- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 2
Read review
Paper summary
This paper proposes Exact Frontier Coupling (EFC), which jointly samples actions from multiple LLM agent policies using multi-marginal optimal transport. By preserving each policy's marginal trajectory distribution while maximizing the reuse of identical model requests, EFC aims to reduce the computational cost of policy evaluation.
Critical strengths
S1. Well-motivated problem and principled formulation. The paper addresses the practical cost of evaluating multiple adaptive LLM agents by combining multi-marginal optimal transport with exact-request sharing.
S2. Good theoretical analysis. The paper provides well-structured theoretical results, including duplicate-law compression, the request-set LP characterization, and marginal trajectory preservation. The three-policy example clearly illustrates the limitations of pairwise tree coupling.
S3. Careful experimental design and practical efficiency. The experiments include strong baselines, paired comparisons, and numerical validation. Exact-request sharing achieve inference savings (29–44% on BFCL52) while preserving observed trajectories and evaluation outcomes.
Critical weaknesses
W1. Limited end-to-end improvement over strong baselines. Although EFC achieves local optimality, it does not outperform optimized quantile in the complete-episode experiments. The planned confirmation study was also not conducted, leaving its practical advantage insufficiently demonstrated. (Sec. 6.1, Table 1; Sec. 8)
W2. Unclear attribution of inference savings. The substantial savings reported on BFCL and AgentDojo primarily result from exact-request deduplication rather than EFC's coupling optimization. The incremental benefit of EFC in realistic agent settings remains unclear. (Sec. 6.2; Sec. 8)
W3. Limited robustness of exact-request sharing. Three Qwen models in BFCL64 were excluded due to output mismatches across execution modes. The causes and implications of these failures are not sufficiently investigated. (Sec. 6.2; Appendix J)
W4. Scalability concerns. The tuple-based LP grows exponentially with the number of distinct action distributions, and the method falls back to approximate coupling for larger problems. Its scalability to larger policy banks remains unclear. (Sec. 3.2; Sec. 8)
W5. Limited experimental coverage. The stochastic agent evaluation relies on a small, manually designed policy bank and a limited number of tasks. More diverse and realistic stochastic agent evaluations are needed to establish the practical value of EFC. (Sec. 5; Sec. 8)
Questions for authors
- Can EFC demonstrate consistent improvements over optimized quantile in complete-episode evaluations?
- How much of the reported inference savings comes from EFC's coupling optimization rather than exact-request deduplication?
- What caused the output mismatches for the three Qwen models in BFCL64, and how do they affect the method's reliability?
- How does EFC scale to larger policy banks, considering the exponential LP complexity and optimization overhead?
- Can the authors validate EFC on larger, more diverse stochastic agent policy banks and held-out tasks?
We thank the reviewer for their time and careful evaluation of our paper. The revised paper addresses the five concerns through a matched complete-episode study, component ablations, and scaling and reliability audits.
W1/Q1. Complete-episode improvement (Section 6.1, Table 1; Appendix C). On 128 tasks with 24 stochastic policies, two models, and three seeds, request-set EFC uses 5,767 physical requests versus 7,099 for optimized quantile: an 18.76% reduction, with a multiplicity-adjusted task-cluster 95% interval of [12.22%, 24.86%]. It also improves on fragmentation-aware quantile, best-root TCAB, and tuple-based EFC. All four adjusted intervals are positive, and every comparison favors EFC in all six model–seed combinations.
The pilot's planned confirmation was not run because its continuation gate was not met (Appendix G). The 128-task evaluation was planned separately; its additional ablations reuse the same tasks.
W2/Q2. Attribution (Sections 3.1–3.2 and 6.2; Table 2). We implement the request-set formulation already presented in Section 3.1 of the original submission, retaining its primary objective and marginal constraints. The implementation in the original submission is retained as tuple-based EFC, including its behavior above the tuple-enumeration limit. All methods use the same exact-request-sharing executor and plan-cache budget, so the comparison measures coupling gains with sharing already available. Applying the tuple-based method's 2% acceptance threshold to request-set EFC retains 18.03% savings over optimized quantile; removing quantile-tie preservation retains 16.69%. Neither difference from full EFC is established after multiplicity correction. Section 6.2 separately attributes BFCL's 29–44% savings to exact-request reuse: its deterministic frontiers offer no coupling advantage.
W3/Q3. Qwen mismatches (Appendix L). The BFCL64 audit finds 14/766, 39/644, and 47/761 mismatching unique matched requests for Qwen3-4B/8B/14B, versus 0/2,981 for ToolACE-2-8B. Recorded inputs, seeds, and token limits agree. All 100 Qwen mismatches occurred with shared batches of eight versus independent batches of four, suggesting sensitivity to batch configuration. Full-trajectory checks detected these mismatches, and the affected runs are excluded from matched-trajectory serving claims. The stochastic study separately completed 24,228 repeated-request checks without mismatch.
W4/Q4. Scaling and overhead (Table 3; Appendix D). The primary LP uses 2K variables for K charged request identities. With four identities and 32 distinct laws, uncached optimization and conditional sampler construction take a median 12.44 ms. This audit retains the sampler implicitly; episode evaluation materializes joint assignments. We also vary request support and frontier fragmentation; complexity remains exponential in request support. Episode executor time including planning is 343.97 seconds versus 350.24 for optimized quantile, a descriptive 1.79% reduction (Appendix C, Table 5).
W5/Q5. Diversity and held-out tasks (Section 5; Appendix C). The bank covers all 24 permutations of four action priorities, with feedback-dependent action probabilities, on Qwen2.5 and Mistral. It was fixed before evaluation following development diagnostics. The 128 GSM8K training questions were held out from method development.