acceptodds
Withdrawn from ICLR 2027

Action Admission and Constraint Evolution: Governing LLM Agent Actions and Learning from Failures

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

LLM-based agents propose actions with consequential effects, some of which violate system constraints. We introduce Action Admission and Constraint Evolution (AACE), a model-agnostic framework that separates action proposal from execution. An Action Admission Controller checks each proposal against predefined constraints and a Constraint Bank. Verified Constraint Evolution generates constraints from observed failures and admits them only after empirical verification and a Candidate Curation Gate. We evaluate AACE in negotiations adapted from AgenticPay and procurement scenarios adapted from RetailBench. In AgenticPay, three of 26 candidate families pass verification and curation, all concerning privacy. On 60 held-out profiles across 540 episodes, the validated Bank reduces attack violations from 57.0% to 31.1% with no observed benign false blocks; admitting all candidates blocks 13.3% of benign episodes. With a Claude executor, the same frozen Bank reduces attack violations from 52.8% to 25.0%. Static admission with deterministic recovery raises valid success from 12% to 96% in the GPT-5.4-mini experiment. In a matched comparison with ToolGuard and a one-revision host, both methods have zero observed executed violations and show no significant difference in valid success (88% versus 92%); AACE has higher mean seller reward and lower latency in this setup. In RetailBench, a verified constraint prevents duplicate procurement in all 12 held-out risky scenarios and blocks none of the 12 benign scenarios. These results support verified constraint updates for specific failure patterns.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
LLM agents, agent safety, pre-execution control, verified constraint evolution, multi-agent systems

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

This paper presents AACE, which checks agent actions before execution and learns new constraints from failures. Proposed constraints are tested for violations and unnecessary blocking before being added to a persistent rule bank. Experiments in negotiation and procurement show fewer observed violations, although improved valid task completion is not statistically established in the main constraint-learning experiments.

Critical strengths

The paper addresses a practical problem: adding safety rules can also block legitimate actions. In Table 2, the validated Bank reduces Qwen attack violations from 57.0% to 31.1%, with no observed false blocks on benign episodes. Adding all candidates instead causes 13.3% benign false blocks. This provides useful evidence for checking rules before deployment.

The evaluation separates action enforcement from recovery after rejection. The recovery ablation and matched ToolGuard comparison make this distinction concrete. I also appreciate the explicit reporting of unsuccessful candidates, uncertainty intervals, and the lack of established gains in valid task completion.

Critical weaknesses

The experiments do not isolate the benefit of the proposed constraint-selection procedure. The main comparison is between three validated rules and all 26 candidates, which changes both selection and Bank size. A simpler selection method retaining three rules could potentially achieve similar results. The Curation Gate accepts every candidate that reaches it, so its additional value is also untested. ToolGuard is a useful static baseline, but it does not address the adaptive contribution; the distinction from adaptive guardrails such as AGrail (Luo et al., ACL 2025) also needs discussion.

The evidence for constraint learning remains narrow and dependent on the semantic evaluator. One learning run produces three privacy rules, with no admitted role or time-wasting rules. Both the gate and judge use Qwen-plus, including when Claude executes the task. Blinding the judge helps, but does not rule out shared semantic errors; Section 4.2 already reports an inconsistent judgment in a case with no intervention. An independent audit is needed to establish how much the main results depend on these labels. RetailBench provides independent deterministic evaluation, but only for one duplicate-procurement pattern.

Questions for authors

What evidence distinguishes the verification procedure from simply retaining a small Bank? A comparison with a simple selection baseline constrained to three rules would directly affect my assessment of the contribution.

Do the violation and false-block reductions remain under an independent human audit or a separately validated evaluator? Please include safe steps blocked within attack dialogues, not only benign episodes.

What does the Curation Gate contribute beyond empirical verification in the evaluated setting, and how does the update procedure differ from existing adaptive guardrails such as AGrail?

