acceptodds
Under review as a conference paper at ICLR 2027

From One Monitor to One Joint: Tractable Probabilistic Inference for Agent Safety with Tensor Networks

Abstract

A safety monitor typically estimates the risk of a completed agent trajectory, but runtime safety must reason as the available evidence changes: observations may be missing, actions assessed before execution, or future violations anticipated from a prefix. Existing methods can answer these queries individually, but do not generally ensure that their answers are probabilistically consistent. The missing ingredient is a shared joint distribution. We therefore propose TTMonitor, which extends a black-box completed-trajectory monitor into a joint distribution over trajectories and safety outcomes, so that different runtime queries become conditionals of a unified probabilistic model. TTMonitor combines an estimated trajectory distribution with monitor distillation and represents the resulting joint as a tensor train, enabling tractable marginalization over various types of evidence. Experiments on InjecAgent and AgentDojo show that TTMonitor maintains high fidelity to the original monitor while ensuring cross-query consistency. Exact tensor-network contraction further avoids the accuracy–latency tradeoff of completion sampling and remains practical relative to direct neural prediction.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
agent safety, tractable probabilistic modeling, tensor networks

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_7oqyOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper propose a joint distribution to solve this question: Given a complete trajectory, how likely is it that it was unsafe? but at runtime you need to ask four different questions. Existing work answer these separately, so the four answers need not to be mutually coherent.

Thus the paper build the missing joint distribution called TTMonitor, p(x,y)= b(x)*w_{theta(y(c=x))}. where w_{theta, y} is the frozen black box teacher monitor's soft score and b(x) is a trajectory distributino estimated from source traces. The main experiments uses beta=1. It uses the first order markov model of agent trajectories, a nonnegative tensor-train model distilled from a frozen teacher monitor, the tensor contractino to condition on observed events and marginalise unobserved events.

On the InjecAgent and Agent Doji, explicit joint models show zero detected cross-query incompatibility, and TT monitor is far more faithful to the teacher than the Markov joints baselines. Query aware distillation cuts runtime fidelity error from 0.028 to 0.007. Separately, exact tensor contraction is faster than the tested monte carlo completion sampling methods and avoids their sampling error.

Critical strengths

1, The problem is good, clear, and practially relevant. The inconsitency exmple is compelling: if every possible resolutino f an unknown event has a risk above 0.6, the unsolved risk cannot be 0.2 under the same joint model.

  1. The construction is elegant and tractable: The nonnegativity and stochasticity constraints make the student a valid conditional distribution. Combining it with a trajectory distributino produces a joint model whose conditionals satisfy probability identifies by construction.

  2. Useful evidence for query aware distillation: On InjectAgent, query MAE decreases from 0.028 with trace-only training to 0.007 with the dual objective, while competed-trajectory MAE remains 0.004 while qury mae moves from 0.028 to 0.007) is a clean separation of what the architecture buys from what the objective buys.

Critical weaknesses
  1. The fidelity comparison is confounded by unmatched training objectives. A major weakeness is that the headline comparison suses tragets that TT monitor was explicitly trained to match, while the baselines were not.

The student is normalized over y for each trajectory. With beta=1. Query MAE is therefore not measuring runtime-inference quality at all- it is the competed-trajetory distillation error, reaveraged under the method's own conditional law. And l post is the empirical version of the evaluation metric(N=8 sampels in training, M=8192 at evaluation).Nothing here is undisclosed and model selection did not touch the test data. And the consequence will be the metric cannot distinguish has a good joint from was trained on this loss.

The author should distill the same frozen teacher into a GRU and a Transformer using the same L comp/L post objective and the same cached completions, and rerun both fidelity and the compatibility audit.

  1. The motivating decision problem is never evaluated. The experiments reports the Brier, MAE and LP feasibility, and never return to the decision. The closest prior work does exactly this (ProGuard reports advance warning time and unsafe-behaviour reduction at preserved task completion), so the bar exists.

  2. b(x) is load-bearing and unvalidated Every marginalization query is an average weighted by b(x) ^bta. b is a first order marlov chain over coarse structural symbols with additive smoothing- and the paper reports no held-out likelihood, perplexity, or comparison against a higher order or neural trajectory model. beta is introduced as a design knob in eq e and then fixed to 1 everywhere; an gamma=0.9 is fixed, unablated, and content-independent.

  3. AgentDojo breaks the paper central promise, and the paper under reports it. On InjecAgent, Complete MAE =0.004 promise kept, On AgentDojo is 0.063 and 0.048, with Query MAE0.074/0.055. A systematic 6 point deviation in probability is not preserving the teacher, it is a differnt monitor. And the text says only that the achievvable fidelity depends on the teacher being approximated, which read as an understatement.

Questions for authors
  1. One decision-level experiment: does inconsistency actually flip interventions, and does the joint fix it. what do distillation matched baselines d? Per A.6 the pooled predictors receive labl supervision only. what are query MAE, joint repair and incompatibility rate for 1)a GRU and ii) a transformer distilled from the same frozen teacher with the same L comp/L post objective and the same cached completiions? This is a control that separates you need an explicit joint from you need the query aware loss.

2: Is the runtime fidelity metric measuring runtime inference, or remeasuring distillation error? (a) Does Query MAE reduce to the completed trajectory distillation gap reweighted by your own trajectory law, and b) is L_post the empirical form of the evaluation metric itself? could you report Query MAE against a reference mu' that your model does not use?

