories. Whenever the two current action classes agree, their next action classes can be maximally coupled, with agreement probability at least 1−δt1-\delta_t. The two complete action histories therefore agree through horizon TT with probability at least ∏t<T(1−δt)\prod_{t<T}(1-\delta_t). We also track an absorbing exit indicator, so any intermediate exit remains recorded through later generations. On the coupling event, the two first-passage indicators coincide, yielding the stated bound.

The correction preserves the roles of the other results. Theorem 4.2 continues to bound marginal quotient-distribution error. Theorem 5.1 establishes the exact killed-operator survival identity, and Theorem 7.1 gives the survival-state sufficiency criterion; their proofs and the empirical first-exit measurements are independent of the former Eq. (38). The contraction factors in Theorem 4.2 apply to endpoint laws, while the new product bound applies to full survival paths.

The revised result retains the intended interpretation: small within-fiber discrepancies constrain differences in future executable survival, and exact lumpability yields equal horizon-specific survival values throughout an executable fiber. We appreciate this constructive correction, which makes the theoretical connection precise.

Revised manuscript: Appendix G, Connection to approximate closure: a path-event bound (Eq. 38 and proof); Section 7, Event-Specific State Sufficiency; Section 5, Theorem 5.1.

AuthorsOpenReview ↗

Response to Reviewer Xt64

W2: Fair prediction baseline and state ablations

Critical weakness:

The prediction results do not yet isolate the contribution of within-basin destination structure. The stationary operator uses transitions throughout the design trajectories, while the repeated-one-step baseline uses only initial retention. Appendix I also identifies the time-product estimator with the empirical design survival curve. A pooled-hazard baseline using all design transitions and explicit state ablations would make the explanation of the gains more convincing.

Question 2:

How much predictive improvement remains over a single-state hazard model fitted to all design transitions, and which state variables contribute that improvement?

A2 — Response:

Thank you for proposing the pooled-hazard comparison and state ablations. They directly address the source of predictive improvement. We now fit pooled stationary hazard and the multi-state killed operator using the same design episodes and the same at-risk source transitions. This controls for the amount of temporal training information available to each estimator.

The revised evaluation compares first-step retention, pooled stationary hazard, the multi-state killed operator, and generation-indexed conditional retention on identical held-out groups. The manuscript reports the following matched-information results across 720 held-out condition/evaluator groups:

PredictorHeld-out MAE
First-step retention0.1356
Pooled stationary hazard0.0461
Multi-state killed operator0.0382
Generation-indexed hazard0.0496

The multi-state killed operator has the lowest MAE in the reported comparison. Its improvement over the exposure-matched pooled hazard is 0.0079 MAE, or approximately 17.1%, with paired 95% interval [0.0008, 0.0175]. This comparison provides direct empirical support for the predictive value of destination information within the surviving executable fiber.

We also evaluated observable-state ablations using receiver agreement, seed equality, and repair status. The A0 pooled representation has MAE 0.0461, the A3 repair-status representation has MAE 0.0403, and the A7 full three-variable representation has MAE 0.0382. Both A3 and A7 improve over A0. The complete A7 representation achieves the lowest MAE among the three, indicating an additional improvement from combining the observable state variables.

The first-passage estimators use source transitions from trajectories that have survived through the source generation. This maintains the event definition throughout estimation. The first-step comparator uses observed generation-one retention, while the other estimators share the appropriate design histories.

We additionally repeated the analysis under held-out receivers that define new executable partitions. Across 192 condition/receiver groups, the reported MAEs are 0.1253 / 0.0482 / 0.0421 / 0.0503 in the same method order. The killed operator retains a small point-estimate advantage over pooled hazard (0.0061, about 12.7%). The comparison uses the same held-out trajectories and executable-event definitions.

These results connect the theory and measurement more directly. The killed operator records movement among surviving message states, and the matched-information experiment evaluates the contribution of that structure to aggregate survival prediction. We appreciate the suggestion, which made the experimental comparison considerably more informative.

