acceptodds
Under review as a conference paper at ICLR 2027

Failure Attribution with Stepwise Uncertainty in Multi Agent Systems

Abstract

Failure attribution in agent systems aims to identify the steps that critically contribute to task failure. Existing approaches either rely on large language models (LLMs) as judges, incurring substantial inference costs, or require additional data and training to build dedicated attribution models. We propose a training-free method for failure attribution that identifies critical failure steps using the outputs of a fixed, lightweight observer model. We find that decisive failure steps leave a measurable signature in the observer's continuous uncertainty. Conventional binary readouts that collapse predictions into binary outcomes discard relative uncertainty across steps, but the proposed method directly uses the observer model's pre-threshold uncertainty as a continuous attribution score, identifying the first step whose uncertainty enters a high-uncertainty regime as the critical failure step. It requires no modification to the observer model or its inference procedure, and introduces no additional training data, step-level failure labels, or auxiliary evaluation models. Recovering the continuous information enables fine-grained, step-level failure attribution at minimal additional cost. Experiments on the Who & When benchmark demonstrate that this attribution signal persists across multiple unsupervised localization rules, and that the proposed method outperforms closed-model LLM-as-a-Judge baselines by up to 10.3%p, while avoiding $2.11-$14.44 in frontier-model API costs per corpus without requiring any additional model calls.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Multi-Agent Systems, Failure Attribution

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

The paper studies step-level failure attribution in multi-agent traces using the continuous uncertainty of a fixed observer instead of its binary judgment. For each trajectory prefix, the observer provides a probability of correctness; the method converts it to uncertainty, applies Otsu thresholding, and selects the first high-uncertainty step. It requires no attribution-specific training or labeled failure steps at inference. Experiments use both Who&When subsets, reporting agent- and step-level exact-match accuracy with and without ground truth. Seven open observers and seven closed judges are compared, with same-observer binary readouts, alternative uncertainty measures, and localization rules. The main result is 38.1% step accuracy on Algorithm-Generated traces without ground truth, versus 27.8% for the strongest closed judge. The paper also reports positive continuous-over-binary gains and higher mean failure-step AUROC.

Critical strengths
  1. The central comparison is well controlled: Section 4.3 uses the same observer, prompt, and inference call for continuous and binary readouts, isolating whether retaining confidence values helps rather than conflating the effect with a different model or extra sampling.
  2. Tables 2 through 4 and Figure 6 vary observer, uncertainty estimator, and localization rule; continuous scores improve 51 of 56 observer-setting pairs.
  3. Section 5 explicitly limits evaluation to Who&When and notes prefix-wise inference costs on long traces.
  4. The method avoids labeled attribution data, which is useful for an inexpensive first-pass diagnostic.
Critical weaknesses
  1. Performance is benchmark dependent: on Algorithm-Generated data, step accuracy is 38.1%, but on Hand-Crafted trajectories it is 23.6%, below the strongest closed judge at 27.0% (Table 1). How well does uncertainty localize failures beyond Who&When?
  2. Absolute step localization remains weak. In Table 3, mean continuous-score AUROC is only 57.3 on Algorithm-Generated and 58.7 on Hand-Crafted data, and several individual observers are near chance or below it. Otsu converts these modest ranking gains into an exact step, but there is no analysis of how often errors are adjacent-step confusions versus distant misattributions.
  3. Baseline currency is a concern for an ICLR 2027 submission. The related-work section discusses Who&When, ECHO, and CHIEF, but Table 1 does not compare against recent structured or causal attribution systems. These are not exact uncertainty baselines, but they are pertinent system-level comparisons for the headline claim of improved attribution.
  4. The uncertainty is a proxy for observer confidence, not a calibrated probability that a step caused failure. There is no calibration or breakdown by trace length and step type. Otsu also assumes a useful two-regime distribution, but the robustness study does not show when this assumption holds.
  5. Scores are exact-match point estimates without confidence intervals or repeated-run variation, leaving gains of only a few points (Table 1) difficult to interpret.
Questions for authors
  1. Can you report results on an additional attribution benchmark or system family, including a comparison with structured methods such as CHIEF and EMFA?
  2. Can you break down step errors by distance from the annotated failure step, trace length, and step type, and report confidence intervals over traces?
  3. What evidence supports Otsu's two-regime assumption, and how often does the selected high-uncertainty regime begin before, at, or after the annotated decisive step?
  4. Can you report calibration of the observer uncertainty against step correctness, or explain why calibration is unnecessary for the thresholding rule?
3 comments
AuthorsOpenReview ↗

