acceptodds
Under review as a conference paper at ICLR 2027

Agent Carcinization: A Testable Framework for Architectural Convergence and Control Trade-offs in LLM Agents

Abstract

LLM agents increasingly rely on external runtime controls to constrain actions, track state, and support recovery. We use agent carcinization to frame the hypothesis that shared operational pressures favor recurring control structures, and ask how their benefits should be evaluated. We study evaluation contracts: the rules that determine which task outcomes, procedural requirements, and boundary constraints count as success. Re-scoring a fixed corpus of 200 episodes across four models, four tasks, and five runtime conditions reveals substantial endpoint sensitivity. On three business tasks, the model-balanced difference between a gated runtime and a placebo control changes from +27.08 percentage points under the original evaluator to −29.86 under an outcome-only ledger, −13.19 when completion is additionally required, and −2.08 when recorded boundary violations are also excluded. These comparisons hold execution trajectories fixed, isolating the effect of scoring rules rather than changes in agent behavior. External comparisons provide additional task context but do not reproduce the reversal, while targeted cancellation experiments demonstrate blocking and recovery under specified conditions. We contribute executable counterexamples, decomposed endpoints, and replayable evidence for auditing evaluation contracts. Our findings show why claims about agent runtime structure should distinguish task outcomes, procedural compliance, and boundary enforcement, while making endpoint definitions and their evidential limits explicit.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Agent

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_NG4yOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines whether the benefits of structured LLM runtimes persist across changes in evaluation criteria. Motivated by “agent carcinization,” it demonstrates that different definitions of task success can lead to opposing conclusions about the value of runtime controls.

The main contribution is an audit of how evaluation choices affect comparisons between runtimes. By rescoring the same execution traces, the authors show that the apparent advantage of capability gating can reverse when evaluation focuses on business completion. However, this reversal is substantially attenuated when compliance is also required.

The authors analyze 200 replay-verified episodes across four models, four tasks, and five configurations. Compared with the schema placebo, gating increases historical success by 27.08 percentage points but decreases business completion by 29.86 percentage points. Requiring completion without recorded violations narrows the difference to −2.08 percentage points.

Critical strengths

The paper provides a valuable evaluation audit by demonstrating how rescoring the same execution traces can reverse conclusions about runtime benefits. This helps clarify what reported improvements actually measure.

Critical weaknesses

The reversal is demonstrated on small synthetic tasks using success criteria that lack independent validation. The external experiments neither reproduce the reversal nor sufficiently exercise the gate, leaving its generalizability unsupported.

Questions for authors
  1. Gating performs slightly worse than the placebo even when completion requires no recorded violations. Could this reflect limitations in the specific gate design used in the experiments?
Reviewer_eH3UOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces “agent carcinization”, the hypothesis that shared operational pressures lead LLM agent harnesses to converge on similar control structures (e.g. state persistence, failure recovery, etc.). The main contribution is an audit of whether a harness component’s apparent benefit depends on the scoring rule. In a synthetic tool-use setting with four local models, the authors compare permission gating against a placebo with the same action schema but no enforcement. The authors rescore the same trajectories under scores that differ in which requirements they include, and show that whether gating or the placebo scores higher depends on the score, with the difference shrinking substantially once an explicit finish or no recorded violation is also required. They also check whether the gate is actually triggered, and do an external study using tau-bench which finds minimal differences, but the gate rarely activates..

Critical strengths

S1. The paper offers an audit method for checking whether the apparent benefit of a harness component depends on the score rewarding following procedural steps (e.g. inspection, validation) rather than task outcomes. The concern is shown to be practical as 48 of 200 runs received full audit credit while failing the task (Section 5.3).

Critical weaknesses

W1. As the authors note (Section 4.3), the outcome ledger deliberately omits validation and finish, which the tasks explicitly request. A reversal between this score and one that includes those steps seems largely expected. When an explicit finish is required, the gap shrinks to -13.19 points, and when no recorded violation is required, it shrinks to -2.08 points. The latter is unstable under leave-one-model-out analysis (Section 5.4). The robust finding then is that scores measuring different things disagree, which is a limited new insight.

W2. The evidence for the remaining findings is also limited: 30 runs per condition, split unevenly across models (12/6/9/3) with equal model weighting, so 3 runs from one model count as much as 12 from another. Additionally, the paper has no uncertainty estimates, its analyses are post-hoc, and in the external study the gate blocked a call in only 4 of 80 pairs, so it cannot test transfer. I appreciate the authors acknowledge some of these limits, but they substantially restrict what can be concluded.

I consider W1 to be the decisive issue: once the expected reversal is set aside, the remaining contribution is an audit method demonstrated on a small synthetic setup with minimal evidence of transfer (W2). Strengthening the findings may require substantial new work that may be beyond what a revision can provide. Since this is not my primary area of research, I may be underestimating the significance of the audit method, so I will keep an eye out for the authors’ response.

Questions for authors

Q1. The authors are motivated by whether runtime structures are useful, but stop at showing that different scores give different conclusions. Do the authors consider any of the scores to be the “correct” measure of success for these tasks? If none can be chosen, what should be the takeaway for practitioners choosing between different harnesses or runtime structures?

Reviewer_Bet2OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper examines whether the benefits of runtime control structures for intelligent agents depend on how evaluation metrics are defined—specifically, whether these metrics reward the very inspection, verification, and other procedural operations that such structures necessitate.

