acceptodds
Under review as a conference paper at ICLR 2027

CoWA: Learning Compact Decision Policies from Controller-Executed Workflows

Abstract

Large language models (LLMs) are moving from open-ended reasoning toward constrained system control, where agents transform structured states into controller-executable actions that optimize system utility under operational constraints. This shift calls for compact LLM agents that can operate under tight latency, cost, and privacy requirements without repeatedly querying cloud models. Existing methods study tool interaction, trajectory supervision, and post-training adaptation, but leave a key design question unresolved: which parts of a domain workflow should be executed by the controller, and which should be learned by the model? We propose Cold-start Workflow Alignment (CoWA), a framework that separates fixed domain computation from final-action learning. CoWA performs task and interface alignment, tool-augmented cold-start supervision over a fixed controller-executed workflow, and GRPO-based objective alignment. The compact policy learns to synthesize the final action from tool-derived observations, while the controller retains responsibility for tool execution and post-generation checks. This design supports constraint-aware decision-making without iterative per-instance utility search. We instantiate CoWA in a representative Wi-Fi channel-allocation case study. Using 750 training trajectories, a CoWA-trained Llama-3.1-8B achieves an average throughput of 293.5 Mbps, compared with 291.8 Mbps for a prompt-engineered GPT-5.2 baseline, with an observed end-to-end wall-clock speedup in our evaluation setup. CoWA also approaches the performance of the search-based baseline without using its search budget.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large language models (LLM), constrained system control, group relative policy optimization (GRPO), Wi-Fi channel allocation

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_s1UsOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies how to use a small LLM to make constrained resource allocation decisions. The concrete setting is channel allocation for Wi-Fi 7 MLO, i.e., deciding which channel each station connects to on each frequency band, which maximizes the total network throughput while satisfying the constraints. The paper proposes CoWA, which splits the task into two parts. A controller calls a set of domain tools (pairwise distances, channel-specific conflict graphs, conflict intensity, channel loads, initial groups, and zero-link STAs) and writes their results into the prompt together with the network state, and then the LLM decides the final allocation. The training includes three stages, (i) organizing the network state, constraints, tools, and output format into a unified text interface, (ii) SFT cold start on tool-augmented trajectories, and (iii) GRPO optimization with a reward composed of the simulator's throughput, format correctness, and allocation validity. The paper runs experiments with Llama-3.1-8B in a simulated environment. Its throughput is close to that of a prompt-engineered GPT-5.2 and of the search-based method BAI-MCTS.

Critical strengths

None. I cannot identify any strength of this paper. In my opinion, the contribution is merely a fairly straightforward adaptation of conventional LLM fine-tuning to a Wi-Fi scenario. The motivation, the challenges, and the intuition of the solution are all unclear, and I do not see any value brought by using an LLM for decision-making.

Critical weaknesses
  1. The motivation for using an LLM is unclear. (i) The LLM's input consists only of numbers such as coordinates, channels, and carrier-sensing radius, plus the distances and conflict results already computed by the tools. All physical knowledge about the channels is handled by the tools and the simulator, and (in my opinion) the LLM itself does not understand this physical knowledge. If so, why use an LLM instead of a GNN, an MLP, or a conventional optimization method? The paper also does not analyze what the LLM actually learns or what it adds on top of the tools' heuristic results. (ii) The paper is motivated by the fact that search methods need thousands of evaluations per instance, but there is no systematic comparison with the cost of search methods. The runtime of BAI-MCTS is not reported at all. Only the runtimes of CoWA and GPT-5.2 are compared, and I think the comparison is unfair, since they run on different hardware (i.e., authors' own device vs. OpenAI server).

  2. The paper keeps emphasizing agents and tool calling, but there is no real agent or tool calling. The 6 tools are called by the controller in a predefined order before LLM inference, all their results are fed to the LLM, and the LLM never chooses a tool. This is closer to prompt-engineered context than to real agentic tool calling. Yet the paper calls it an LLM agent in the abstract, the introduction, and Figure 3, and discusses it alongside ReAct, Toolformer, AgentTuning, and similar works in the related work. The framing does not match what is actually done. Without this framing, the method is essentially SFT and GRPO on a prompt augmented with domain features, and the novelty is limited.

  3. The experiments are insufficient to support the paper's conclusions. (i) I feel there are a lot of naive baselines that should be compared. The tools already compute an initial allocation, and the LLM mainly completes it and adds links on top of it. However, the paper does not compare with simply using the tool results plus a simple completion rule, nor does it give the same tool results to GPT-5.2 or to an untrained Llama-8B. If these alternatives perform similarly, the LLM step contributes almost nothing. (ii) The improvement is too small. As shown in Figure 4, the average throughputs of CoWA, GPT-5.2, and BAI-MCTS differ by only ~1%, and I also notice that the BAI-MCTS baseline achieves higher throughput with 6 and 12 STAs. (iii) The experiments consider only one simulated case with a fixed setting of 3 BSSs and 3 channels. The paper needs to consider more diverse settings to demonstrate the generalization of the proposed method, especially when the LLM faces more BSSs and more channels, where the large amount of numerical input may cause the LLM to overlook part of the information during reasoning and make wrong decisions.