Official Comment by Authors (1/2)

Thank you for highlighting our controlled comparison of continuous and binary readouts and the practical value of our training-free approach. Below, we address your concerns one by one.


W1, Q1. Benchmark Dependence and Generalization

The lower accuracy on Hand-Crafted trajectories is partly associated with their greater length. Most Hand-Crafted logs exceed 10 steps, where all methods, including closed-source judges, achieve below 14% step-level accuracy. Nevertheless, our best observer outperforms six of seven closed-source judges on Hand-Crafted logs. Beyond Who&When, our method achieves the highest step-level accuracy on two additional benchmarks, Who&When Pro and TraceElephant, outperforming the strongest baseline by 3.9pp and the binary readout by 12.9pp on the pooled evaluation.

BenchmarkRandomBinary readoutBest baselineOurs
Who&When Pro21.836.447.8 (Binary Search)51.2
TraceElephant5.99.69.6 (All-at-Once)11.5
Pooled19.432.441.4 (Binary Search)45.4

Under the same observer, our method also outperforms CHIEF and OAT overall (see W3). We could not evaluate EMFA because its implementation is not publicly available.


W2. Error Distance Analysis

Localization errors are predominantly near misses on Algorithm-Generated logs but more distant on Hand-Crafted logs. Among incorrect predictions, 53% fall within ±2 steps of the gold step on Algorithm-Generated logs, compared with only 16% on Hand-Crafted logs.

Errors within ±2 steps (%)Hand-CraftedAlgorithm-Generated
Ours1653
Binary readout1247

This difference is associated with trajectory length: all Algorithm-Generated logs contain at most 10 steps, whereas most Hand-Crafted logs exceed 10 steps. Unlike AUROC, which evaluates ranking across all steps, our method localizes the onset of elevated failure-evidence scores. Thus, moderate AUROC can coexist with near-miss localization errors (Appendix B.6).


W3. Comparison with Recent Attribution Systems

Our method achieves the highest overall step-level accuracy (30.4%) under the same observer (Qwen3.6-35B-A3B), outperforming Scope Delineation (+9.4pp), CHIEF (+10.5pp), and OAT (+13.6pp). On Algorithm-Generated logs, it also surpasses CHIEF by +12.7pp and OAT by +17.8pp.

MethodHC AgentHC StepAG AgentAG StepAll Step
All-at-Once29.15.554.017.513.8
Binary Search47.310.930.214.313.3
Scope Delineation40.012.736.524.621.0
CHIEF43.67.360.325.419.9
OAT-8.7-20.316.8
Ours41.812.750.038.130.4

Unlike CHIEF's multi-stage generation and retrieval or OAT's training on successful trajectories, our method requires no training and only a single-token probability readout per step. We have added these comparisons to Table 2. We also tried to compare with EMFA, but its implementation is not publicly available.


W4, Q3. Validity of the Two-Regime Assumption

The two-regime assumption is empirically supported by a clear increase in failure-evidence scores after the gold step. Mean scores rise from 0.41 to 0.62 on Algorithm-Generated logs and from 0.38 to 0.61 on Hand-Crafted logs, with higher post-failure averages in 84% and 93% of logs, respectively. Regarding Q3, predicted onsets tend to precede the gold step, particularly on longer Hand-Crafted trajectories, as shown below.

Predicted onset (%)Before goldMatches goldAfter gold
Hand-Crafted63.612.723.6
Algorithm-Generated36.538.125.4

Change-point detection mitigates early-onset errors by identifying shifts in mean scores rather than isolated threshold crossings, improving Hand-Crafted accuracy from 12.7% to 16.4% (Appendix B.9).

