acceptodds
Under review as a conference paper at ICLR 2027

Belief Is Not Authority: Warranted State Transitions for Black-Box Language Models

Abstract

Guardrails check rules and state trackers infer latent needs, but neither decides when an uncertain belief may authorize a different action. We introduce RAVR-W (Retrieval-Augmented Verification and Repair with Warrants), a runtime-assurance layer for frozen black-box language models built on authority-flow semantics. Any discovered factor may enter an open shadow state, but it influences prediction, recommendation, or execution only through distinct, earlier-epoch receipts for each tier; retrieval supplies evidence, never authority. A typed transition warrant switches away from a known fallback only with fallback-relative effect evidence, binds every argument to its evidential lineage, and commits external effects only after a read-back witness. We prove authority non-interference (unauthorized factors cannot change the decision), tier non-escalation, fallback-relative switch validity, witness-closed commit, and preservation of learned preconditions, and we introduce blind causal transition induction, which grows the executable state schema from an empty graph without letting any episode authorize itself. Across 14 benchmarks in counseling, negotiation, and stateful tool use, against more than a dozen matched baselines, RAVR-W wins wherever evidence licenses a switch. It improves calibration on all six core counseling tasks with clustered intervals below zero (296,796 held-out instances; median Brier reduction 14.7%, up to 30.8%) and raises τ-bench state balanced accuracy from 0.585 to 0.726. On ToolSandbox it improves success in all four model/provider configurations (+0.0391 [0.0239, 0.0531]), on an independent 105-scenario set, and by +0.154 [0.050, 0.262] on 60 further scenarios; the complete live kernel with learned precondition closure improves a bare agent by +0.111 [0.052, 0.177] on 184 scenarios, and induced preconditions lift held-out tool-call success from 0.40 to 1.00. On Gaia2 it solves 5/6 workflows versus 0/6 for the default agent. Where evidence is insufficient, it returns the exact fallback (148/148 on MHDialog), turning refusal into an auditable outcome rather than a silent failure.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
trustworthy LLMs, runtime assurance, LLM agents, action authorization, formal guarantees, causal precondition learning, tool use, safe policy improvement

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_rr4qOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper presents a mechanism for separating self-reported model assessments from what is actually permitted by a policy. They provide results showing the mechanism improves results on several benchmarks across multiple tasks.

Critical strengths
  1. The core idea is well supported and relates to the much discussed topic of self-assessment by language models.
  2. The presented evaluation results are thorough with relevant ablations.
Critical weaknesses
  1. The presentation of the method is slightly disorganized and includes more jargon than necessary. In addition, the three theorems are more linked to safety claims than the performance claims of the paper.
  2. Many of the reported performance improvements reported are fairly small. For example, the Brier reductions in Table 2 are around 0.01 and the ToolSandbox gain in Table 3 is under 0.04 improvement.
Questions for authors
  1. What is the token cost overhead for this mechanism?
  2. How do you estimate the latency would change when used in more complex schemas than the ones evaluated here?
  3. How were the thresholds in Section 4.2 chosen?
Reviewer_PpPpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes RAVR-W (Retrieval-Augmented Verification and Repair with Warrants), a runtime assurance framework for frozen black-box LLMs that separates inferred beliefs from the authority to make decisions or execute actions. RAVR-W introduces a tiered authority mechanism that separates information gathering and prediction from decision-making and execution, requiring additional evidence before information can be used to recommend or execute actions. Experimental results demonstrate improvements in prediction calibration and tool execution, while reverting to specified fallback actions when sufficient evidence is unavailable.

Critical strengths
  • The problem of distinguishing uncertain model beliefs from the authority to make decisions and execute actions is important and well-motivated.
  • The paper provides a comprehensive framework that integrates tiered authority management, action verification, and causal transition induction. The experiments cover a broad range of benchmarks and include comparisons with multiple baselines.
Critical weaknesses
  • The proposed framework is quite complex, involving multiple authority tiers, evidence receipts, transition warrants, and verification mechanisms. It is unclear how easily the framework can be generalized to new domains, particularly when domain-specific rules, fallback actions, and evidence requirements must be specified.
  • Despite the broad experimental coverage, the evaluations are somewhat fragmented across different benchmarks, with different components and authority tiers evaluated separately. It is unclear to what extent the results demonstrate the effectiveness of the complete framework in realistic end-to-end applications.
Questions for authors

How easily can RAVR-W be adapted to new domains, given its reliance on domain-specific rules, fallback actions, evidence requirements, and execution contracts? How much manual effort is required to configure these components?

