acceptodds
Withdrawn from ICLR 2027

TurnRoute: Learning Modular Agent Capabilities for Dynamic Composition

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Language-model agents coordinate multiple capabilities over long interactions, yet typically learn them jointly within a monolithic set of adapted parameters. Can these capabilities instead be decoupled into separate parameter modules and dynamically recomposed during execution? We introduce TurnRoute, a framework that trains capability-specific LoRA experts through turn-level loss masking of labeled trajectories over a shared frozen backbone. During execution, a router selects one adapter to generate each turn while maintaining a shared interaction history, with an unmasked adapter providing a fallback. On software engineering tasks, capability probes show that masked experts retain comparable competence to monolithic adaptation on their assigned capability while exhibiting behavioral separation across adapters. Although individual specialists underperform as standalone agents, TurnRoute achieves strict pass@1 of 15.8% at 7B and 28.4% at 14B on 500 SWE-bench Verified instances, outperforming monolithic LoRA by 1.6 and 2.0 pp. It reaches 98% and 92% of full fine-tuning performance with roughly 9.4–11.7× fewer trainable parameters than full fine-tuning, while activating only one expert adapter per turn for response generation. In addition, at 7B, TurnRoute increases non-empty patch production from 77.8% to 94.4%, reducing empty-patch failures by 74.8%. Together, these results demonstrate the feasibility of decoupling agent capabilities into separately trainable parameter modules and dynamically recomposing them while retaining end-to-end task competence, providing a practical approach to modular adaptation of language-model agents.

Primary area:
foundation or frontier models, including LLMs
Keywords:
LLM agents, parameter-efficient fine-tuning

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_QorwOpenReview ↗
Rating
2
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes TurnRoute, a framework that decomposes agent capabilities into independently trained LoRA experts and dynamically selects them at each turn. Experiments on SWE-bench show that TurnRoute achieves better performance than monolithic LoRA while approaching full fine-tuning performance.

Critical strengths

Decomposing agent capabilities into independently trainable modules is an interesting research direction, and this paper provides an initial exploration of its feasibility. The paper conducts detailed experiments to support its claims.

Critical weaknesses
  • Unclear practical motivation. I am not convinced of the practical benefits of modularizing agent capabilities. TurnRoute does not reduce inference cost compared to monolithic LoRA, introduces additional router overhead and memory usage, and still underperforms full fine-tuning. It remains unclear what concrete advantages justify this additional complexity.

  • Performance gains are not attributable to modularization. In Table 2, monolithic LoRA achieves 14.2%, TurnRoute achieves 15.8%, while monolithic LoRA with the Foresight routing signal achieves 16.8%. This suggests that the improvement mainly comes from the routing signal rather than capability-specific LoRA experts. The paper therefore does not convincingly demonstrate the benefits of parameter-level modularization over a simpler routing-guided monolithic agent.

Questions for authors

see Weaknesses

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

This paper studies whether capabilities exercised within software-engineering agent trajectories can be learned as independently composable parameter modules. Turn-level loss masking is used to train three LoRA specialists—localization, verification, and code generation/tool use—over a shared frozen backbone, together with an unmasked fallback adapter. At inference time, a router selects an adapter for each assistant turn based on the interaction history. On SWE-bench Verified, TurnRoute achieves resolution rates of 15.8% and 28.4% at 7B and 14B, compared with 14.2% and 26.4% for monolithic LoRA. I view the main contribution as a feasibility study of capability decomposition and dynamic recomposition, rather than evidence that modular systems generally outperform non-modular alternatives.

Critical strengths

S1. The method addresses a well-defined modularity question and differs substantively from related approaches.

S2. The experiments provide useful evidence of end-to-end feasibility.

S3. Several controls are diagnostically useful, and the authors transparently discuss results that do not favor their proposed configuration.

Critical weaknesses

W1. The independent contribution of modularization remains insufficiently established given the routing-instruction control.

In Table 2, providing the same Foresight router’s decisions as instructions to monolithic LoRA achieves 16.8%, compared with 15.8% for TurnRoute. The reported parameter counts correspond to approximately 323.0M trainable parameters for the former and 807.5M for the latter. The authors correctly acknowledge that the routing signal explains much of the performance gain. My concern is therefore not that this important control is missing, but that its result places greater importance on establishing what expert modularization contributes beyond being executable.

Preserving comparable performance in exchange for useful modular properties can certainly constitute a contribution; achieving the highest resolution rate is not necessary. However, the experiments do not directly demonstrate that updating or replacing one specialist preserves the effectiveness of the remaining capabilities and the composed system. Section 3.1 establishes that adapters can be retrained or replaced independently at the parameter level, but parameter independence does not itself establish behavioral compatibility after replacement. Meanwhile, Appendix E reports that training the four adapters costs nearly four monolithic training runs and that the 14B system requires approximately 1.6 times the resident memory of monolithic LoRA.