AuthorsOpenReview ↗

Official Comment by Authors (2/2)

Q2. Analysis by Trace Length and Step Type

Step-level accuracy declines with trajectory length across all methods, including closed-source judges. Our method achieves the highest exact-match and ±2-step accuracy on 6–10-step trajectories, which account for the majority of the benchmark.

Length (steps)OursBinary readoutScope Delin.Gemini 3.5 Flash LiteGPT-4.1
1–571.4 / 10057.1 / 10042.9 / 71.442.9 / 85.771.4 / 100
6–1035.0 / 68.326.8 / 60.224.4 / 56.926.8 / 61.820.3 / 57.7
11+2.0 / 17.67.8 / 17.69.8 / 27.57.8 / 21.613.7 / 19.6

(Values indicate exact-match / ±2-step accuracy (%).)

Since Who&When lacks step-type annotations, we analyze failures by the responsible agent as a proxy. WebSurfer-related failures are the most challenging, with all methods achieving at most 13% accuracy. Trace-level 95% confidence intervals are provided in Appendix B.6.


Q4. Calibration of the Observer Score

Our method uses observer scores as relative failure evidence rather than calibrated probabilities of failure. Otsu determines a threshold from the score distribution within each trajectory, without requiring probability calibration. Although raw scores are poorly calibrated (ECE: 0.543 on HC and 0.449 on AG), Platt scaling substantially reduces ECE to 0.001 and 0.017, respectively. Thus, probability calibration is not a prerequisite for our training-free localization approach (Appendix B.8).


W5. Confidence Intervals and Run-to-Run Variance

Repeated experiments confirm consistent improvements on Algorithm-Generated logs. Across four runs, our method outperforms the binary readout in step-level accuracy in every run, achieving 35.9% vs. 31.7% (+4.2pp) with less than half the standard deviation (1.6 vs. 3.6).

ReadoutAG stepAG agentHC stepHC agent
Ours (Otsu)35.9 ± 1.646.0 ± 1.311.8 ± 3.140.0 ± 2.6
Ours (map)32.1 ± 0.544.4 ± 0.611.8 ± 2.342.7 ± 1.0
Binary readout31.7 ± 3.643.3 ± 2.613.2 ± 4.045.5 ± 3.0

On Hand-Crafted logs, performance varies more across runs, with no consistent advantage over binary readout. On Algorithm-Generated logs, paired bootstrap tests further confirm significant improvements over binary readout (+10.3pp, p=0.004), all same-observer baselines, and GPT-4.1 (p≤0.022) in the reported single-run comparison. Paired bootstrap results are reported in Appendix B.1, and run-to-run variance in Appendix B.2

AuthorsOpenReview ↗

Brief Summary of Our Response and Gentle Reminder

Thank you again for your thoughtful and constructive feedback, and for recognizing the value of our controlled comparison between continuous and binary readouts and the practical benefits of our training-free approach.


Regarding your concerns:

Benchmark dependence and generalization (W1, Q1): We extended our evaluation to Who&When Pro and TraceElephant, where our method achieves the highest step-level accuracy on both benchmarks, outperforming the strongest baseline by 3.9pp and binary readout by 12.9pp on the pooled evaluation.

Error distance analysis (W2): We analyzed localization errors and found that 53% of incorrect predictions on Algorithm-Generated trajectories fall within ±2 steps of the gold step, compared with 16% on longer Hand-Crafted trajectories, clarifying the relationship between localization accuracy and trajectory length.

Comparison with recent attribution systems (W3): We demonstrated that our method achieves the highest overall step-level accuracy (30.4%) under the same observer, outperforming Scope Delineation (+9.4pp), CHIEF (+10.5pp), and OAT (+13.6pp), without additional training.

Validity of the two-regime assumption (W4, Q3): We provided empirical evidence of increased failure-evidence scores after the gold step in 84–93% of trajectories and showed that change-point detection mitigates early-onset errors, improving Hand-Crafted accuracy from 12.7% to 16.4%.

