acceptodds
Under review as a conference paper at ICLR 2027

From Local Errors to System Failures: Modeling and Benchmarking Error Propagation in Multi-Agent LLM Systems

Abstract

Errors in LLM-based multi-agent systems (MAS) can propagate through intermediate messages and compromise final answers, yet outcome-based evaluation obscures how agents amplify, attenuate, or correct them. We propose Response-Aware Error Propagation among Agents (REPA), a structural model that connects measured local responses to system-level error propagation. REPA separately estimates input assembly and agent processing across multiple error channels, while theoretical analysis relates local modeling errors to endpoint prediction error. Experiments span three Qwen base models, six MAS workflows, and ten knowledge and reasoning settings in ErrPropBench-10. REPA achieves a mean Spearman correlation of 0.84 across base models, compared with 0.47 for direct regression. Controlled experiments also demonstrate its utility for targeted recovery. REPA-guided RAG improves accuracy over no retrieval by an average of 19 percentage points with 16% additional tokens. By linking local agent behavior to system susceptibility, REPA provides an interpretable foundation for studying error propagation and developing targeted optimization in MAS.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Multi-Agent Systems, Large Language Models, Error Propagation

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_JntgOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes REPA, a framework for modeling error propagation in multi-agent LLM systems. REPA uses LLM judges to quantify semantic differences between reference and perturbed agent trajectories, then learns local linear models to capture how errors propagate through agent interactions. By composing these models along the workflow, REPA predicts system-level error propagation and guides targeted interventions, such as prompt refinement and RAG. Experiments across multiple LLMs, MAS workflows, and datasets show promising results in prediction correlation, fault localization, and system recovery.

Critical strengths
  1. Meaningful problem. The paper addresses an important problem of understanding error propagation in multi-agent LLM systems.

  2. Interpretable framework. REPA provides a structured approach to modeling local error responses and predicting system-level error propagation.

  3. Promising experimental results. The method achieves strong improvements in prediction correlation and demonstrates practical benefits in fault localization and targeted recovery.

Critical weaknesses
  1. Validity and cost of error measurement. The framework relies on LLM judges to quantify semantic differences between agent trajectories. It is unclear whether these numerical scores can faithfully capture complex error propagation. Additionally, the cost of LLM-based evaluation could be substantial for larger MAS workflows.

  2. Limited MAE improvement. Despite significant improvements in Spearman correlation, the reduction in MAE is relatively modest. This raises concerns about the framework's ability to accurately predict the magnitude of error propagation.

  3. Insufficient justification for linear modeling. The framework relies on linear transformations to model error propagation. Given the complexity of MAS interactions, it is unclear whether this assumption is sufficient. More empirical justification or comparisons with nonlinear alternatives would strengthen the paper.

Questions for authors
  1. Could the authors provide additional validation of the numerical error measurements and clarify the associated evaluation costs?

  2. Could the authors provide instance-level prediction results to better evaluate the accuracy of error magnitude estimation?

  3. Have the authors considered nonlinear alternatives to validate the linear modeling assumption?

  4. Could the authors clarify their reproducibility plans, including whether the code, datasets, and evaluation pipelines will be released?

3 comments

The authors thank the reviewer for your careful review and helpful suggestions.

1. Validity and cost of error measurement

REPA measures changes in error severity between original and perturbed runs using fixed scoring rules for each channel. In Fig. 2a, we form the joint reference by averaging A/B results within GPT and Claude, then averaging the two models. CCC measures agreement in both the pattern and size of these changes. These experiments support the observer's reliability and cost efficiency in error measurement. We also conduct a human validation study. Our human check covers 50 text pairs. Ten people each rate 10 pairs independently on a 0–4 scale, giving 100 pair-level ratings. We use the mean human change for each pair as the reference.