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

The action admission problem that the paper focuses on is important in agent systems. The paper proposes AACE, a framework working between action proposals and environment execution, which maintains a constraint bank derived from failures to evolve the modules. The proposed method is validated in two benchmarks and demonstrates its advantages.

Critical strengths
  1. The focused problem is important and practical, and the overall method design is rational and self-consistent.
  2. The experiment verifies the method through improved performance in the safety-utility tradeoff and demonstrates the advantages compared with baselines.
Critical weaknesses
  1. Experimental verification can be improved: the constraint evolution should be compared with some simple policy- or rule-based methods; only the privacy-related constraints in the final bank, which weaken the generality of the method.
  2. The quality of the figures can be improved: fig2 misses some process details (e.g., whether the bank updates when the controller chooses revise/escalate to re-check); fig3 looks vague; fig4 looks like a sketch
Questions for authors
  1. All 23 rejections occur in the verification stage, so how to prove the functions of the curation gate?
  2. How reliable are the safety labels?
  3. The gate and judge both use the same LLM family. How is performance affected?
  4. The recovery policies of AACE and ToolGuard are different. How convincing of the performance comparison?
  5. Can the method be applied to general agentic tasks?
Reviewer_rmoTOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

The paper aims to solve the problem of evolving the constraints to be imposed on LLM agents, using a technique called Action Admission and Constraint Evolution (AACE). Specifically, the paper mentions Economic Multi-Agent Tasks as one of its focuses. The evaluation presents results from using AACE and its different ablations in different settings (open and closed weight model harnesses).

Critical strengths

The paper aims to solve the problem of evolving the constraints to be imposed on LLM agents, using a technique called Action Admission and Constraint Evolution (AACE); the evaluation presents results from using AACE and its different ablations in different settings (open and closed weight model harnesses).

Critical weaknesses

The paper seems to aim at solving an interesting problem, but suffers from unreadability due to several unexplained concepts, underspecified methodology, and a lack of proper justification for many choices made in the technical section.

Take, for example, "a proposal": the paper defines a proposal to be one of "natural-language response, a structured decision, a tool call, or a control operation". It goes on to say that there is a set of constraints C that governs execution. The paper does not give a general formulation or characterization for C in any more detail. Moreover, what could the expressivity of the constraint definitions be that it is able to range over all four types of the proposal mentioned previously?

The paper then defines an "Economic Multi-Agent Task" using phrases like "governed agent task in an economic setting", where "agent actions may include offers, prices, discounts, refunds, commitments, recommendations, or tool calls". The paper does not define in clear terms the capabilities of such an agent or even the task setting itself.

Figure 1 mentions "Predefined Hard Constraints" and includes "pricing, privacy and authority rules" as some examples. But it does not explain in the paper what these hard constraints can be in general, in AACE.

Given the above instances of underspecification (and other similar instances), the paper makes it hard to understand and appreciate the precise problem and the proposed solution approach.

Seeing past the above problems, there seems to be an important but unstated assumption made in the technique that a "failure" is always attributed to a shortcoming of the LLM agent and not any other part of this multi-agent system (user LLM, deployment environment, etc.). Moreover, the notion of "failure" itself is not clearly defined in the paper.

The evaluation does not adequately justify the effectiveness of the proposed technique. The evolution of the policy considered for evaluation shows that only 3 new candidate constraints are accepted, and even they do not prevent the substantial failure cases according to the results.

The paper also uses "verification of a policy" to mean "testing the policy against a set of interactions", as opposed to the notion of verification that provides a guarantee of correctness against a specification. This usage of "Verification" creates an impression of a stronger guarantee that the technique does not actually provide.

Questions for authors

Can you elaborate on the language used to express the constraints used in AACE?

What is its expressivity, and can it handle all four types of the proposal mentioned in the problem formulation?

What does a failure mean in the context of AACE?

How complete is the testing procedure that validates a candidate rule?

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

Related papers

Discussion (0)

Sign in to comment.