Analysis by trace length and step type (Q2): We demonstrated that localization accuracy declines with trajectory length across all methods, while our approach achieves the highest exact-match and ±2-step accuracy on 6–10-step trajectories. We also provided agent-level failure analysis and confidence intervals.

Calibration of the observer score (Q4): We clarified that our method relies on relative failure evidence rather than calibrated probabilities, and showed that probability calibration is not required for our trajectory-adaptive thresholding approach.

Confidence intervals and run-to-run variance (W5): We conducted four repeated runs and confirmed consistent improvements over binary readout on Algorithm-Generated trajectories (35.9% vs. 31.7%), while acknowledging the higher variance and lack of consistent improvement on Hand-Crafted trajectories.


As today is the final day of the rebuttal period, we hope that these additional experiments and clarifications address your concerns, and would sincerely appreciate it if you could reconsider your assessment of our manuscript in light of these updates.

Reviewer_oJ49OpenReview ↗
Rating
3
Confidence
2
Informativeness
3(+1)
Presentation
3
Soundness
3(+1)
Read review
Paper summary

This paper introduces a simple but efficient framework to identify the earliest error step in a multi-step agentic setting. It adopts P(True) to estimate the uncertainty score of the error at each step and uses the score to find a division of two groups. The experiment compares the proposed method (continuous) against frontier model evaluation (binary) and open-weight models (also binary), and the continuous method achieves a large gain in various settings.

Critical strengths
  • The proposed method is very simple but effective and easy to implement in practice. Especially, it doesn't depend on the closed-source model API and achieves better results compared to the frontier model API judge.

  • It studies a practical debugging question (where did the failure begin?), and it needs no model training or new attribution labels.

  • The paper is well written and easy to follow; I really enjoyed reading the paper. :)

Critical weaknesses
  • The evaluation is conducted on a relatively small dataset, which harms the paper's reliability.
  • The idea of measuring an uncertainty score continuously instead of binary is not novel; some similar works, like Deepthink (ICLR2026), adopt a similar idea to measure the possible failure of generation. Although this paper is in a multi-step agentic setting, I think a more comprehensive evaluation task is needed to validate the results.
Questions for authors
  • It seems that the gain is largely from the continuous score for reach step estimation. Some frontier APIs from the frontier model also return token probabilities, so it would be possible to design a simple continuous method for the frontier method as well and see if this brings improvements compared to the binary judge.

  • In Table 1, there is a big gap (-10) compared to the closed LLM at the automated agent level. Do you have any insight into this? It looks suspicious that this is the opposite of the other settings.

  • I would be happy to improve the score if the author addressed my concerns and questions.

3 comments

Thank you for highlighting the simplicity and practical value of our method. Below, we address your concerns one by one.


W1. Evaluation Scale

To address concerns about evaluation scale, we extend our experiments to two additional benchmarks with decisive-step annotations: Who&When Pro and TraceElephant. Under the same observer and without GT, our method achieves the highest step-level accuracy on both benchmarks, outperforming the strongest baseline by +3.9pp and the binary readout by +12.9pp on the pooled evaluation.

BenchmarkRandomBinary readoutBest baselineOurs
Who&When Pro21.836.447.8 (Binary Search)51.2
TraceElephant5.99.69.6 (All-at-Once)11.5
Pooled19.432.441.4 (Binary Search)45.4

Per-system results are provided in Appendix B.3.


W2. Novelty Relative to Continuous Confidence

Our key distinction lies in whose confidence is measured and how it is used. DeepConf uses a model's own confidence to filter its reasoning, whereas our method uses an independent observer's confidence to localize failures in multi-agent trajectories. This distinction is important because self-correction without external feedback is often unreliable (Huang et al., 2024), and agents accessed through APIs may not expose their logits. We have clarified this distinction in Section 2.


Q1. Continuous Readout for Frontier Judges

Continuous readout also improves failure localization with frontier LLM judges. Using the same Step-by-Step prompts, extracting P(Yes) instead of binary verdicts consistently improves step-level accuracy for both GPT-4o-mini and GPT-4.1.