I am not asking the authors to validate every potential use of modularity. Nevertheless, some direct evidence is needed to establish what these additional costs buy that the routing-instruction control does not already explain. Reporting patch production for that control would help determine whether the paper’s clearest behavioral improvement also primarily comes from routing. A focused specialist-update or replacement experiment would instead test a practical benefit of modularity. This is a central concern for my recommendation.

W2. The evidence more directly establishes behavioral differentiation than semantically well-defined capability decoupling.

Table 1 provides useful initial evidence, but differences between specialists do not necessarily show that the corresponding capabilities have been effectively isolated into their designated modules. For example, on the verification probe, the verification specialist and monolithic LoRA both achieve 6.2%, the base model achieves 6.0%, and the off-capability adapter achieves 3.5%. This is also consistent with some adapters preserving a base-model capability while others degrade it; the between-adapter gap alone does not distinguish these explanations.

The random-supervision-token control is valuable: it shows that the code specialist’s lower end-to-end performance cannot simply be attributed to receiving fewer supervised tokens. However, this control is evaluated only through end-to-end resolution, not through the capability probes. It therefore does not directly establish whether capability-based masking produces more clearly defined specialization than a non-semantic partition with a comparable supervision budget.

Furthermore, Figure 4 analyzes the actual behavior of monolithic and capability-agnostic LoRA, rather than TurnRoute, while Figures 5–6 report adapter selections and switching patterns. Selecting an adapter named “verification” does not automatically establish that the resulting turn performs verification. An analysis of routing-label versus actual-action correspondence in existing TurnRoute trajectories, together with probe results for the existing random-supervision control, would more directly support the central interpretation that these modules perform their intended roles.

Questions for authors

Q1. What is the non-empty patch production rate of Foresight plus monolithic LoRA? Is there existing evidence directly demonstrating an independent benefit of modularization, such as updating or replacing one specialist while holding the other specialists and router fixed?

Q2. In actual TurnRoute trajectories, how closely do the selected capability labels correspond to the functions of the generated turns? How does the existing capability-agnostic LoRA perform on the probes in Table 1? These results would help distinguish intended-role specialization from selective capability degradation.

A numerical discrepancy requiring clarification: The reported 14B non-empty patch rates are 97.4% and 85.2%. Under the same definition and the same set of 500 instances, these imply a difference of 61 patches, whereas the discordant-pair counts in Table 10 imply 67−4=6367-4=63. Please clarify whether the instance sets or counting conventions differ, or correct the relevant numbers.

Reviewer_fA2YOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes a LoRA training framework named TurnRoute, for multi-turn LLM agent, that trains decomposed capabilities into different LoRA adapters, and dynamically select an adapter at each agent turn. The authors evaluate the training framework on a software engineering benchmark, and show modest effectiveness gain over monolithic LoRA baselines.

Critical strengths
  1. The idea of learning modular capabilities through different LoRA adapter is conceptually sound. The authors demonstrate the possibility of separate training and dynamic composition of these adapters.
  2. The evaluation uses models with two different scales (i.e., 7B and 14B).
Critical weaknesses
  1. Based on the results from Table 2, Foresight router (7B) + the proposed modular LoRAs achieves 15.8% pass@1, while the Foresight router (7B) + baseline monolithic LoRA achieves 16.8% pass@1, which means the simpler baseline approach method outperforms the proposed modular LoRAs, yet being more efficient (with 3 * 161.5M less trainable parameters). It raises a concern that if the modular adapter training is practically meaningful.
  2. The authors only evaluate TurnRoute on a specific task in software engineering: resolving GitHub issues. The evaluation testbed is very limited and it is unclear whether the proposed training framework could generalize to other agentic tasks, especially in agentic tasks where the boundaries of capabilities are ambiguous. In addition, the authors only experiment on capability-specific adapters with two model sizes from the same model family.
  3. Hindsight and foresight routers are trained using very different setups, specifically the teacher model, the number of training examples and the number of training epochs are different. Therefore, it is unclear whether the improvement comes from differences in the training setup or the design of the foresight router. In addition, it uses a stronger model to annotate the trajectories for router training, given the router predicts the capability of the next action using such annotated data. Therefore the performance gain could just come from the annotation from a stronger model.
  4. The reported gains over baseline LoRA are relatively small. There is no error bar shown in the main pass@1 results, nor statistical significance or confidence intervals are shown. It remains unclear that the 1.6-2.8 pp improvements are statistical significant to show the superior performance of the propose methods.
Questions for authors
  1. What is the design decision of using one adapter for both the code generation and tool use capabilities, and what is the empirical justification?

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

Related papers

Discussion (0)

Sign in to comment.