ComparisonFactHalluScFaithTaskIrr
GPT vs Claude (Fig. 2a)0.930.750.690.81
DS vs joint reference (Fig. 2a)0.870.600.330.46
Claude vs humans0.890.720.640.80
DS vs humans0.850.660.400.61

The first two rows report model agreement; the last two compare the observer measurements with human judgments. Against humans, DS is 0.04 below Claude on Fact and 0.06 below it on Hallu. Fig. 2a estimates observer scoring costs at USD 0.17 per 100 question pairs for DS and USD 0.38 for GLM. Both observers have the same Fact/Hallu CCC in Fig. 2a, so DS provides these agreement levels at a lower cost. For larger workflows, the cost of scoring paired runs grows with the number of nodes, rounds and the length of messages. Once the maps are fitted, forward prediction uses source measurements and numerical computation without LLM judging at each downstream node. Together, these results support our use of DS as the observer in REPA. DS combines agreement with human ratings on Fact and Hallu with a lower scoring cost.

2. MAE and instance-level predictions

As described in the paper, REPA estimates statistical error-propagation patterns across samples. It does not identify a separate propagation rule for each individual sample. MAE evaluates the predictions of this shared model on individual test records. For each question and test condition, we compute the absolute error in each of the four endpoint channels, then average across channels and records. Spearman measures the ranking of condition-level gains. A large ranking improvement therefore need not imply an equally large MAE reduction. REPA has the lowest MAE among the four methods in Table 1. Both Fact and Hallu also have lower MAE than other methods.

3. Why use local linear maps?

Agent behavior can be nonlinear. REPA models average error responses across paired samples. For a smooth local mean response, a first-order Taylor expansion gives a linear term plus higher-order effects. This motivates the local linear maps in Section 3.2. We fit these maps with nonnegative ridge regression. Their residuals include higher-order effects, measurement error and fitting error. The first-order form uses few parameters and lets us combine local maps along the workflow. The prediction results in Table 1 and fault-localization results in Fig. 4 support its practical use in the tested settings. Therefore, we use a first-order approximation to study error propagation in this work.

4. Code, data and evaluation pipeline

We will release the code, datasets and evaluation pipeline. It also provides clear instructions and keep expanding the datasets to cover more tasks and agent settings.

Thank you for the detailed response and additional validation. However, concerns regarding measurement reliability, limited MAE improvements, and the lack of nonlinear comparisons remain only partially addressed. Therefore, I will maintain my current rating.

Dear reviewer, The authors first thank the reviewer for your careful review. We add a nonlinear comparison on Q9 using polynomial regression with ridge regularization. We use a second-degree polynomial with squared and interaction terms.

MethodSpearman ↑MAE ↓
REPA-Polynomial0.800.28
REPA (ours)0.850.23

REPA achieves a higher correlation and a lower MAE in this comparison. One possible explanation is overfitting. The additional polynomial terms increase the number of parameters and may fit noise in the available training data, even with ridge regularization. These results support our use of the first-order approximation.

The authors also wish to clarify that relatively few studies directly connect local agent errors to system-level propagation predictions, as discussed in Appendix B. The current results provide empirical support for the proposed error propagation model in multi-agent systems. We will extend this work with additional experiments beyond the scope of the present conference paper.

Reviewer_88n6OpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper discusses errors in LLM-based multi-agent systems, specifically how these look at error propagation. It introduces REPA (Response-Aware Error Propagation among Agents) to show how local responses connect to error propagation by using a structural model that composes estimated agent responses on the topology. They introduce a benchmark to cover different knowledge domains and datasets and evaluate across multi-agent and single-agent baselines. They also have theoretical analysis on the error propagation, depending on perturbation placement, output selection and communication topology.

Critical strengths
  1. REPA uses a structural method that composes local responses according to the execution topology which is a good way to view the different components that could cause error propagation.
  2. REPA has good performance when compared with baselines across many different domains.
  3. Improvements in fault localization, targeted prompt restoration, and RAG placement outside of prediction.
Critical weaknesses
  1. Uses LLMs as candidate observers for cost purposes as well as proprietary reference models, but sometimes there is not very high CCC like 0.6 for hallucination. I'm wondering if there could be differences when different LLMs are used? If human validation could be provided for a subset, this could strengthen this evaluation as well.
  2. What are the impacts of input assembly vs agent processing along the executed workflow? It is not clear how the components contribute.
Questions for authors

Local error responses are approximated to linear maps but have you tried any nonlinear alternatives? An agent might not act in a linear way?

