Artificial Intelligence · arXiv:2609.01360 v1 · submitted 1 Sep 2026 · 35 references
EDGE: Error Dependency Graph-Guided Multi-Error Attribution in Multi-Agent LLM Systems
The paper says
“Large language model (LLM) agent failures often contain multiple related errors rather than a single mistake. Existing attribution methods usually identify a responsible agent, step, or root cause, but do not explicitly model dependency between errors.”
more of the abstract
We introduce EDGE, an Error Dependency Graph-guided multi-Error attribution framework. EDGE constructs an error dependency graph from observed error events and validates a reliable causal subset through counterfactual rollout. The inference graph guides a two-stage LLM-as-judge detector for error attribution, and the intervention-validated subgraph provides a more reliable basis for explanation and repair analysis. Experiments on TRAIL and MAST show that EDGE improves category-level multi-error attribution across most evaluated models and settings. Experiments with adapted Who&When-style prompts show that the graph helps across prompting strategies. These results suggest that dependency structure is a useful diagnostic prior for agent failures beyond isolated root-cause prediction.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 9 Sep 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up, with conditions. EDGE presents a practical dependency-aware approach to multi-error attribution in agent traces, but its causal validation pipeline warrants scrutiny. This line is wrong
4 major objections · 7 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 35 references say.
The paper says, at Section 4.1, Limitations
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
All reported results are single-run point estimates without any error bars or confidence intervals. Given stochasticity in LLM inference (the paper states 'default sampling temperature'), the reported differences (e.g., +14.17 on TRAIL-GAIA) may not be statistically significant. The authors should provide error bars (e.g., over multiple seeds, bootstrap resampling of traces, or paired tests) to support the claim that EDGE consistently improves over baseline.
more of objection 1
This is a load-bearing methodological issue because the manuscript draws conclusions about cross-model and cross-benchmark effectiveness from unquantified point estimates.
The paper says, at Section 4.1, Appendix A.5
The review did not quote this passage.
Pith says: Objection 2 of 4 · major
The inference-graph threshold τ is selected by maximizing model-averaged weighted-F1 over the full evaluation set (τ=0.35 for TRAIL, τ=0.50 for MAST). While Appendix A.5 reports a sensitivity sweep on the full data, there is no indication that τ was tuned on a held-out validation set.
more of objection 2
The held-out 80/20 split (Table 3) uses the same τ as the full-data experiment, so it does not validate the threshold selection procedure. This raises a risk of overfitting the evaluation set. The authors should either perform τ selection on a separate validation split or provide a more rigorous argument (e.g., stability across splits) that the chosen τ is general.
The paper says, at Section 3.3, Appendix A.2, Table 6
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
The intervention-based validation pipeline for MAST relies entirely on GPT-4o for both the simulated counterfactual rollouts (because AG2 traces are recorded and not re-executable) and the Judge-B effect evaluation. Table 6 shows that 70% of Judge-B labels on MAST are 'not_observable', meaning the effect cannot be assessed.
more of objection 3
The remaining signal (23% 'disappeared') comes from the same model family that likely produced the original traces. This creates a circularity risk: the validated graph GV may encode the simulation model's own expected error propagation rather than an independent causal fact. The paper's interpretation of GV as a 'causal anchor' for explanation and repair analysis (Section 4.4) depends on the validity of this pipeline. The authors should test the stability of GV under different evaluation models (e.g., a different LLM for Judge B) or provide additional evidence that the validated edges reflect genuine causal propagation rather than model self-consistency.
The paper says, at Table 1, MAST results
The review did not quote this passage.
Pith says: Objection 4 of 4 · major
On MAST, EDGE improves weighted-F1 on 7 of 8 backbones, but accuracy decreases on 7 of 8 backbones (ranging from -0.33 to -3.74), while recall increases substantially. This pattern suggests that EDGE may be making the detector more permissive rather than genuinely improving attribution quality.
more of objection 4
The paper acknowledges this trade-off (footnote 2 mentions an always-'no' classifier achieves 69.5% accuracy) but does not analyze whether the recall gains come at the cost of false positives that harm precision. The authors should report precision-recall curves or F2 scores to assess whether the gains in recall are meaningful, and discuss whether the accuracy drop is acceptable for the intended use case (diagnosis aid).
