acceptodds
Under review as a conference paper at ICLR 2027

Learning from Success: Online Early Failure Detection for Agentic Systems

Abstract

LLM-based agent systems are increasingly deployed for complex tasks, yet their execution success rates remain limited, making early failure detection critical for raising alerts or intervening in erroneous executions before completion. Existing studies primarily focus on post-hoc failure attribution, which requires complete execution trajectories and forgoes the opportunity to intervene when the task is still running, while the few efforts on early failure detection typically rely on supervised learning from failure trajectories with step-level error annotations. In practical deployments, however, annotating a sufficient number of failure trajectories for supervised failure detection is costly, whereas successful trajectories are naturally accumulated through routine execution. This motivates an important yet underexplored setting: safe-only early failure detection, where an effective monitor identifies failures online by learning exclusively from successful execution trajectories. We introduce a simple framework for this setting, grounded in the insight that successful trajectories provide complementary evidence about both how an execution should evolve and whether the current transition is consistent with previously observed behaviors. Accordingly, our framework combines a safe-dynamics predictor that models the temporal evolution of successful trajectories with a non-parametric safe transition memory that evaluates contextual consistency against historical safe transitions. When a limited number of unsafe trajectories become available, our framework naturally incorporates them as failure memory through causal retrieval and safe-unsafe transition contrast, without additional retraining or fine-tuning. Extensive experiments demonstrate that our framework achieves strong online failure detection performance under the safe-only setting and can further leverage limited failure supervision to approach the performance of fully supervised detectors while requiring fewer annotated failure trajectories.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
online failure detection, multi-agent systems, only safe trajectories

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_CuuJOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2(+1)
Presentation
2
Soundness
2(+1)
Read review
Paper summary

This paper studies safe-only online early failure detection for LLM-based agentic systems, aiming to identify execution failures before tasks are completed without requiring extensive failure trajectory annotations. The authors propose Learning from Success (LFS), which combines a parametric safe-dynamics predictor with a non-parametric memory of successful transitions to detect deviations from successful execution patterns. When limited failed trajectories are available, LFS incorporates decisive failure transitions as a contrastive memory without retraining the predictor. Experiments on AgentProcessBench and AFTraj-2K demonstrate competitive detection and localization performance, with additional improvements from limited failure supervision.

Critical strengths
  1. The proposed framework is intuitive and computationally lightweight, combining temporal prediction with context-based retrieval to capture complementary signals of execution anomalies.
  2. The results show promising performance in the safe-only setting and demonstrate that a relatively small failure memory can substantially improve detection and localization on certain domains.
Critical weaknesses
  1. The methodological novelty appears limited, as next-step prediction, nearest-neighbor transition retrieval, and memory-based failure comparison are established techniques whose combination may not constitute a sufficiently distinctive contribution.
  2. The transition retrieval mechanism primarily uses the immediately preceding step embedding, leaving it unclear whether it adequately captures the longer-range, task-dependent context emphasized in the paper's motivation.
  3. The evaluation metrics focus partly on ranking annotated decisive failure steps against successful execution steps, which may not fully characterize practical online monitoring performance, particularly false alarms and missed detections.
Questions for authors
  1. How does LFS differ fundamentally from existing safe-only trajectory modeling and anomaly detection methods, particularly OAT and ATS, beyond combining a predictor with transition retrieval?

  2. Why does the transition memory retrieve neighbors using only the immediately preceding step embedding rather than the task query and a longer execution-history representation? Have the authors evaluated these alternatives?

  3. How much does each component contribute relative to simpler baselines, including the dynamics predictor alone, nearest-neighbor transition anomaly detection alone, and a straightforward safe-versus-failure memory comparison?

  4. Can the authors provide additional online evaluation results measuring false-alarm rates, missed detections, and detection timing under distribution shifts or varying task and trajectory-length distributions?

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

The paper proposes LFS for online failure detection in agentic systems. It combines a self-supervised next-step embedding predictor with successful-transition retrieval to detect deviations from successful execution, and optionally incorporates failure examples without retraining.