3 comments

The authors thank the reviewer for your careful review and helpful suggestions.

1. Observer reliability and human validation

Scores vary by observer and channel. Fig. 2a compares A/B prompts within each reference model, GPT with Claude, and five candidate observers with a joint reference. For each channel, we measure the change from the original to the perturbed output. We average the A/B results within GPT and Claude, then average the two models to form the joint reference. CCC measures agreement in both the pattern and size of these changes. These experiments support the observer's reliability and cost efficiency in error measurement. Following the reviewer's suggestion, we conduct a human validation study. Our human check covers 50 text pairs. Ten people each rate 10 pairs independently on a 0–4 scale, giving 100 pair-level ratings. We use the mean human change for each pair as the reference.

ComparisonFactHalluScFaithTaskIrr
GPT vs Claude (Fig. 2a)0.930.750.690.81
DS vs joint reference (Fig. 2a)0.870.600.330.46
Claude vs humans0.890.720.640.80
DS vs humans0.850.660.400.61

The first two rows report model agreement; the last two use human ratings. GPT and Claude also agree less on Hallu than on Fact. Against humans, DS is 0.04 below Claude on Fact and 0.06 below it on Hallu. Together, these results support our use of DS as the observer in REPA. DS combines agreement with human ratings on Fact and Hallu with a lower scoring cost.

2. Input assembly and agent processing

A and M describe different parts of the workflow. A maps changes in source inputs and upstream messages to changes in the input an agent actually receives. M maps that input change to the agent's output change. We fit them separately from paired observations. M also captures whether the agent passes on, reduces or amplifies an input error. Separating the stages helps us distinguish inherited changes from the current node's response. We will clarify the two stages and add a node-level example from saved runs in Section 3.2.

3. Linear approximation and nonlinear effects

Agent behavior can be nonlinear. REPA models average error responses across paired samples. For a smooth local mean response, a first-order Taylor expansion gives a linear term plus higher-order effects. This motivates the local linear maps in Section 3.2. We fit these maps with nonnegative ridge regression. Their residuals include higher-order effects, measurement error and fitting error. The first-order form uses few parameters and lets us combine local maps along the workflow. The prediction results in Table 1 and fault-localization results in Fig. 4 support its practical use in the tested settings. Therefore, we use a first-order approximation to study error propagation in this work.

4. Clarity of the method

The overview in Fig. 1 follows three steps: measure paired runs, fit the local A and M maps, and combine them along the executed workflow to predict the final response. In the final revion of our paper, we define the input and output of each step before the equations. We also explain the statistical approximation and distinguish fitting, forward prediction and diagnosis. We will make the explanation of the proposed method as clear and complete as possible.

Thank you for the clarification. While the Taylor expansion provides a theoretical motivation for local linear approximation, my concern is primarily empirical. Have the authors evaluated nonlinear alternatives, such as polynomial regression or MLPs, to determine whether the linear assumption adequately captures agent error propagation? Additionally, can you provide this node-level example?

The authors thank the reviewer for this helpful follow-up.

Following the reviewer's suggestion, we add a nonlinear comparison on Q9 using polynomial regression with ridge regularization. We use a second-degree polynomial with squared and interaction terms.