7 minor objections
The paper says, at Abstract
Pith says: minor objection 1 of 7
The abstract states 'Experiments on TRAIL and MAST show that EDGE improves category-level multi-error attribution across most evaluated models and settings.' This is accurate but would benefit from a brief caveat about the MAST accuracy trade-off.
The paper says, at Figure 2 caption
Pith says: minor objection 2 of 7
The caption is dense and uses notation (GS, GC, GV, Gτ) without fully defining them in the caption itself. Consider adding a brief parenthetical explanation for each graph symbol.
show 5 more
The paper says, at Equation (6)
Pith says: minor objection 3 of 7
The risk-difference is defined as an expectation over T in T_ver. But the quantity inside the expectation is a boolean difference. This is effectively the proportion of traces where B is present in the original but absent in the counterfactual. Clarify that this is a sample average, not a formal expectation over a distribution.
The paper says, at Appendix A.2
Pith says: minor objection 4 of 7
The patch library JSON listings (TRAIL and MAST) occupy many pages. While the detail is appreciated for reproducibility, consider moving the full JSON to a supplementary file and summarizing key design principles (patch side, slot schema, forbidden actions) in the appendix text.
The paper says, at Table 6
Pith says: minor objection 5 of 7
The row 'Judge-B disappeared' shows 57.5% for TRAIL and 23% for MAST. But the text says 'The labels emerged and strengthened contribute negatively'. The calculation of Δ(A→B) only uses disappeared? Clarify how the other labels are handled in the equation (6) computation.
The paper says, at Section 4.5, Table 16
Pith says: minor objection 6 of 7
The case study table (Table 16) uses metric deltas for single traces. These are not meaningful for generalizing, and the narrative seems anecdotal. Consider either removing or reframing as illustrative examples with a disclaimer.
The paper says, at Limitations
Pith says: minor objection 7 of 7
The limitations paragraph is thorough, but it does not mention the single-run point estimate issue. Add a sentence acknowledging that all numbers are single-run and that future work should include uncertainty quantification.
What would settle it
Replace GPT-4o with a different frontier model (e.g., GPT-5 or o3-mini) for the MAST counterfactual rollout simulation and Judge B evaluation, keeping all other pipeline components fixed. Recompute the validated graph GV and compute the Jaccard similarity of its edge set with the original GV. If similarity < 0.5, the causal interpretation is model-dependent and the claim that GV captures true propagation is weakened. If similarity > 0.8, the validation is robust to the evaluation model choice.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 9 Sep 2026
- Reader holds up, with conditions
- Skeptic agrees
- Editor wrote the verdict above
- Referee major revision: 4 major objections, 7 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
EDGE addresses a real problem: multi-error attribution in multi-agent LLM traces, where errors cascade. The core idea—building a category-level dependency graph from observed event statistics and using it to guide a two-stage detector—is sensible and produces clear F1 gains across both TRAIL and MAST benchmarks. The paper does several things well: it adapts Suppes/CAPRI progression modeling to agent errors, adds a counterfactual validation step for causal edges, and includes useful ablations (random-edge controls, a held-out split, static-graph comparison). The improvements are consistent across many backbones, and the limitations section is unusually candid.
The main weakness is the causal validation on MAST, as the stress test flags. The entire pipeline for MAST—patch generation, Judge A, Judge B, and the simulated rollout—relies on GPT-4o. With 70% of Judge B labels being not_observable, the remaining signal may reflect model self-consistency rather than true causal propagation. The paper does not test whether the validated graph is stable under a different evaluator model. This matters because the claim that dependency structure provides causal insight beyond correlation rests on this validation. Additionally, all reported numbers are single-run point estimates without confidence intervals, and the threshold τ is tuned on the evaluation set (though the held-out split partially mitigates this). Localization gains are small, especially on SWE-Bench, suggesting the graph mainly helps category detection rather than pinpointing spans.