Critical strengths
  1. The method supports online monitoring through prediction and retrieval, with low additional overhead after embedding extraction.
  2. The self-supervised objective uses successful execution logs without requiring step-level error annotations, reducing dependence on failure supervision.
Critical weaknesses
  1. The evaluation provides limited evidence that LFS distinguishes context-dependent errors from unfamiliar but valid behavior. Cross-trajectory score comparisons do not isolate this distinction, particularly for unseen tasks.
  2. Although the threshold is calibrated on successful validation trajectories, the false-alarm rate on successful test trajectories is not reported. This leaves the risk of interrupting valid executions, especially on unseen tasks.
Questions for authors
  1. For AFTraj-2K, are the safe training, validation, and test sets disjoint by task/query? What evidence distinguishes error sensitivity from sensitivity to task or trajectory familiarity?
  2. At the calibrated P95 threshold, what fraction of successful test trajectories trigger an alarm? Are such results available for tasks absent from training and calibration?
Reviewer_taamOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies safe-only online early failure detection for LLM agent systems: a monitor trained only on successful trajectories must decide, from each trajectory prefix, whether to raise an alarm. It proposes LFS, which encodes steps with a frozen Qwen2.5-3B-Instruct encoder and averages the discrepancy from an LSTM safe-dynamics predictor with the distance to a non-parametric safe transition memory (Eq. 6, 12, 13). Alarms fire above a 95th-percentile threshold calibrated on safe validation trajectories. When annotated failures are available, their decisive transitions are added as a contrastive memory without retraining (Eq. 16–23). Validation on AFTraj-2K and AgentProcessBench compares LFS with prompting/ICL and adapted ATS, OAT and Safe-FSM monitors on AUROC, AUPRC, Exact-F1, Step-Acc and ASS, and compares failure-assisted LFS (5–50% failure support) with AgentForesight-7B and other supervised detectors.

Critical strengths
  1. The safe-only early-detection setting is practical, and the prefix-only constraint is enforced consistently (Sec. 3, App. B.5)
  2. The method is simple and cheap: a frozen encoder, an LSTM next-step predictor and a transition memory, fused by a parameter-free average (Eq. 5, 12, 13). Its overhead is small next to encoding (Fig. 2)
  3. Safe-only gains on AFTraj-2K are large: average AUROC is 93.15 vs. 84.64 and AUPRC 68.36 vs. 49.16 against the best baseline, OAT (Table 1)
  4. Failure memory needs no retraining. With 10% or more support, Math Exact-F1 exceeds AgentForesight-7B (Table 12)
Critical weaknesses
  1. False alarms on safe trajectories are not measured. Exact-F1, Step-Acc and ASS are computed only over unsafe trajectories (Eq. 24–28), and test-set FPR is never reported.
  2. AUROC/AUPRC overstate discrimination. Positives are decisive steps t*, while negatives are only safe-trajectory steps; pre-t* steps of failing trajectories are excluded (B.3).
  3. The safe-only score has no ablation. "Safe dynamics" and "Transition memory" are defined as baselines in Table 5 but have no rows in Table 1, so the complementarity claim is untested.
  4. The supervised comparison is not like-for-like. AgentForesight numbers are copied from the original evaluation, and prompted auditors are labelled "100%" failed support (Table 12 vs. B.2). Failure-assisted LFS still trails overall (61.76 vs. 66.44 Exact-F1; 0.869 vs. 0.59 ASS).
  5. Failure memory sometimes hurts, without discussion. At 5% support, Agentic Exact-F1 falls from 14.45 to 10.63 and Coding ASS rises from 0.566 to 1.200. Coding ASS never recovers to its safe-only level (Table 12).
Questions for authors
  1. What is the trajectory-level false-alarm rate on the safe test trajectories of both benchmarks, under A_safe and A_assist?
  2. What do A_dyn alone and A_mem alone score on Table 1's metrics?
  3. Can AgentForesight-7B, or a supervised probe on your encoder, be evaluated under your exact protocol and test split?
1 comment

Dear Reviewer taam,

Thank you very much for your time and effort throughout the review process.

We sincerely appreciate your valuable suggestions and comments. Regarding all your concerns, we have provided detailed response and supplemented new experimental results, including:

1. Additional FAR evaluations on both benchmarks.

2. Ablation studies on the "Transition memory" and the "Safe Dynamics" component.

3. An analysis on Coding and Agentic domains of the non-monotonic performance improvements with respect to the failure memory.

4. A fully supervised comparison between AgentForesight-7B and a fine-tuned Qwen2.5-3B classifier.

We have also provided detailed explanations addressing your other questions.

We fully understand that you may have a busy schedule. Nevertheless, we would greatly appreciate it if you could take a moment to review our responses and share any further feedback you may have.

Thank you again for your time and consideration.

Sincerely,

Authors

Comments on the paper

1 comment

Dear Reviewers,

Thank you very much for your time and effort throughout the review discussion.

We would like to kindly ask whether you could consider our response, in which we provide detailed replies point by point for your concerns.

We would sincerely appreciate your feedback you may have. Your insights are highly valuable and will help us improve our work.

Thank you again for your careful consideration.

Sincerely,

Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.