MethodSpearman ↑MAE ↓
REPA-Polynomial0.800.28
REPA (ours)0.850.23

REPA achieves a higher correlation and a lower MAE in this comparison. One possible explanation is overfitting. The additional polynomial terms increase the number of parameters and may fit noise in the available training data, even with ridge regularization. These results support our use of the first-order approximation in the tested Q9 setting. We also provide a concrete example from the Reasoner node in a saved Q9 sequential workflow. The question asks for Space Mountain's opening date. The upstream Verifier message contains the incorrect date “January 15, 1985” and a later correction to “January 15, 1975.” Input assembly combines this message with the original question. The Reasoner uses the correction and returns “January 15, 1975.” The observed Fact discrepancy between the clean and perturbed runs is 0.75 at the node input and 0.00 at its output. Here, A describes how the upstream message enters the assembled input. M describes the node's response to that input, including error reduction. This example illustrates why receiving an error does not necessarily mean passing it on. The maps remain statistical fits across paired samples.

Reviewer_eRs2OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies the error propagation in an LLM-based multi-agent system (MAS). This is an important research problem as the current agent system grows larger and more complex where an intermediate error can be propagated and amplified and finally compromise the system and final answer. The authors propose REPA, a structural model of error propagation in MAS based on continuous-valued responses across multiple error channels. It predicts error propagation from a pair of observations through input assembly which transforms into discrepancy channel maps and execution-aware composition and readout which summarizes the system level predictions. The authors also construct a benchmark, called ErrPropBench-10, covering seven domains and three multi-hop question-answering datasets, to evaluate the proposed REPA. Lastly, the authors also show that REPA can improve RAG performance by a large margin with a few additional tokens.

Critical strengths
  1. This paper studies a very important problem in multi-agent LLM systems about how local errors affect downstream agents. It is indeed very critical as the current agent system grows more complex and understanding system failures is important to improve the system. This paper proposes a reasonable and interpretable solution to address the problem. The proposed method can be used to help better understand and improve complex MAS systems.
  2. The proposed method is quite effective in predicting system responses from local agent behvavior, achieving 0.8 mean condition-level Spearman correlation compared to 0.47 of regression baseline.
  3. The RAG experiment is a nice addition, showing that the framework can be useful beyond error prediction and help improve real use cases.
Critical weaknesses
  1. Most of the experiments focus on the HotpotQA dataset. More diverse agent datasets and benchmarks would be helpful to better understand the capability of the proposed REPA. Besides the dataset, only Qwen models are used. More model families or API models would be more helpful and convincing.
  2. It would be interesting to see whether the approach can be applied in RL training for agents, for example, using the prediction results to provide training signals or support self-evolving agents.

In general, I like the interpretable way to explain how local behavior affects a multi-agent system. However, the dataset and models are somehow limited in the context of current agent frontier such as coding, CUA, etc.

Questions for authors

N/A

1 comment

The authors thank the reviewer for recognizing the value of our work.

1. Datasets and model families

Our main experiments use three Qwen models. ErrPropBench-10 covers HotpotQA, 2WikiMultiHopQA, MuSiQue and seven knowledge domains (e.g., Roman history, photosynthesis, and machine learning). Following the reviewer's suggestion, we provide additional results using the API model GPT-4o-mini. In this evaluation, REPA achieves a Spearman correlation of 0.86 and an MAE of 0.21. These values are close to those for the Qwen models in Table 1. They provide further support for REPA's robustness across the tested datasets and base models.

2. RL training and self-improvement

REPA can guide where and how to improve an agent system. In Fig. 4, it locates the faulty node in 46/60 cases, compared with 26/60 for the local difference baseline. In Fig. 5a, a single-node prompt repair restores accuracy from about 17% to 31%, close to the healthy system's 32%. Across the same six workflows using Q122, accuracy rises from 51.11% without RAG to 66.94% with REPA-RAG. These intervention results provide a practical starting point for agent training and self-improvement. For RL, we propose predicted downstream errors as an auxiliary training signal alongside the task reward. Diagnostic scores can prioritize weak nodes, while high-error conditions help select training examples. For RSI, the proposed loop runs the system, locates a weak node, updates its prompt, retrieval strategy or model setting, and evaluates the change. The loop includes refitting local response maps after agent updates. Our RL and RSI experiments remain in progress due to limited computing resources and the short rebuttal period. We will expand the discussion of how REPA can support RL training and agent self-improvement in the revised paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.