Questions for authors
  1. I am not sure about the rationale and intuition for using an LLM for this kind of problem, i.e., Wi-Fi resource allocation. Specifically, without the fine-tuning proposed in the paper, does the LLM already understand this physical information? Do more capable LLMs understand this information better? This would be one motivation for the proposed method. If they already understand it, does the paper's fine-tuning teach them how to allocate resources? If they do not, does the fine-tuning teach them to understand this information, or does it merely make them memorize these numbers and make decisions?
  2. I suggest that the authors compare the end-to-end cost of conventional methods and the LLM. In particular, can the thousands of searches in conventional methods be parallelized? Please make a fair comparison, for example on the same hardware as well as on several typical deployment platforms (e.g., edge devices).
  3. In the real world, can any system tolerate such a long LLM inference time (36.50 seconds per allocation on average)?
Reviewer_qs77OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents Cold-start Workflow Alignment (CoWA), a framework for a Wi-Fi channel-allocation task that separates fixed domain computation done by ordinary software from what the model should learn. CoWA performs task/interface alignment, tool-augmented cold-start supervision over a fixed controller-executed workflow, and objective alignment based on GRPO. The paper fine-tuned Llama-3.1-8B with CoWA and achieved a throughput comparable to GPT-5.2 in the Wi-Fi 7 MLO allocation problem.

Critical strengths
  1. The separation of what the model should learn vs. what is done by other software tools is a fundamental problem in many domains, and this paper presents one design point for it in the Wi-Fi channel allocation task.
Critical weaknesses
  1. The paper does not sufficiently establish the benefit of using an LLM for this task. In Figure 4, CoWA does not consistently outperform the non-LLM BAI-MCTS baseline. Although CoWA avoids BAI-MCTS's iterative per-instance search, the paper does not directly compare their runtime or total computational cost.
  2. Another concern is about train/test separation. It is unclear from the paper how evaluation topologies are kept disjoint from instances used in training. Also, the generalization study is limited to changes in the number of client devices within the same simulated task. It does not evaluate generalization to different numbers of access points or channels, different simulation conditions, or other tasks.
Questions for authors
  1. How does CoWA compare with BAI-MCTS in end-to-end runtime and computational cost, including the upfront cost of data generation and training? How do they compare at the unseen network sizes evaluated in Figure 5?
  2. How are training and evaluation instances separated across SFT, GRPO, and testing? Have you evaluated transfer to configurations with different numbers of access points or channels, or to other tasks?
Reviewer_mcAWOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

CyberGraft improves agent harnesses by extracting reusable mechanisms from external implementations and saving them as reusable assets. Each asset maintains a benefit-risk profile tracking its historical performance. To verify code viability upon transplantation, the framework employs a 3-way validation mechanism. CyberGraft achieved good improvements. In particular, a CoWA-trained Llama-3.1-8B achieved an average throughput of 293.5 Mbps vs a 291.8 Mbps for a prompt-engineered GPT-5.2 baseline, while providing an end-to-end speedup of 2.69×.

Critical strengths
  • Nice separation of concerns: the controller performs deterministic operations while the model learns the final allocations. This avoids iterative utility search.
  • The reported improvements are decent, i.e., 293.5 Mbps versus 291.8 for prompted GPT-5.2 and 297.2 for BAI-MCTS,
  • Nice ablations quantifying both the benefits of tool context and GRPO: removing them reduces throughput by 29.5 and 30.7 Mbps, respectively (Table 1).
Critical weaknesses
  • It is unclear how the system handles invalid allocations. Table 1 reports that no link is allocated to about 0.05 of stations, though, if I interpret Eq 17 correctly, every station must receive at least one link. While the controller checks validity, you do not explain what happens when this check fails or if the results include any correction costs. As such, it is unclear whether the reported throughput and latency are representative for a system that always satisfies the stated constraints.

  • You do not quantify the impact of the learned policy. You should provide results for deterministic completion and non-LLM policy baselines that leverage the same information. Without this, the experiments do not fully establish the benefits of using LLMs.

  • There are some gaps in the description of the evaluation. Appendix B.1 does not specify the topology sampling distribution, the separation of SFT/GRPO/test topologies, or the algorithm used to compute the reference allocations. These omissions weaken your generalization claims.

Questions for authors
  • Are the results reported in Table 1 before or after validation? Do the reported throughput/runtime include any rejection, repair, or fallback?
  • How do a deterministic completion heuristic and GPT-5.2 perform when using the same tool outputs? Please report throughput and latency numbers, including BAI-MCTS latency, to quantify the benefits of the learned policy.
  • How do you generate and assign topologies across SFT, GRPO, and testing?
  • How do you compute the reference allocations?
  • What happens when an output fails the validation test?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.