acceptodds
Under review as a conference paper at ICLR 2027

FLOWROUTER: EXPERIENCE-GUIDED ROUTING FOR QUERY-ADAPTIVE MULTI-AGENT WORKFLOWS

Abstract

Large language model (LLM)-based multi-agent systems can improve problem solving through collaborative reasoning, yet a uniform workflow may incur redundant computation for some queries while providing insufficient reasoning capacity for others. Current methods include both manually designed and automatically optimized workflows, but adapting complete workflow configurations to individual queries remains challenging. To address this, we propose FlowRouter, an experience-guided framework for query-adaptive multi-agent workflow routing. FlowRouter adopts an explicit graph-based factorization of workflow construction into conditionally dependent decisions over topology, collaboration scale, and position-wise LLM assignment. The resulting routing decisions are jointly optimized under a shared execution-level performance–cost reward, with retrieved workflow experience providing auxiliary guidance. Across four benchmarks, FlowRouter achieves the highest macro-average performance among the evaluated methods while maintaining a favorable inference-cost trade-off. Further analysis shows that topology selection varies significantly across MATH subfields but not across official difficulty levels, suggesting that workflow adaptation reflects characteristics beyond scalar difficulty.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Multi-Agent Systems, LLM Routing, Agentic Workflows

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_iUsDOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

Multi-agent workflows spend different amounts of computation on different queries. The paper routes each query through three dependent decisions, the collaboration topology, the number of agent positions, and the language model at each position. A frozen sentence encoder feeds small trainable heads, and retrieved records of past executions bias the routing distributions. All decisions are trained with a policy gradient under one reward that subtracts inference cost from task performance. On four benchmarks the method is compared with routing and workflow baselines on accuracy and training cost.

Critical strengths
  1. The three decisions are trained under one execution level reward instead of per stage proxies. The ablation includes a shuffled assignment control that keeps model composition and breaks placement. That is the right test of whether positions matter.
Critical weaknesses
  1. The paper never defines what a node does. There is no prompt template per position, no edge payload specification, and no definition of the aggregation step, including whether it is a model call counted in the reward cost. Every number depends on these choices.

  2. The ablation undercuts the headline dimension. Random topology sampling scores within about half a point of learned routing on MATH at lower cost. Fixed topologies span roughly one point, and no table carries seeds or confidence intervals.

  3. The learned behavior is mostly benchmark level. One backbone receives about four fifths of positions on MMLU and under four percent elsewhere. The memory retrieves only same domain records, so most behavior is reproducible by a per benchmark lookup. The subfield association in Figure 9 has no per cell counts, and the null difficulty result is equally consistent with a router that does not track difficulty.

Questions for authors
  1. Specify the per position prompt, the edge content, and the aggregation step. Is aggregation a model call, which model runs it, and does it count inside the cost term.

  2. With five seeds and intervals on the main and ablation tables, does learned topology routing beat random topology sampling on MATH, and does the method stay above MasRouter on MATH.

  3. What happens when every position is fixed to each of the other three backbones. The current fixed assignment arm uses only one backbone, and not the strongest one on the two benchmarks where it is reported (Table 1 has gpt-oss-120b above nemotron-3-ultra on MATH and MBPP), so it conflates placement with model choice.

  4. What are the test set sizes and the per token prices behind the cost axis, and what does the router itself cost per query next to a routed single call.

Reviewer_ue13OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper studies query-adaptive construction of LLM multi-agent workflows. It represents a workflow by three coupled decisions — collaboration topology T (one of six families: chain, star, tree, full, layered, random), collaboration scale k (number of agent positions, 1–6), and a position-wise LLM assignment over a four-model pool — and routes them for each query with a factorized policy F(S|Q,R) = F_t(T|Q,R) · F_k(k|Q,T,R) · ∏_i F_m(M_i|Q,T,k,i,R). A retrieval-based experience memory of previous training executions supplies multiplicative guidance factors on each routing distribution. Because the decisions are discrete and execution is a black box, the router is trained with REINFORCE under an execution-level reward R = U(output, answer) − λ_c · C(cost), with an EMA baseline and an entropy term. The method is evaluated on GSM8K, MATH, MMLU and MBPP against single-agent prompting, fixed multi-agent collaboration (MacNet), automated workflow search (AFlow) and routing methods (RouteLLM, RouterDC, MasRouter), reporting accuracy, USD cost per query, training API cost, ablations over the three routing dimensions and the memory, and an analysis of routing behaviour across benchmarks and MATH subfields. Reported macro-average accuracy is 92.94%, claimed to exceed the best single backbone by 7.74 points and MasRouter by 2.08 points.