Who is this for? Researchers working on LLM agent trace analysis, failure diagnosis, or debugging tools. The framework is taxonomy-agnostic and transferable, making it useful as a starting point for further work. It deserves peer review because it tackles a practical problem with a clear method and generally solid experiments, but the authors should be asked to address the causal validation concern, ideally by using a different model for evaluation or by providing human-validated event positions, and to report error bars from repeated runs.
Recommendation: send to review, but flag the causal validation as a required revision point.
The strongest objection
The MAST causal validation pipeline (Section 3.3, Appendix A.2, Table 6) uses GPT-4o for both the simulated counterfactual rollout (since AG2 traces are recorded and not re-executable) and the Judge B effect evaluation. Table 6 shows that 70% of Judge-B labels on MAST are 'not_observable', meaning the effect cannot be assessed. The 23% 'disappeared' labels that drive edge retention come from the same model family that likely produced the original traces (multi-agent LLM systems also commonly based on GPT-4-family models). This creates a circularity: the validated graph GV may encode the simulation model's own expected error behavior rather than an independent causal fact. The paper's central interpretation of GV as a 'causal anchor' for explanation and repair analysis (Section 4.4) is load-bearing on the validity of this pipeline. While Gτ augments GV with observational edges, the paper claims that 'dependency structure is a useful diagnostic prior beyond correlation', and that claim depends on the causal interpretation of the validated subgraph. The Limitations note that 'repair verification, effect evaluation, and MAST event annotation rely on LLM judgments, which may introduce noise', but do not test whether GV is stable under different evaluation models. Without such a test, the causal edges on MAST may be artifacts of the evaluation model rather than genuine error propagation.
The reader's weakest assumption correctly identifies the central vulnerability: the causal validation pipeline (especially for MAST) relies on LLM-generated judgments and simulated rollouts, creating a risk that GV encodes annotation artifacts rather than true causal propagation. The paper's own Limitations acknowledge this. However, the paper provides multiple independent supports for the broader claim: (1) Held-out 80/20 split (Table 3) shows graph-guided gains transfer to unseen traces, which cannot be explained by overfitting. (2) Random-edge controls (Figure 3) show that graph structure matters, not just edge count. (3) Who&When-style adaptation (Figure 4, Tables 13-14) shows gains across different prompting strategies. (4) The paper separates prediction (Gτ) from explanation (GV), so the main detection gain does not depend exclusively on the causal interpretation. Therefore, even if GV were partially unreliable, the central claim remains supported. The concrete test (cross-model validation) would strengthen or challenge the causal interpretation but does not invalidate the detection gains or the overall contribution. Hence the verdict should remain CONDITIONAL, as the reader proposed.
The referee's report
If the central claim holds—that error-category dependency graphs learned from trace corpora and partially validated by counterfactual intervention provide a useful prior for multi-error attribution—then EDGE offers a principled way to move beyond isolated root-cause prediction toward dependency-aware diagnosis in LLM agent systems. The framework is taxonomy-agnostic and is demonstrated on two benchmarks with different annotation granularities and error taxonomies. The separation of an observational inference graph (for detection coverage) from an intervention-validated subgraph (for explanation and repair analysis) is well motivated. However, the strength of the evidence is tempered by the lack of uncertainty quantification, the potential for overfitting during graph-threshold selection, and the reliance on LLM judgments for causal validation, especially on the MAST benchmark.
