- The paper introduces a large-scale benchmark that injects controlled errors to systematically trace failure origins in LLM agents.
- It employs multimodal and multi-agent testing with a unified error taxonomy, yielding high-fidelity, step-resolved annotations.
- Empirical analysis reveals that LLMs excel at symptom detection but struggle with root-cause attribution, especially in complex, long-trace scenarios.
WHO&WHEN PRO: A Large-Scale Benchmark and Empirical Analysis of Automated Failure Attribution in LLM Agents
WHO&WHEN PRO addresses a critical problem in the development and deployment of LLM-powered agentic systems: the attribution of subtle failures in complex, multimodal, and multi-agent environments. As agent capabilities scale, failures become less conspicuous but more consequential for iterative improvement and safety. Prior approaches to failure attribution have been constrained by their focus on text-only tasks, limited scale, or lack of high-fidelity supervision due to the difficulty of systematic, step-level annotation without intensive human involvement.
The core research question is whether current LLMs, both closed-source and open-weight, can reliably attribute failures in diverse agentic scenarios—specifically, can LLMs identify the responsible agent, the exact step where the failure emerges, and classify the failure mode with actionable granularity. WHO&WHEN PRO is positioned as a comprehensive diagnostic resource to empirically answer this question, and to dissect the limitations of existing models and protocols.
Benchmark Design and Pipeline
The benchmark construction pipeline is predicated on the decisive-error formulation: for a given agent trajectory, the error attribution target is the earliest action that, if corrected, would have averted the failure outcome. WHO&WHEN PRO introduces several methodological innovations to ensure label fidelity and representativeness:
- Controlled Warm-Start Error Injection: Successful agent trajectories are replayed up to a selected step, at which a taxonomy-conditioned error is injected using a context-preserving prompting strategy. The agent then continues execution from this state, ensuring post-injection behavior is an authentic reaction to a single, deliberate fault, and strictly isolates the failure’s origin.
- Multimodal and Multi-Agent Coverage: Source trajectories are acquired from 15 agentic frameworks spanning 26 benchmarks, covering text, image, and video modalities, and both single- and multi-agent topologies.
- Unified Failure Taxonomy: High-quality error-mode taxonomy is derived from expert analysis, encompassing 18 failure modes across perception, reasoning, planning, action, verification, and coordination.
- Quality Assurance: An extensive human review phase demonstrates high agreement with injected labels (step: 94%, agent/error family: 90%, κ=0.73), confirming the validity and unambiguity of failure traces.
The final benchmark contains 12,326 failure traces, with each trace precisely labeled for the error-inducing agent, step, and mode.
Empirical Evaluation and Analysis
Attribution Methods
Three evaluation protocols are tested:
- All-at-Once: The LLM observes the entire agent trajectory and predicts agent, step, and error mode jointly.
- Step-by-Step: The LLM evaluates one step at a time, flagging the first it deems erroneous.
- Binary Search: The LLM recursively narrows the failure region by comparing trajectory segments.
All-at-once consistently outperforms incremental methods in joint and step-level accuracy due to its global context.
Attribution Results and Error Analyses
Comprehensive evaluation of 10 LLMs reveals several critical trends:
- Attribution Remains Suboptimal: The best models achieve up to 73.9% accuracy in deciding the correct step on text tasks, but performance is consistently lower in error-mode classification (max 22.2% macro-F1), and attribution on multimodal/video is substantially more challenging.
- Divergence Across Modalities: Step-localization accuracy degrades with additional modalities—long context and interleaved observations obscure the decisive error step, whereas error-mode classification benefits from richer multimodal cues.
- Salient Symptom vs. Root Cause: LLMs tend to attribute failures to the most visible symptom (often reasoning) rather than tracing back to the originating cause (such as planning or verification), especially as errors propagate across agent actions. The confusion matrix shows frequent mislabeling of planning/verification/coordination failures as reasoning errors.
- Trace Length Bottleneck: Step-localization accuracy collapses on longer traces, dropping to nearly 50% on traces >12k tokens, demonstrating that context length and action density directly impede diagnostic accuracy.
- Open-Weight Cost Efficiency: Several open-weight models (e.g., Qwen3.5) approach or surpass closed-source performance at significantly reduced inference cost, suggesting practical utility in attribution tasks for resource-constrained deployments.
Ablations
- Protocol Comparison: All-at-once provides a significant performance advantage (+7.6 percentage points for step accuracy vs. step-by-step) and dominates the cost/accuracy Pareto frontier.
- Ground Truth Provision: When the ground-truth task answer is included, attribution for perception errors improves, but reasoning error detection often degrades, with judges over-indexing on outcome matching rather than process deviation. This reveals that automated self-improvement protocols leveraging outcome signals can misdiagnose root causes.
Theoretical and Practical Implications
WHO&WHEN PRO empirically demonstrates that current LLMs, even at scale, do not saturate the failure attribution task. The results imply several key theoretical and engineering challenges:
- Diagnostic Generalization Barrier: The skills required for high-quality attribution (long-range causal reasoning, multimodal state tracking, fine-grained taxonomy discrimination) are distinct from the next-token prediction and instruction-following capabilities saturated in current LLM pretraining. Model scaling alone is unlikely to close the diagnostic gap.
- Strong Supervision Needed for Self-Evolving Agents: Automated, externally validated, and step-resolved failure signals are necessary for robust agent self-improvement. Coarse or purely self-generated feedback (as in many current self-evolving systems) will propagate and entrench attribution errors, limiting long-horizon adaptability.
- Benchmark Realism: Large-scale, multimodal benchmarks with ground truth rooted in actual agentic process perturbations are critical for both measuring progress and revealing modality- and context-induced weaknesses.
Future Directions
Further research directions catalyzed by this work include:
- Causal Attribution Algorithms: Methods capable of reconstructing or intervening on the causal chain of agent decisions will be required to approach robust attribution, supplanting models that anchor to surface-level similarities or immediate symptoms.
- Attribution-Aware Training: Jointly training LLM agents with attribution feedback and explicit supervision on the taxonomy is an open field, as is the design of architectures or prompting strategies tailored to highly compositional, multi-agent, and multimodal error propagation.
- Expansion into Embodied and Interactive Environments: As agent deployments expand into real-world, continuously evolving, and collaborative contexts, the taxonomy and benchmarks must adapt dynamically, with emphasis on interactive diagnostics and continuous error-trace integration.
Conclusion
WHO&WHEN PRO constitutes a rigorous foundation for diagnosing and understanding failures in LLM-agent systems. By scaling error attribution benchmarks across tasks, modalities, and agentic architectures, and systematically exposing the limits of current LLM-based attribution protocols, it clarifies the frontier and exposes the technical requirements for building reliable, self-improving, and traceably safe AI agentic ecosystems (2607.09996).
