acceptodds
Under review as a conference paper at ICLR 2027

Learning What to Skip: Counterfactual Credit Assignment for Efficient Multi-Agent LLM Workflows

Abstract

Multi-agent LLM workflows use planning, execution, verification, and summarization to improve task performance, yet the value of each component depends on the state already produced. Executing every component can waste computation or overwrite a correct intermediate answer. We formulate component omission as counterfactual credit assignment: full-workflow logs reveal the executed trajectory's reward, while controlled skip interventions reveal the consequences of omitting a future step. We introduce Learning What to Skip (LW2S), which learns action-specific safety models from these interventions and combines held-out calibration with domain-native guards to select skips. When an early skip is rejected, the controller can continue execution and reconsider a later component. Across mathematical reasoning, multiple-choice QA, and code generation with two instruction-model families, LW2S reduces recorded token cost while preserving or improving full-workflow reward in the evaluated settings. Scale-up and second-topology experiments further examine component redundancy, while shared-error cases reveal why agreement alone is insufficient for skip selection. These findings connect efficient workflow execution to learning the conditional utility of individual components.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Graph Learning, Workflow, Agent

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_h5zPOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

This paper frames omission of future components in multi-agent LLM workflows as a counterfactual credit-assignment problem. The proposed LW2S method collects full-workflow and single-component skip traces, trains action-specific skip-safety predictors, calibrates acceptance thresholds, and combines them with domain-specific guards and sequential fall-through decisions. The method is evaluated on mathematical reasoning (MATH and GSM8K), multiple-choice question answering (MMLU), and code generation (MBPP), using Qwen2.5-7B and Mistral-7B. The authors report reduced component-token cost while matching or improving aggregate accuracy in the evaluated settings. The paper further studies model-family transfer, a larger model diagnostic, and a debate-refine workflow topology.

Critical strengths
  • I consider inference-cost reduction in multi-agent LLM workflows an important problem, particularly when workflows execute several components even when some may be redundant.
  • I appreciate the use of matched full-workflow and component-skipping traces to construct supervision for skip decisions.
  • I appreciate the inclusion of mathematical reasoning, general knowledge, and code generation benchmarks, as well as multiple model families.
Critical weaknesses
  • I am concerned that the safety objective measures non-degradation rather than correctness. Since Za=I[Ra≥Rfull]Z^a=\mathbb{I}[R^a\ge R^{\mathrm{full}}], a skip action is considered safe even if it preserves an incorrect answer. I would like the authors to report how often skipping preserves existing errors, repairs errors, or damages correct answers, clarifying whether LW2S improves task correctness or merely preserves baseline reward.
Questions for authors
  • How much of the method performance improvement comes from executable-test guards rather than learned skip selection?
  • How robust is the method when task distributions, component implementations, or workflow topologies differ?
  • What are the total data-collection and training costs, and how many future workflow executions are required for the method to relieve these costs?
Reviewer_gNGdOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a framework for identifying components that can be omitted from agentic trajectories in multi-agent systems. It formulates this problem as counterfactual credit assignment and trains action-specific safety models to estimate whether skipping an action would preserve the full-workflow reward, given the observed trajectory prefix. The proposed LW2S framework further incorporates held-out calibration and domain-native guards to determine whether a predicted skip should be accepted. The method is evaluated using a fixed five-role multi-agent framework across four benchmarks.

Critical strengths

The paper is generally well written and clearly structured.

Critical weaknesses

Limited novelty: The motivation is relatively weak, and the proposed method does not demonstrate sufficient novelty over existing approaches.

Insufficient grounding in prior work: The literature review omits several key milestone papers on LLM-based workflows and multi-agent systems. Consequently, the authors' characterisation of existing research does not adequately reflect the achievements and capabilities of prior approaches. This weakens the motivation and theoretical grounding of the proposed method.

Insufficient experimental evidence: The experimental evaluation is not sufficiently convincing to establish the effectiveness and advantages of the proposed framework.

Questions for authors

What motivates the specific design choices for held-out calibration and domain-native guards? Why were these mechanisms selected over alternative approaches, and what evidence supports their suitability for the proposed framework?

Reviewer_8mdpOpenReview ↗
Rating
1
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes LW2S, a controller that decides, after observing a partial multi-agent workflow (planner → two executors → verifier → summarizer), whether a future component can be skipped without reducing reward.

Critical strengths
  • Framing component omission as a counterfactual question that full-workflow logs cannot answer
  • The shared-wrong-consensus and superficial-code-agreement regimes (Table 13), and especially the finding that AST equality selects 94/100 tasks with two degradations while executable tests block them (Table 14), are the paper’s most informative empirical results.
  • Open-weight models, a full resource ledger, released traces and code.
Critical weaknesses
  • The paper’s own tables suggest the learned safety model contributes almost nothing
  • Named baselines are in-house reimplementations
  • Test sets are small and improvements are within noise
Questions for authors

What is the per-action AUROC of the safety classifiers, and on how many tasks does the learned score change the decision the guard alone would make? Why is coverage 100% in five of six settings? Does the policy ever return to the full workflow, and if not, in what sense is it selective? What are single-call accuracy and token cost for each backbone on each benchmark? Why does fixed verifier skipping dominate LW2S on the debate-refine workflow?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.