Pith's reading of the paper (a machine's summary, not the authors')
Large language model agents often produce multiple related errors rather than a single mistake, but existing attribution methods treat each error in isolation. This paper introduces EDGE, a framework that constructs an error dependency graph from observed error events across a corpus of failed agent traces. EDGE validates causal edges through counterfactual rollouts and uses the resulting graph to guide a two-stage LLM-as-judge detector. Experiments on TRAIL and MAST benchmarks show that EDGE improves category-level multi-error attribution across most evaluated models and settings, with gains of up to 14 points in weighted F1. The dependency structure provides a useful prior for diagnosing cascading failures and supports more explainable debugging and repair prioritization.
What carries the argument. The central object is the error dependency graph, built in two phases: an observational correlation graph GS from temporal precedence and probability raising, a causal-candidate graph GC via CAPRI pruning, and a validated causal graph GV from counterfactual rollout with controlled direct effects. The inference graph Gτ unions GV with score-filtered edges from GS. The key operational mechanism is the two-stage detector that first predicts an initial error set, then uses the trace-specific edge subset Rτ(T) (derived from the graph and Stage-1 predictions) to propose and verify downstream errors in a second LLM call. This progressive conditioning turns the corpus-level graph into a per-trace diagnostic prior.
Core claim. EDGE claims that modeling the propagation structure between error categories, rather than attributing each error independently, improves multi-error detection in LLM agent traces. The framework builds an observational correlation graph via Suppes screening and CAPRI pruning, then validates a causal subgraph through controlled counterfactual interventions that patch source errors and measure downstream effects. The thresholded inference graph, which unions validated causal edges with score-filtered observational edges, is consumed in a two-stage detector: first predicting initial errors, then verifying graph-suggested downstream errors. On TRAIL, EDGE raises weighted F1 over baseline by up to 14.17 points on GAIA and 12.31 points on SWE-Bench; on MAST, gains reach 9.37 points. The graph guidance transfers across prompting strategies and holds up under an 80/20 held-out split, confirming that the signal comes from the structure of learned dependencies rather than dataset artifacts.
The premise it rests on. The causal graph relies on LLM-generated judgments for counterfactual validation and step annotation, which could introduce systematic bias if the LLM judges are unreliable.
If the paper is right (Pith's inferences, not the authors')
System developers can prioritize repairing upstream errors that have many validated downstream effects, reducing the number of fixes needed to address cascading failures.
Multi-error attribution becomes tractable for long-context agent traces where exhaustive step-by-step search is infeasible, because the graph narrows the search space to plausible propagation paths.
The validated causal subgraph GV serves as an inspectable artifact for post-hoc explanation and can be consumed by automated repair loops, potentially enabling self-healing agent systems.
The graph-construction pipeline is taxonomy-agnostic, meaning it can be applied to any error classification scheme with observed events, generalizing beyond TRAIL and MAST taxonomies.
The dependency graph may reveal common failure archetypes across different agent architectures—for example, that planning errors consistently trigger execution errors—which could inform agent design improvements.
Because the validation step relies on LLM-generated patches and judgments, the graph likely encodes biases of the LLM used for those tasks; using stronger or more diverse judge models could either sharpen or weaken the causal interpretation.
The method implicitly assumes that error categories are stable across traces; if new error types emerge during deployment, the graph would need periodic recomputation, which could limit practical adoption in rapidly changing environments.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Large language model (LLM) agent failures often contain multiple related errors rather than a single mistake. Existing attribution methods usually identify a responsible agent, step, or root cause, but do not explicitly model dependency between errors. We introduce EDGE, an Error Dependency Graph-guided multi-Error attribution framework. EDGE constructs an error dependency graph from observed error events and validates a reliable causal subset through counterfactual rollout. The inference graph guides a two-stage LLM-as-judge detector for error attribution, and the intervention-validated subgraph provides a more reliable basis for explanation and repair analysis. Experiments on TRAIL and MAST show that EDGE improves category-level multi-error attribution across most evaluated models and settings. Experiments with adapted Who&When-style prompts show that the graph helps across prompting strategies. These results suggest that dependency structure is a useful diagnostic prior for agent failures beyond isolated root-cause prediction.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-09-23
The authors' main claim, as the classifier picked it out 0.97
“These results suggest that dependency structure is a useful diagnostic prior for agent failures beyond isolated root-cause prediction.”