  1. Does inconsistency change decisions? On the audited query groups where a baseline is incompatible, at what rate does the incomptibilty flip the intervention decision at a realistic threshold? and for q future, can you report detection earliness against a prefix trained monitor? a low number is still a result and would be worth stating plainly
Reviewer_k7ygOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper introduces TTMonitor, an application of the Tensor Train (TT) idea to agent safety monitoring, where it uses TT as the primary modeling tool to learn from a teacher monitor to make predictions about agent trajectories with missing information or under query-based settings. The resulting TTMonitor method shows strong prediction adherence to the teacher model it is trained on, and meanwhile exhibits consistency by construction between the various predictions that it makes.

Critical strengths
  • The proposed approach leverages Tensor Train, resulting in a principled and efficient solution that can be leveraged at agent runtime to make timely decisions about future agent actions by designing loss functions that distill only from full-trajectory assessments of the teacher at training time.
  • Empirical results show superior adherence to the teacher, consistency between the predictions that TTMonitor makes, and lower latency compared to various baselines that can be supported by CPU.
Critical weaknesses
  • The method makes an assumption about the well-calibratedness of the teacher, which is unclear if it holds well in practice. This seems to require that the probabilities assigned by the teacher to different trajectories are already well-calibrated across these trajectories -- this is an unchallenged assumption and I don't think the experiments or discussion cover the case if the teacher is not calibrated.
  • It is unclear how this method handles complexities like large event / state spaces and sparse observations. The construction of the tensors seem to heavily rely on count statistics (Eqn. 5-8), which can itself quickly become intractable / difficult to estimate when the state space is complex or exponential.
  • The writing is a bit difficult to follow. The motivating example itself is unclear, and in the methodology section it is not immediately clear which variables are observational vs trainable. Some technical details also remain opaque. For instance, it's unclear what the dimensionality rr stands for or how it is chosen on Line 191, and what is the size of the various vectors and matrices in Eqn 5-8. The paper could also use an introduction of the two benchmark datasets in the main body to help the reader better understand the results.
Questions for authors
  • If the teacher were not calibrated in the first place, do the results actually show that TTMonitor is just overfitting to arbitrarily badly calibrated teachers better than baseline methods? Does the TT formulation offer any regularization in this case?
Reviewer_ULvFOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes TTMonitor, which combines a trajectory model with a tensor train approximation of an existing safety monitor. The resulting joint distribution handles predictions from completed trajectories, missing events, candidate actions, and unfinished prefixes. The experiments evaluate agreement with the original monitor, consistency across related queries, outcome prediction, and inference cost.

Critical strengths
  • The problem setup is practical. Runtime safety queries are usually handled by separate predictors, and treating them as conditionals of a single joint distribution gives them shared semantics while making consistency explicit.

  • The approach is simple and practical. It extends a frozen monitor without retraining it, and exact inference handles missing, candidate, and future evidence without sampling possible completions.

Critical weaknesses
  • The fidelity comparison does not isolate the benefit of the tensor train setup. TTMonitor is trained to match teacher outputs on sampled completions, while the direct predictors and joint baselines are not distilled from the teacher. Their higher teacher error might simply reflect this target mismatch. The runtime comparison samples the distilled student rather than the original teacher, so it does not establish the latency gain over answering the target queries directly with the teacher.

  • The estimated runtime risks depend heavily on how unobserved events are modeled. A first order Markov model weights possible completions, and a geometric prior sets the remaining sequence length. Matching the teacher under this fitted model does not prove that the risks are accurate when real trajectories have longer dependencies or stop based on their current state.

  • The method requires a fixed set of event types. In practice, whether an action is unsafe often depends on the instruction and tool arguments. Adding these details expands the event set, which increases inference cost and leaves less data per Markov transition. The evaluation uses monitors over discrete event encodings, so applicability to monitors that process raw instructions and tool arguments is not established.

  • The practical benefit of cross query consistency is not clearly established. Direct neural predictors yield better or comparable outcome predictions and show very little inconsistency. The paper motivates consistency using contradictory intervention decisions, but it does not show that observed inconsistencies actually alter decisions in practice. It remains unclear if the drop in prediction accuracy is worth the tradeoff.

  • The positioning relative to prior work is unclear. Methods like GeLaTo [1] and Ctrl G [2] also distill a large model into a tractable representation to compute future probabilities from a prefix. Explaining how TTMonitor differs would clarify its technical contribution. In addition, discussing runtime monitors for rule enforcement [3], policy verification [4], and failure prediction from prefixes [5] would help show what joint consistency adds in practice.

[1] Tractable Control for Autoregressive Language Generation, ICML 2023.

[2] Adaptable Logical Control for Large Language Models, NeurIPS 2024.

[3] AgentSpec: Customizable Runtime Enforcement for Safe and Reliable LLM Agents, ICSE 2026.

[4] ShieldAgent: Shielding Agents via Verifiable Safety Policy Reasoning, ICML 2025.

[5] AgentForesight: Online Auditing for Early Failure Prediction in Multi Agent Systems, arXiv 2026.

Questions for authors
  1. Is there a comparison against direct predictors trained on the same teacher targets? What runtime cost is needed to estimate the target queries directly with the original teacher?

  2. Is there an evaluation showing how well the Markov model and stopping prior fit held out executions? How sensitive are the runtime risks to these modeling choices?

  3. How would cost and fidelity change with a finer event set? How would distillation work for a monitor that processes raw text directly?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.