Critical strengths

S1. A clean, composable formulation of query-level workflow routing. The paper decomposes workflow construction into topology → scale → position-wise model assignment and keeps the three decisions conditionally dependent, which is a natural and useful extension of prior routing work that selects a single model or a collaboration mode. Because the chain rule keeps each factor individually parameterized while preserving structural context, the formulation is reusable beyond this specific instantiation. This is consequential because it is the main conceptual contribution of the paper, and it is delivered coherently in Section 3 and Figure 1.

S2. The evaluation treats cost as a first-class object and analyses routing behaviour within a benchmark. Reporting inference cost on a Pareto plot (Figure 2 and Appendix D), training API cost (Table 2), and a λ sensitivity study (Figure 4), rather than only accuracy, is more than most multi-agent orchestration papers do, and the subfield-level analysis in Section 4.3 goes beyond the usual "different benchmarks route differently" observation. This matters because cost efficiency is the paper's second headline claim and the design space it explores is only meaningful under a budget.

Critical weaknesses

W1. The method as described cannot be reproduced, and the two most distinctive components are under-specified. (a) The instantiation G_Q = Γ(T_Q, k_Q) is never defined: for a given family and k, which edges exist, how "star"/"full"/"layered" differ, and why random counts as a topology family are left to the reader ("For the random topology, a chain backbone is retained and additional forward edges are sampled"). (b) Node roles and prompts are absent from the method and the appendix, yet the case studies (Figures 10–11) show role names such as "Mathematical Analyst", "Arithmetic Verifier", "Code Reviewer" that are not among the routed decisions; nor is the aggregation/"Aggregate / output" step defined. Since the routed object is only (T, k, {M_i}), the reader cannot tell what actually differs between two sampled workflows beyond the model names. (c) In the experience module, w_j is introduced as "its signed experience weight" (Appendix A) but never defined, and Eq. 9 computes B = β/(Σ_j s_j + ε) · Σ_j w_j I[·], which makes the guidance vanish as retrieved evidence grows — the opposite of the intended "more evidence, stronger prior". This is consequential: the memory is one of three claimed contributions and its ablation (Table 3) is used as evidence for it, but the mechanism cannot be checked or re-implemented from the manuscript, and no code or prompt release is promised.

W2. The empirical comparison does not yet establish the paper's two headline claims. (a) No seeds, no standard deviations, no confidence intervals appear anywhere; Table 1 reports differences of 0.94 (MATH) and 2.08 (average) points over MasRouter, which for a stochastically routed workflow with sampled topologies and models cannot be distinguished from run-to-run noise. (b) The multi-agent and workflow-search baselines are run with a single homogeneous backbone ("For homogeneous baselines, we generally use nemotron-3-ultra as the shared backbone") while FlowRouter and all other routers get the four-model pool; MacNet is additionally pinned to "a fixed seed-1 connected random DAG with four actor nodes", i.e. not the configuration MacNet's own study uses. (c) AFlow's MBPP result (45.67%) is implausibly low and is the denominator of the "30% lower training cost" claim in Table 2, so that claim is currently unsupported. (d) Because the paper's selling point is the cost–performance frontier, the missing control is a cost-matched simple baseline: at FlowRouter's average cost per query, Vanilla nemotron-3-ultra / gpt-oss-120b with best-of-N or self-consistency is the obvious thing to beat, and Vanilla gpt-oss-120b already reaches 89.48% on MBPP at a fraction of the cost. Without it, the +4.67 MBPP margin cannot be attributed to structured collaboration rather than to spending more tokens. This is consequential because it is the difference between "adaptive structure helps" and "more inference compute helps".

(Deliberately not listed as critical, but relevant to the recommendation: the delta over MasRouter is "topology + scale + position-aware assignment", and the ablation shows the weakest of the three is the position-wise one ("w/ Shuffled LLM Assignment" loses 0.37 on MATH and 2.00 on MBPP), while the largest effect is heterogeneous-vs-homogeneous model use. Graph-structure learning work such as G-Designer and AgentPrune is neither cited nor compared.)

