Skip to content

DoVer: Intervention-Driven Auto Debugging for LLM Multi-Agent Systems

🕒 Published (v1): 2026-01-01 · Source: ICLR · Venue: ICLR 2026 · link

Why this paper was selected

Microsoft Research; intervention-driven debugging for LLM multi-agent system failures

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

DoVer is an intervention-driven auto-debugging framework for LLM multi-agent systems that moves beyond passive log analysis by actively editing agent messages or plans at suspected failure points and re-executing the trajectory to verify whether the fix resolves the failure. It reframes debugging as hypothesis-test-validate rather than single-step attribution, recovering 18–49% of failed trials across benchmarks.

Problem

Existing LLM-based multi-agent debugging relies on log-only failure attribution—identifying the single agent/step responsible for a failure—but this approach has two fundamental flaws: (1) hypotheses are never validated, leaving untested guesses, and (2) single-point attribution is ill-posed because modern agentic sessions contain multiple ReAct-style trials with distinct failure points, and inter-agent misalignment makes clean responsibility assignment ambiguous. Empirically, 14 of 29 GAIA cases in the WW benchmark have uncertain ground-truth labels, and even strong models (GPT-4o) achieve only 6–44% step-attribution accuracy.

Method

DoVer (Do-then-Verify) is a four-stage pipeline:

  1. Trial Segmentation: An LLM parses the full session log \(\tau = \{(a_t, m_t, \sigma_t)\}_{t=1}^T\) into sub-traces \(\tau^i\) by identifying re-planning steps as cut points, producing independently interveable causal chains.
  2. Failure Attribution: For each trial \(\tau^i\), an LLM generates a hypothesis \(h_i = (\hat{a}^i_{\hat{t}}, r^i_{\hat{t}})\) pinpointing a suspected agent and step with a natural-language rationale, using an "All-at-Once" prompt over the segmented log.
  3. Intervention Generation: Hypotheses are converted into concrete edits—either (a) modified instructions from the orchestrator to sub-agents, or (b) orchestrator plan updates (reorder/decompose/replace steps). Interventions operate at the message-passing layer to remain architecture-agnostic.
  4. Intervention Execution: The system replays from the intervention point forward, preserving all prior steps, yielding a counterfactual trace \(\tilde{\tau}_I = \{\tau_1, \ldots, \tau_{i-1}, \tilde{\tau}_i\}\). Each intervention is repeated 3 times to reduce LLM stochasticity.

Evaluation metrics: Trial Success Rate (full task recovery); Progress Made \(P\text{rog}(\tau \to \tilde{\tau}_I) = \frac{A(\tilde{\tau}_I) - A(\tau)}{K}\) measured via up to \(K \leq 5\) LLM-extracted milestones; and hypothesis outcome classification (Validated / Partially Validated / Refuted / Inconclusive) using an "is intervention fulfilled" boolean signal.

Key Contributions

  • DoVer framework that operationalizes debugging as hypothesis→intervene→verify, eliminating dependence on noisy human attribution labels
  • Empirical study of GT uncertainty in the WW benchmark, showing that ~48% of GAIA cases have ambiguous ground-truth annotations and that multi-trial session structure is the primary cause
  • Two minimal prompt refinements (explicit step indices + annotator-guidance reminders) that raise GPT-4o step attribution accuracy from 6% to 24% on WW-HC
  • Trial-level progress metric using LLM-extracted milestones, enabling partial-success measurement where binary pass/fail is too coarse
  • Demonstration of generality across two agent frameworks (MagneticOne/M1 and AG2) and three datasets (GAIA, AssistantBench, GSMPlus)

Results

  • MagneticOne + AssistantBench: 18% of failed trials flipped to successes
  • MagneticOne + GAIA: 28% of failed trials flipped to successes; up to 16% milestone progress on non-recovered cases
  • MagneticOne + GAIA/AssistantBench: 30–60% of failure hypotheses validated or refuted (depending on task complexity)
  • AG2 + GSMPlus: 49% of failed trials recovered, demonstrating cross-framework generality
  • Prompt refinements alone boosted GPT-4o attribution accuracy from 6% → 24% on WW-HC (All-at-Once setting); for the 15 "certain" WW cases, GPT-5 reaches 53% step-attribution accuracy vs. 24% on uncertain cases

Limitations

  • Interventions are restricted to the orchestrator message-passing layer; direct sub-agent capability improvements (e.g., extending a WebSurfer with in-page search) are out of scope
  • Progress Made metric requires human-annotated intermediate solution steps (available in GAIA but not universally); LLM-as-a-judge trace comparison is left for future work
  • The system does not handle asynchronous multi-agent execution logs
  • Three independent re-runs per intervention reduce but do not eliminate LLM stochasticity; some trials are marked Inconclusive when agents fail to follow the intervened instruction
  • Evaluation is on benchmark tasks during development; online deployment scenarios (user thumbs-down signals) are discussed but not experimentally validated

Relevance to Agentic AI / LLM Agents

DoVer directly addresses a critical bottleneck in deploying reliable multi-agent systems: how to close the loop between failure observation and verified repair without human-in-the-loop intervention at scale. The framework's insight that single-step attribution is ill-posed—because ReAct-style replanning creates multiple independent failure chains within one session—is an important empirical finding for the field. The counterfactual trace replay mechanism (preserve-and-rerun) generalizes across agent frameworks and is architecturally compatible with checkpoint/interrupt systems like LangGraph, pointing toward automated self-healing pipelines. For researchers building orchestrator–sub-agent hierarchies, the taxonomy of orchestrator-level vs. sub-agent-level interventions offers a practical design principle for where automated debugging can act without deep system modification.