1 comment

Thank you for the authors' response. The paper would benefit from another round of revisions to more thoroughly address the feedback. At this stage, I will maintain my original assessment.

Reviewer_gqJsOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
1
Soundness
2(+1)
Read review
Paper summary

From my understanding, the paper looks at the problem of deciding when a language model should be allowed to take an action based on the belief it has formed form observations about the current state of the environment. They propose a framework that has different levels of authority for information that the agent uses. The agent has predefined fallback actions, an in order for the system to allow the agent to change its action from the predefined fallback, the systems requires "evidence". the paper talks about several theoretical results about the proposed framework, but the formulation and proposed methodology are very difficult to follow, and I am not confident the technical contributions are faithfully communicated.

Critical strengths

The problem is important to ensure the agents take actions based on reliable information.

Critical weaknesses

Paper is very very hard to follows. Starting from the abstract and introduction, the authors introduce a bunch of technical terms without sufficiently explaining any of their meaning or purpose. I found it very difficult to understand how the different components relate or fit together, and what the central contribution is. The preliminaries are also very heavy in unnecessary notation, and several quantities are introduced without sufficient intuition. the authors need to provide a simple running example that clearly explains the formulation.

Questions for authors

The theoretical results seem to just follow from definitions. I don't understand the significance of them. Theorem 1 states that unauthorized factors cannot affect the decision, but the decision kernel is defined to only use authorized information. There is no technical insights inside these theorems.

All the gaurantees rely on very heavy meaningless assumptions.

empirical evaluations are also very hard to interpret, theres a large collection of benchmarks covering very different tasks, and each use different evaluation metrics, baselines and settings. its not clear what the overall results establish at all.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision 2: restructured guarantees, lighter notation, and one correction

We thank all reviewers for the discussion. We have uploaded a second revision (all changes in blue) that acts on the remaining points.

  1. Guarantees, each with what it adds (rr4q, gqJs). Non-interference is now presented as a property that holds by construction; its value is that it can be audited (5,000/5,000 perturbation trials), not that it is deep. The main text keeps three results with content, each followed by a short "What it adds" paragraph:
  • Theorem 1 (no self-authorization): requiring evidence from an earlier epoch caps false authorization at the test level alpha_0 however adaptive the proposer is, versus up to 2^B alpha_0 when the proposer sees B bits of the certification data. This matches our measurements: 84.7% vs 4.7% false passes in simulation, and 100% vs 0.5% placebo promotion.
  • Theorem 2 (no worse than the fallback): links safety to performance. RAVR-W's expected utility is at least the fallback's plus m * P(switch) minus alpha * R, so every warranted switch adds at least the entry margin, and the only way to underperform the fallback is certificate failure.
  • Theorem 3 (safe schema growth and witnessed commit): rules out the two characteristic agent failures, acting on an unmet precondition and reporting success that did not happen, while letting learned preconditions grow from data. Routine statements and their proofs moved to the appendix.
  1. Lighter preliminaries (gqJs). The 14-component transition tuple and state decomposition moved to a new appendix section; the main text now defines a warrant in plain language as six checks, alongside the glossary (Table 1), the one-sentence summary, and the running example in the introduction.

  2. Complexity (PpPp). The layers are separable: the calibration studies use only the prediction tier, the ToolSandbox executor uses only the execution layer, and the live agent combines them. Wrapping an agent requires a tool-call adapter (about 310 lines for ToolSandbox) around the domain-independent kernel.

  3. Correction to our earlier response to rr4q. We stated that one threshold configuration is shared by all domains. More precisely, a default configuration (entry margin 0, harm budget 0, at least 20 groups, overlap at least 0.8) applies unless a study's freeze manifest fixes a task-specific contract before execution; the ToolSandbox kernel study uses 60 calibration groups per task, full overlap, and a harm budget of 0.10. No threshold was tuned on outcomes, and the domain packs contain no numeric thresholds. The revised text states this accurately.

We are grateful for the constructive feedback and are happy to clarify anything before the discussion closes.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.