Questions for authors
  1. Workflow semantics. Please define Γ precisely for all six families and all k ∈ [1,6] (edge sets), specify the system prompt / role of each position and how intermediate outputs are passed and aggregated, and state whether code and prompts will be released. Without this, none of the reported numbers can be reproduced or attributed.

  2. Experience guidance. How is the signed weight w_j computed (sign, magnitude, dependence on y_j and c_j)? Is Eq. 9 intended as β · Σ_j s_j in the numerator rather than β/(Σ_j s_j + ε)? Does the current form not make the guidance weaker as the retrieved evidence grows?

  3. Experimental protocol. How many random seeds underlie each number in Tables 1–3, and can you report mean ± std (or confidence intervals)? Which λ_c and which E were used for the main results, and which training/validation split was used for MMLU? Was any test-set information used for model or hyperparameter selection?

  4. Baselines. AFlow's 45.67% MBPP pass@1 is far below its published behaviour — is the re-implementation misconfigured? Please report AFlow and MacNet with their own recommended configurations (and MacNet with the heterogeneous pool where applicable), and re-state the Table 2 training-cost comparison in that light.

  5. Cost-matched control. At FlowRouter's average cost per query, what do Vanilla nemotron-3-ultra and gpt-oss-120b achieve with best-of-N / self-consistency? If a simple budget-matched baseline matches or exceeds FlowRouter, what is the remaining evidence that the gain comes from structural adaptation?

  6. Routing analysis. How many test queries fall in each MATH subfield cell of Figure 9 (Number Theory is routed 100% to star), and what power does the test have for the null result on difficulty (p = 0.393)?

Criteria under which my score would rise: a precise specification of Γ, node prompts and aggregation plus a code release; corrected definition of Eq. 9; multi-seed results with variance; baseline re-runs at their own configurations, including a cost-matched best-of-N control. Meeting (1) and (3)–(5) would move me to 3 (weak acceptance); the current manuscript, as written, does not let me verify the claims.

Reviewer_LDTvOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This work proposes a new algorithm for automatic workflow design. The policy networks are trained to select workflow configurations based on user queries. Compared with prior work, the proposed approach expands the workflow design space to include workflow topology, number of collaborating models, and the choice of underlying LLMs. Furthermore, the authors employ RL to train these policy networks, with rewards aimed at enhancing task utility and inference cost.

Critical strengths

The work is well written and easy to follow. It also introduces a novel method for automating the design of LLM workflows.

Critical weaknesses
  • The benchmarks used in the evaluation seem outdated. Recent work suggests that stronger prompting techniques (e.g., when prompted to decompose tasks), can establish competitive baselines [1]. Xu et al. [1] show that multi-agent systems often yield only marginal improvements over existing approaches on commonly used benchmarks, and that appropriate prompting can further narrow this gap. While I understand that the selected benchmarks were also used in the compared baselines, I would encourage the authors to evaluate all the methods on more challenging tasks to better demonstrate the advantage of the proposed approach.

  • Some experimental details are unclear. For example, although the paper states that the reward consists of a task-utility function and execution cost, it does not specify how task utility and execution cost are quantified numerically. Similarly, the authors mention that candidate workflows include chain, star, etc. While readers can often picture the workflows corresponding to these names, it is suggested to include their formal definitions and implementations.

  • The proposed method appears to extend prior work by introducing a more flexible routing policy. While I am not aware of prior work adopting exactly the same approach, the extension seems relatively straightforward, and the novelty is somewhat limited in the absence of additional insights. To strengthen the contribution, I encourage the authors to articulate the key challenges caused by the considered setting and explain what new techniques are developed to address them.

[1] Xu et al., Rethinking the Value of Multi-Agent Workflow: A Strong Single Agent Baseline

Questions for authors
  1. What does kQk_Q represent? Does it refer to the number of LLMs involved in a task or the number of substeps in the workflow?
  2. How is the entropy of the factorized distribution FθF_\theta computed? Could the authors provide the corresponding formulation or further implementation details?
  3. AFlow leverages search algorithms (e.g., MCTS) to optimize workflows in a relatively open search space. In contrast, the candidate workflow topologies in this work appear to be predefined. Does this design choice limit the method's ability to discover more diverse workflows (e.g., loops like those in [1])? How would the method handle tasks that benefit from hybrid workflows combining multiple predefined topologies (e.g., chain + tree)?

[1] Du et al., Improving factuality and reasoning in language models through multiagent debate

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.