Revised manuscript: Section 9, Observation 3, Figure 4 and Table 1; Section 9, Observation 4 and Figure 5(b); Appendix M, estimation definitions; Appendix N, held-out error construction; Appendix P, Survival-Prediction Estimates, Figure 11 and the operator tables; Appendix Q, held-out receiver analysis and Figure 5(b); Abstract and Section 10.

Thank you authors, for your detailed response. I believe your responses substaintially addressed my major concerns, I would remain my positive suggestion of score unchanged. I would like to drill down with one last follow-up question. Could you clarify, since repair status alone accounts for most of the improvement, is the additional gain from A3 to A7 statistically reliable? Thanks.

Dear Reviewer Xt64,

Thank you for your encouraging feedback and insightful follow-up.

We conduct a paired bootstrap analysis across the 720 held-out groups. The additional MAE reduction from A3 (0.0403) to A7 (0.0382) is 0.0021, with a 95% confidence interval of [0.0005, 0.0037] and p = 0.011.

This indicates a modest but statistically significant improvement from the additional state variables beyond repair status alone.

We appreciate your suggestion, which strengthens our statistical assessment.

Best regards,

The Authors

Comments on the paper

12 comments
Reviewer_BxdSOpenReview ↗

Concern of AI use from my fellow reviewers Xt64, 89tn, WZMt, 56dJ and the paper authors

Dear Area Chairs,

You may have noticed my earlier comment/concern that this appears to basically be an AI-generated paper.

I now wish to add that I am also highly alarmed by the other reviewers Xt64, 89tn, WZMt, and 56dJ.

Their reviews also come across as AI-generated!

I'm not sure what is causing this, but if you read what the other reviewers say, they all have the same "score 3 confidence 2", and all delve right into the same poorly motivated and non-standard terminology that spurred my AI-overuse concerns about the paper itself in the first place.

I thought to raise this urgent concern on time, given that ICLR is known to have high standards among conferences.

Many thanks,

Reviewer_BxdSOpenReview ↗

3rd response on AI overuse and overreliance in this paper

Dear Authors (cc. Area Chairs),

Thank you for stating that you used GPT-6 Astra. However, describing the AI use here as just for "equation consistency checking" significantly understates the role AI clearly played in this paper. This would be evidently clear to anyone who reads it.

My concern is not about the use of AI for equation checking. It is about the weak underlying research question and the writing that looks like it was generated by AI. Even just looking at the vast tables in the appendix that look like a direct copy-paste from AI output.

The paper introduces extensive formalism without clearly explaining why the problem matters or what we learn from it. As stated in my first review, why would a useful system repeatedly rewrite the same request for 16 or 32 rounds? What do we learn beyond the already-studied observation that repeated rewriting can distort information?

Simply adding lots of proofs, or showing substantial computation and ablations, does not automatically constitute novel contribution (especially when those artefacts look auto-generated by AI).

Also, simply adding the "cherry on top" of a human created looking Figure 1 (a figure which is also not very scientifically compelling) does not meet the bar for fixing this obvious AI overuse for basically everything else.

I appreciate the clarification about your process, but these substantive concerns remain unanswered. Per my extensive reviewing experiences, I maintain my score of 1.

Reviewer_BxdSOpenReview ↗

4th response on AI overuse and overreliance in this paper (& re-requesting on the weak research motivation)

Dear Authors (cc. Area Chairs),

Well, I have already made clear my concern that the authors understated their AI overuse. When I referred to “equation consistency checking,” I was using your own words in the OpenReview reply right above. Also, your "AI Usage Statement" does itself look AI-generated (to be quite honest). This is moreso a big issue because of the honestly shell-shocking AI overuse in virtually every other part of this paper.

But besides that, you also did not answer the serious questions about scientific motivation here.

A well motivated and well written paper should have been submitted in the first place. Promising to revise the writing doesn't automatically resolve things.

Thanks,

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.