How the main claim is stated: hedged possibility 1.00
What the abstract itself states, as probabilities: claims first 0.09 · states limitation 0.09 · human participants 0.05 · controlled comparison 0.56 · randomized assignment 0.02 · states quantitative result 0.06 · compares with prior results 0.75 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Large language model (LLM) agent failures often contain multiple related errors rather than a single mistake. | background 1.00 | 0.10 | 0.03 | asserted 0.98 |
| Existing attribution methods usually identify a responsible agent, step, or root cause, but do not explicitly model dependency between errors. | gap 0.99 | 0.03 | 0.03 | asserted 1.00 |
| We introduce EDGE, an Error Dependency Graph-guided multi-Error attribution framework. | contribution 1.00 | 0.91 | 0.06 | asserted 0.99 |
| EDGE constructs an error dependency graph from observed error events and validates a reliable causal subset through counterfactual rollout. | method 0.98 | 0.92 | 0.06 | asserted 1.00 |
| The inference graph guides a two-stage LLM-as-judge detector for error attribution, and the intervention-validated subgraph provides a more reliable basis for explanation and repair analysis. | method 0.99 | 0.82 | 0.13 | asserted 0.99 |
| Experiments on TRAIL and MAST show that EDGE improves category-level multi-error attribution across most evaluated models and settings. | result 1.00 | 0.96 | 0.25 | asserted 1.00 |
| Experiments with adapted Who&When-style prompts show that the graph helps across prompting strategies. | result 1.00 | 0.93 | 0.06 | asserted 1.00 |
| These results suggest that dependency structure is a useful diagnostic prior for agent failures beyond isolated root-cause prediction. | interpretation 1.00 | 0.69 | 0.78 | hedged possibility 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 1.00
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-09-23: 1 request, 6,094 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (6)
Showing 5 of the paper's 6 figures. Figures are the authors', reproduced from the paper.
References (35)
- [1] Trail: Trace reasoning and agentic issue localization , author=. arXiv preprint arXiv:2505.08638 , year= on PithPith's reviewarXiv
- [2] Why do multi-agent llm systems fail? , author=. Advances in Neural Information Processing Systems , volume= on Pith
- [3] Zhang, Shaokun and Yin, Ming and Zhang, Jieyu and Liu, Jiale and Han, Zhiguang and Zhang, Jingyang and Li, Beibin and Wang, Chi and Wang, Huazheng and Chen, Yiran and Wu, Qingyun , title =. Proceedings of the 42nd International Conference on Machine Learning , articleno =. 202... on Pith
- [4] Where did it all go wrong? A hierarchical look into multi-agent error attribution , author=. arXiv preprint arXiv:2510.04886 , year= on PitharXiv
- [5] Zhu, Chenyang and Hong, Spencer and Wu, Jingyu and Chawla, Kushal and Tang, Yuhui and Yin, Youbing and Wolfe, Nathan and Babinsky, Erin and Liu, Daben. RAFFLES : Reasoning-based Attribution of Faults for LLM Systems. Proceedings of the 19th Conference of the E uropean Chapter ... on Pithdoi:10.18653/v1/2026.eacl-long.359
- [6] Automatic Failure Attribution and Critical Step Prediction Method for Multi-Agent Systems Based on Causal Inference , author=. arXiv preprint arXiv:2509.08682 , year= on PithPith's reviewarXiv
- [7] Abduct, act, predict: Scaffolding causal inference for automated failure attribution in multi-agent systems , author=. arXiv preprint arXiv:2509.10401 , year= on PitharXiv
- [8] From Flat Logs to Causal Graphs: Hierarchical Failure Attribution for LLM-based Multi-Agent Systems , author=. arXiv preprint arXiv:2602.23701 , year= on PitharXiv