JudgeAG BinaryAG ContinuousHC BinaryHC Continuous
GPT-4o-mini15.119.8 (+4.8)14.516.4 (+1.8)
GPT-4.120.627.8 (+7.1)16.421.8 (+5.5)

These results demonstrate that the benefit of continuous failure-evidence scoring extends beyond open-weight models to frontier LLM judges (Appendix B.4).


Q2. Agent-Level Gap on Algorithm-Generated Logs

The agent-level gap arises from how agent attribution depends on step localization. Our method assigns the agent associated with the predicted step. Since agents frequently alternate between steps in Algorithm-Generated logs, even an off-by-one localization error can lead to incorrect agent attribution. When the predicted step is correct, our agent accuracy reaches 97.8%.

Agent accuracy (%)Step correctStep wrong
Ours97.817.3
All-at-Once63.654.0

In contrast, All-at-Once predicts the agent independently of the step, allowing correct agent attribution even when step localization fails. This difference explains the observed agent-level gap (Appendix B.7).

Thank you for the detailed responses and new experiments for my questions and for other reviewers. Overall, I think it is an interesting and practical contribution. I am not very familiar with this research problem; it seems that the new baseline comparison results were provided in response to reviewer nbR6, which would strengthen the soundness. I would expect the paper to be more informative after the next revision. I revised my scores in these two aspects to support the paper.

Thank you for your positive feedback and for taking the time to review our additional experiments and responses to other reviewers. We sincerely appreciate your reconsideration of the soundness and informativeness scores and your support for our work.

Reviewer_nbR6OpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper argues that finding which agent and which step broke a multi-agent run can require either expensive frontier LLM judges or training custom evaluator models. This paper proposes a lightweight, training-free alternative using small open-weight models. The method prompts an LLM judge at each step of a trajectory prefix to ask if the current action contains an error, extracts the raw token probabilities for the yes and no answers instead of taking a hard greedy decision, and then applies Otsu thresholding over the sequence of scores to flag the first step that jumps into a high-error regime. Experiments on the Who and When benchmark show that keeping the soft token probabilities works much better than using hard binary outputs from the same model and beats several closed-source LLM judges under the step-by-step evaluation setup.

Critical strengths
  1. Using soft tokens as a ranking signal is very interesting idea.
Critical weaknesses
  1. The baselines in Table 1 are too weak, and the paper did not compare with methods it cites in related works. Section 2 talks about full-trajectory judges, binary search, Scope Delineation, AgenTracer, and AEGIS, but none of them show up in the experiments. Instead, the main table only compares against the basic step-by-step prompt from Who&When, which prior work already showed is pretty weak. It also leaves out other recent ways of doing attribution, like anomaly scoring from successful runs [1] or the fault localization step in DoVer [2]. Right now, we only know that soft probabilities beat hard thresholding on a naive prompt, not how the method stacks up against real baselines.
  2. What the paper calls uncertainty is really just error probability, and the Otsu rule looks brittle for long trajectories. In Eq. 6, the score is the normalized probability assigned to the error token. When the model is 100% sure a step is wrong, the score is 1.0, whereas real model uncertainty would peak at 0.5 when the model is torn between yes and no. Framing this as uncertainty is confusing. On top of that, scoring every prefix one by one means token costs grow quadratically with trajectory length, and because Otsu always forces a two-group split, picking the very first step over the threshold means one early hiccup in a long run will throw off the prediction.

[1] Tracing Agentic Failure from the Flow of Success.

[2] DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems.

Questions for authors

The work itself is valuable in my opinion. However, the lack of vertical comparison against other established failure attribution baselines weakens the contribution significantly.

2 comments

Thank you for recognizing the value of soft token probabilities as a ranking signal. Below, we address your concerns one by one.


W1, Q1. Comparison with Existing Failure Attribution Methods

Our method achieves the highest overall step-level accuracy (30.4%) under the same observer (Qwen3.6-35B-A3B), outperforming Scope Delineation (+9.4pp), CHIEF (+10.5pp), and OAT (+13.6pp).

MethodHC AgentHC StepAG AgentAG StepAll Step
All-at-Once29.15.554.017.513.8
Binary Search47.310.930.214.313.3
Scope Delineation40.012.736.524.621.0
CHIEF43.67.360.325.419.9
OAT-8.7-20.316.8
Ours41.812.750.038.130.4

Unlike CHIEF's multi-stage generation and retrieval or OAT's training on successful trajectories, our method requires no training and only a single-token probability readout per step. We have added these comparisons to Table 2. We also tried to compare with additional methods, but these comparisons were infeasible: AgenTracer and AEGIS have not released their trained attribution models, ECHO lacks a public implementation, and DoVer requires intervention and re-execution, which Who&When's recorded logs do not support.


W2-1. Clarifying the Uncertainty Score

Our score UtU_t follows the P(True) formulation (Kadavath et al., 2022), where 1−P(True)1-P(\mathrm{True}) measures confidence that a step is erroneous (Vashurin et al., 2025). Unlike entropy, which peaks at 0.5, UtU_t increases with the observer's confidence in an error. To clarify this distinction, we now refer to UtU_t as a failure-evidence score in Section 3.3 and compare it with entropy-based measures in Figure 6.


W2-2. Computational Cost of Prefix Scoring

Prefix caching reduces the number of newly processed tokens from quadratic to approximately linear growth with trajectory length by reusing shared prefixes. As shown below, only about 500 new tokens are processed per step, even for long trajectories.

Length (steps)No cache (tokens/log)With cache (tokens/log)New tokens/step
1–710,0322,757445
8–1424,8665,495566
15–2990,73310,498510
30+843,55737,826498

For trajectories with 30+ steps, caching reduces processed tokens by 95.5% and scoring time from 110s to 40s on a single GPU (Section 4.5, Appendix B.5).


W2-3. Robustness to Early Spikes

Our method does not depend on Otsu; the continuous failure-evidence scores can be used with different decision rules. As shown below, all continuous-score rules outperform binary readout on Algorithm-Generated trajectories. For longer Hand-Crafted trajectories, change-point detection achieves the highest accuracy (16.4%) by identifying shifts in mean scores rather than relying on the first threshold crossing, making it less sensitive to isolated early spikes.

RuleHand-CraftedAlgorithm-Generated
Binary readout10.927.8
Otsu12.738.1
GMM-29.133.3
Percentile9.132.5
Change-point16.430.2

These results show that the benefit of continuous failure-evidence scoring is not specific to Otsu. We retain Otsu as the default and recommend change-point detection for longer trajectories (Section 3.4, Appendix B.9).

AuthorsOpenReview ↗

Brief Summary of Our Response and Gentle Reminder

Thank you again for your thoughtful and constructive feedback, and for recognizing the value of soft token probabilities as a ranking signal.


Regarding your concerns:

Comparison with existing failure attribution methods (W1, Q1): We demonstrated that our method achieves the highest overall step-level accuracy (30.4%) under the same observer, outperforming Scope Delineation (+9.4pp), CHIEF (+10.5pp), and OAT (+13.6pp), without additional training.

Clarification of the uncertainty score (W2-1): We clarified that our score follows the P(True) formulation and represents failure evidence rather than predictive entropy, and updated the terminology and comparisons accordingly.

Computational cost of prefix scoring (W2-2): We demonstrated that prefix caching reduces newly processed token growth from quadratic to approximately linear, reducing processed tokens by 95.5% and scoring time from 110s to 40s for trajectories with 30+ steps.

Robustness to early spikes (W2-3): We confirmed that the benefits of continuous failure-evidence scoring are not specific to Otsu, and showed that change-point detection improves accuracy on longer Hand-Crafted trajectories (16.4%) by reducing sensitivity to isolated early spikes.


As today is the final day of the rebuttal period, we hope that these additional experiments and clarifications address your concerns, and would sincerely appreciate it if you could reconsider your assessment of our manuscript in light of these updates.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.