The authors analyze 200 trajectories across four models, four categories of synthetic tasks, and five runtime conditions. On three specific operational tasks, a "capability-gating" mechanism outperformed a "schema placebo" baseline by 27.08 percentage points in terms of historical success rates; however, when evaluated using an independent ledger of operational outcomes, it scored 29.86 percentage points lower. This gap narrowed to −13.19 percentage points when explicit termination requirements were introduced, and further to −2.08 percentage points when constraints against unrecorded violations were applied.

The paper also reports on 160 trajectories generated by the Youtu model and 32 additional trajectories from the Qwen model within the τ-bench retail environment, though no significant "rank reversal" (a shift in relative performance) was observed in these cases. The primary contribution of this work lies in its integrated approach—combining an analysis of metric provenance, the re-scoring of fixed trajectories, sensitivity to constraints, and the actual triggering of mechanisms—to critically assess the validity of claims regarding the benefits of runtime control structures.

Critical strengths
  • The boundaries of the conclusions are clearly defined. The authors acknowledge that the "ledger" omits certain explicitly required procedures and do not posit it as the sole, definitive standard; nor do they interpret the near-zero discrepancies observed in external experiments as evidence of functional equivalence. Overall, the paper’s claims regarding its own results remain measured.
  • A distinction is made between the existence of counterexamples and their actual frequency of occurrence. For instance, the fact that repeated inspections can artificially inflate an "audit" score—a counterexample at the mechanism level—is presented alongside the empirical finding that 48 out of 200 trajectories in the dataset resulted in a "perfect audit score" despite the failure of the underlying task. Reporting these separately prevents the common pitfall of inferring widespread impact solely from the existence of a vulnerability.
  • The mechanism for triggering audits holds practical value. In external experiments, only 4 out of 80 trajectories for Youtu (and 0 out of 16 for Qwen) reached the permission boundaries. The authors correctly point out that since most of these experiments did not explicitly test for the enforcement of blocking actions, the absence of observed discrepancies cannot be taken as proof that the gating mechanism is without value.
  • The experimental process is disclosed with a high degree of transparency. The paper documents the replay of cached responses, state reconstruction, failed experimental trials, protocol modifications, and interruptions followed by restarts; it also distinguishes between scripted probes and the model’s autonomous behaviors. Such documentation facilitates future verification and replication.
Critical weaknesses
  • The reversal of rankings is insufficient to prove that the original evaluation metrics were flawed. The ledger mechanism removed the "validation" and "finish" steps—which users had explicitly requested—and relaxed constraints on other processes and violation handling. Consequently, this reversal primarily demonstrates that different objective definitions yield different rankings; it does not prove that the gains from "gating" stem from unreasonable scoring. Notably, when only the "inspection" requirement was removed, the advantage of gating actually increased, suggesting that the observed reversal cannot be simply attributed to bias in the inspection process.
  • Crucial independent semantic verification remains incomplete. All labels are derived from procedures defined by the authors. While independently implementing the ledger and verifying the consistency of the replay allows for an audit of the computational process, it cannot determine which procedural requirements are truly necessary. The paper defers independent human assessment to future work, yet this step is vital for distinguishing between legitimate compliance requirements and scoring biases.
  • The control conditions fail to isolate the specific effect of the gating mechanism. Beyond the enforcement of execution constraints, the gating and placebo conditions may differ in terms of observation filtering and access to historical information. The paper reports that in 28 out of 80 relevant trajectories, unfiltered inspection history was exposed to subsequent model requests. Therefore, the observed performance differences may stem from a combination of information exposure, available actions, and execution constraints.
  • External validation is insufficient to support broad generalization. Re-running experiments across eight tasks does not expand the scope of task types, and the occurrence of edge-case triggers is too infrequent to adequately assess recovery capabilities following a rejection or the trade-offs involved in the gating mechanism. In the Qwen extension, both the agent and the user simulator were replaced, making it impossible to isolate the influence of the agent model itself. Furthermore, the fact that the test tasks were already encountered during the development phase undermines the rigor of the held-out validation.
  • There is a gap between the conceptual framework and the empirical scope. The concept of "Agent Carcinization" concerns the recurring emergence of control structures and their adaptive value; however, the paper does not examine whether this reflects structural convergence or independent emergence driven by shared pressures. The current evidence is better suited to support an audit of a specific evaluation methodology rather than a broad proposition regarding architectural adaptability.
Questions for authors
  • Could you provide a comprehensive mapping that links each task prompt to its corresponding tool prerequisites and scoring criteria, and have an independent reviewer assess whether each procedural requirement is truly necessary? Does the rank reversal persist even after this independent evaluation?
  • Could you report results that account for all requirements simultaneously—successful task completion, validation, the "finish" action, all necessary checkpoints, and full-trajectory permission compliance based on a unified definition? Current approaches—which only incorporate the "finish" action or constraints regarding past violations—fail to capture the full scope of the task requirements.
  • Could you isolate the effect of the execution-gating mechanism by keeping all other variables—such as visible information, interaction history, and tool schemas—constant, while ensuring that any attempted violations are logged before the gate is applied?
  • Could you incorporate a fixed set of scenarios—covering legitimate actions, unauthorized attempts, and recoverable errors—along with a standardized user simulator, to separately evaluate performance loss during normal operation, the effectiveness of blocking unauthorized actions, and the system's ability to recover after a request is denied?
  • Are the evidence packages mentioned in the paper available to the reviewers? Furthermore, is it possible to verify the immutable versions of the historical models, as well as the minimal reproducible examples for the anomalous scoring observed in τ-bench Task 106?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.