- [9] AgentRx: Diagnosing AI Agent Failures from Execution Trajectories , author=. arXiv preprint arXiv:2602.02475 , year= on PithPith's reviewarXiv
- [10] A Probabilistic Theory of Causality , author =. 1970 , publisher = on Pith
- [11] CAPRI: efficient inference of cancer progression models from cross-sectional data , author=. Bioinformatics , volume=. 2015 , publisher= on Pith
- [12] TRONCO: an R package for the inference of cancer progression models from heterogeneous genomic data , author =. Bioinformatics , volume =. 2016 , doi = on Pith
- [13] Where llm agents fail and how they can learn from failures , author=. arXiv preprint arXiv:2509.25370 , year= on PitharXiv
- [14] Wang, Junjie and Wang, Yawen and Chen, Mengzhuo and Xie, Xiaofei and Chen, Chunyang and Mu, Fangwen and Liu, Zhe and Wang, Qing , title =. IEEE Transactions on Software Engineering , year =. doi:10.1109/TSE.2026.3717765 , url = on PitharXiv
- [15] THPs: Topological Hawkes Processes for Learning Causal Structure on Event Sequences , author=. IEEE Transactions on Neural Networks and Learning Systems , year= on Pith
- [16] Qiao, Jie and Cai, Ruichu and Wu, Siyu and Xiang, Yu and Zhang, Keli and Hao, Zhifeng , title =. 2023 , isbn =. doi:10.24963/ijcai.2023/633 , booktitle = on Pithdoi:10.24963/ijcai.2023/633
- [17] Learning granger causality from instance-wise self-attentive Hawkes processes , author=. International Conference on Artificial Intelligence and Statistics , pages=. 2024 , organization= on Pith
- [18] Toward Temporal Causal Representation Learning with Tensor Decomposition , author=. ArXiv , year= on Pith
- [19] Causal Temporal Regime Structure Learning , author =. Proceedings of The 28th International Conference on Artificial Intelligence and Statistics , pages =. 2025 , editor = on Pith
- [20] Your Autoregressive Model Already Reveals the Causal Graph , author=. arXiv preprint arXiv:2602.01135 , year= on PithPith's reviewarXiv
show 15 more
- [21] Interpretable Clustering with Adaptive Heterogeneous Causal Structure Learning in Mixed Observational Data , author=. ArXiv , year= on Pith
- [22] The impact of prior knowledge on causal structure learning , author=. Knowledge and Information Systems , year= on Pith
- [23] Direct and indirect effects , author=. Probabilistic and causal inference: the works of Judea Pearl , pages= on Pith
- [24] Inferring tree causal models of cancer progression with probability raising , author=. PloS one , volume=. 2014 , publisher= on Pith
- [25] Swe-bench: Can language models resolve real-world github issues? , author=. International Conference on Learning Representations , volume= on Pith
- [26] Gaia: a benchmark for general ai assistants , author=. International Conference on Learning Representations , volume= on Pith
- [27] 2025 , month = mar, howpublished = on Pith
- [28] gpt-oss-120b & gpt-oss-20b Model Card , author=. 2025 , eprint= on Pith
- [29] Gemma 3 Technical Report , author=. 2025 , eprint= on Pith
- [30] QwenLong-L1: Towards Long-Context Large Reasoning Models with Reinforcement Learning , author=. arXiv preprint arXiv:2505.17667 , year= on PithPith's reviewarXiv
- [31] QwQ-32B: Embracing the Power of Reinforcement Learning , author=. 2025 , howpublished= on Pith
- [32] Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities , author=. arXiv preprint arXiv:2507.06261 , year= on PithPith's reviewarXiv
- [33] GPT-4o System Card , author=. 2024 , eprint= on Pith
- [34] OpenAI GPT-5 System Card , author=. 2026 , eprint= on Pith
- [35] OpenAI o3-mini System Card , author= on Pith
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
- DeFA: Dependency-Guided Failure Attribution for LLM Agents cs.AI · Oct 2026
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment