acceptodds
Withdrawn from ICLR 2027

P-FedSkill: A Personalized Federated Skill Learning Framework under Task Heterogeneity

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Skill learning has become an important way to improve agent performance. However, agent trajectories often contain sensitive information and cannot be shared across clients, hindering collaborative skill learning. To tackle this issue, federated skill learning (FSL) has emerged as a solution by distilling trajectories into compact skill representations and uploading them to a central server for sharing. However, existing FSL methods typically share skills based on skill similarity without considering their contributions to heterogeneous tasks across clients, thereby degrading agent performance. To address this challenge, we propose P-FedSkill, a personalized FSL framework that comprises three key components: local skill extraction (LSE), personalized skill selection (PSS), and task-aware skill routing (TASR). Specifically, LSE first extracts skill patches (i.e., task applicability conditions and execution strategies) and task execution outcomes from trajectories. PSS then uses these outcomes to evaluate each skill patch's performance on client tasks and incorporates its task applicability condition to construct a personalized skill set for each client. Finally, TASR adaptively selects a task-relevant subset of skill patches based on their applicability conditions to guide agent execution. We further provide a theoretical guarantee for personalized skill selection in P-FedSkill, and extensive experiments across four benchmarks also demonstrate its effectiveness.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Federated Learning, Skill Learning, LLM Agents

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_H4kLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

This paper studies federated skill learning for LLM‑based agents under client‑side task heterogeneity. Existing federated skill‑learning methods such as FedSkill share skills mainly based on embedding similarity, which may introduce irrelevant or conflicting skills when clients hold different task distributions. The authors propose P‑FedSkill with three components: Local Skill Extraction (LSE) that extracts skill‑patches grouped by task type, Personalized Skill Selection (PSS) leveraging Bayesian Beta posterior with uncertainty penalty to build client‑specific skill sets, and Task‑Aware Skill Routing (TASR) for adaptive on‑the‑fly skill selection at inference time. The paper also provides finite‑sample sub‑optimality bounds for PSS

Critical strengths

Meaningful practical motivation. The paper addresses an important real‑world scenario: edge agents generate private interaction trajectories that cannot be directly shared, and clients exhibit heterogeneous task distributions. It points out the limitation of similarity‑only skill sharing in existing federated skill learning.

Critical weaknesses
  1. Missing centralized‑oracle upper‑bound baseline. All experiments run under federated task partitioning where no single client can access the full training set. There is no centralized‑oracle control: a single agent with access to all training tasks, using the same LSE pipeline and evaluated on the identical held‑out test set. Without this reference, we can only confirm relative improvements versus other federated baselines. We cannot quantify how much performance degradation is purely caused by federated data fragmentation, making it hard to judge how close P‑FedSkill approaches the ideal full‑data capability. This gap exists across all four benchmarks.
  2. Confounding factors in non‑IID simulation setup. The paper simulates client‑side category heterogeneity (different clients prefer different business task types) by partitioning tasks according to raw task types via Dirichlet sampling. However, across the adopted benchmarks (especially τ2\tau^2-Bench), different task types have large intrinsic differences in difficulty. As a result, non‑IID partitioning conflates category heterogeneity (target of the paper) and difficulty heterogeneity (confounder) — these are two distinct phenomena. Worse, training trajectories are dynamically rollout‑generated by each client’s local agent. Clients assigned many hard tasks will produce more failed low‑quality trajectories and inferior skill‑patches, further amplifying the confounder. The counter‑intuitive result that non‑IID accuracy slightly exceeds IID accuracy cannot be reliably attributed to the benefit of personalized skill selection; difficulty‑distribution shift provides an alternative plausible explanation. Only IID comparisons are relatively clean.
  3. Uneven experimental evidence across benchmarks
    • Component‑wise ablations and canary‑based privacy leakage evaluation are only performed on DS‑1000 and GTA. No ablation or privacy evaluation is conducted on τ2\boldsymbol{\tau^2}‑Bench and ALFWorld. Evidence for component effectiveness and privacy protection is extrapolated rather than directly measured on these datasets.
    • On τ2\tau^2-Bench, only aggregated dataset‑level accuracy numbers are reported (59.17% IID / 60.83% non‑IID). Per‑domain numerical results for Retail / Airline / Telecom are only shown in figures without tabulated values. Therefore it is unclear whether performance gains hold uniformly across easy and hard sub‑domains, in particular the challenging Telecom domain.
Questions for authors

Q1. How large is the performance gap between P‑FedSkill and the centralized oracle in your simulation setting? Could this gap explain most of the remaining error under federation? Q2. In your non‑IID setup, to what extent is client performance correlated with the fraction of intrinsically hard tasks assigned to that client? Could the slightly higher non‑IID accuracy be explained purely by difficulty distribution shifts? Q3. On τ2\tau^2-Bench, does the improvement persist in the hardest Telecom sub‑domain, or is the overall gain mainly driven by easier Retail cases? Q4. What fraction of test episodes actually activate the adaptive‑branch logic in TASR? How many samples effectively reduce to fixed‑K selection?

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

This paper proposes P-FedSkill, a personalized federated skill learning framework for heterogeneous client tasks. It extracts compact skill patches from local trajectories, constructs client-specific skill sets using execution statistics and task applicability, and routes relevant patches for individual tasks. The paper provides a finite-sample suboptimality bound for personalized skill selection under its defined scoring objective and evaluates the framework on four benchmarks under IID and non-IID settings.

Critical strengths

The paper addresses the differing utility of shared skills across heterogeneous client tasks. A skill that benefits one client may be irrelevant or harmful to another client with a different task distribution. Considering both client-level skill selection and task-level routing is a useful direction for collaborative skill learning.

Critical weaknesses

The motivation does not sufficiently distinguish task–skill mismatch from differences in skill utility after appropriate task matching. Figure 1 primarily illustrates a shared skill being useful, irrelevant, or misleading for different task types. This motivates task-aware matching, but does not establish what additional difficulty remains once skills are accurately matched to their applicability conditions. The ablations provide evidence for the complete framework, yet do not directly isolate the benefit of outcome-based personalization over selection based solely on semantic task applicability. This distinction matters for assessing the specific contribution of PSS beyond improved task–skill matching.

Questions for authors
  1. Skill mismatch and similarity-based selection. Could the authors provide a concrete example showing how similarity-based sharing selects an inappropriate skill? Would selection based solely on task applicability, using the same semantic encoder and routing procedure, resolve the failure, or does outcome-based personalization provide an additional benefit?
  2. LSE implementation. Lines 173–176 outline the extraction process. Could the authors provide the prompts or rules used to extract applicability conditions and execution strategies? How are source trajectories associated with each extracted patch, and how are task-specific operations filtered out?
  3. Success labels. How are the binary outcomes yk,iy_{k,i} in Equation 3 determined for each benchmark? Are they obtained from official task evaluators, agent judgments, or another procedure, and what criteria distinguish successful from failed trajectories?
  4. Figure 4 configuration. What task distributions and skill patches are used for the clients in Figure 4? Since the clients share the same backbone, could the authors clarify whether the performance differences reflect task-type proportions, instance difficulty, or other differences?
  5. Score weighting. What motivates the unit coefficients in Equations 8–9? Could the authors report sensitivity to the relative weights of performance, overall task relevance, and failed-task relevance, or compare against a normalized combination?
  6. Variability across seeds. Appendix C.3 states that results are averaged over three random seeds. Could the authors report standard deviations, confidence intervals, or per-seed results for the main comparisons and ablations? This would help assess whether the reported improvements are consistent across runs.
Reviewer_Wty4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose P-FedSkill, a personalized federated skill learning framework that considers task heterogeneity across clients. This framework aims to address the limitations of existing Federated Skill Learning (FSL) methods, which primarily share skills based on skill similarity and thus fail to sufficiently reflect client-specific task characteristics. To this end, P-FedSkill consists of three modules. First, Local Skill Extraction (LSE) extracts reusable skill patches and task execution outcomes from trajectories. Then, Personalized Skill Selection (PSS) constructs a personalized skill set suitable for each client based on the extracted information. Finally, Task-Aware Skill Routing (TASR) selects skill patches relevant to the current task to support agent execution.

Critical strengths
  1. The authors propose a novel idea of personalized skill sharing that considers task heterogeneity across clients.
  2. The effectiveness of the proposed method is validated through experiments on various benchmarks.
  3. The ablation study reports higher average task accuracy compared to the ablation variants of LSE, PSS, and TASR.
Critical weaknesses
  1. Section 3.2 and Algorithm 1 in Appendix A lack sufficient details on the specific skill extraction procedure. Although the authors provide some explanations, such as trajectory grouping by task type and the components of skill patches, it remains unclear how individual skills are identified and extracted. In particular, Figure 2 appears to show multiple skills being extracted from trajectories grouped by task type, but the criteria used to distinguish these skills are not specified. This lack of explanation makes it difficult to understand the proposed LSE.
  2. The authors use the number of successes and failures in the source trajectories associated with each skill to evaluate its performance. However, it is questionable whether this approach accurately reflects the actual performance of the skill. A single trajectory may involve multiple steps, and the skill being evaluated may not be directly related to the success or failure of the task. This potential mismatch may also affect the reliability of Personalized Skill Selection discussed in Section 3.3. Therefore, further explanation is needed regarding the appropriateness of the proposed method for evaluating individual skill performance.
  3. The authors explain that they first group trajectories by task type before extracting skills to reduce the mixing of execution strategies from different task types. However, one of the main objectives of skill extraction research is also to discover generalizable skills that can be applied across multiple tasks. Therefore, the proposed method may face difficulties in performing tasks that were not used during skill extraction.
  4. The baselines used in the experimental section focus primarily on FSL variants, without comparisons to existing skill learning and personalized federated learning methods. Although the authors provide comparisons with FedSkill-based variants, additional comparisons with some of the methods discussed in the related work would be necessary.
Questions for authors
  1. The current experiments fix the number of clients at five. However, the number and composition of skills collected by the server may vary depending on the number of clients, which could also affect Personalized Skill Selection. I wonder how the performance of P-FedSkill would change with fewer or more clients.
  2. Can the previously learned skills be effectively utilized for new task types that were not encountered during training? Since the current experiments primarily evaluate heterogeneity within existing task types, I am also curious about the generalization performance on unseen task types.
  3. The proposed method uses task types to group trajectories. I wonder whether it can also be applied in environments where task types are unavailable or where different task types are difficult to clearly distinguish.

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

Related papers

Discussion (0)

Sign in to comment.