acceptodds
Under review as a conference paper at ICLR 2027

Learning Reusable Partial-Order Skills from Agent Traces

Abstract

Agent traces are sequential, but the procedures they record may allow alternative action orders, repetition, and repair. We introduce Reusable Partial-Order Skills (RPOS), a probabilistic model that jointly infers skill boundaries, assignments, and shared precedence constraints from action-labelled traces. The action dictionary, number of skills, are specified in advance; skill decompositions and partial orders are inferred. Each skill instance tracks the validity of action outputs, modelling repetition and repair without adding cycles to its shared graph. Semi-Markov dynamic programming sums exactly over admissible segmentations and skill assignments to evaluate structural proposals. Experiments on synthetic data show improved recovery and prediction in the completed runs. On coverage-controlled TaskBench simulations, RPOS predicts unseen valid action orderings better than the tested total-order and sequence models. On -Retail, RPOS learns an explicit library of partial-order skills and transitions between them, making within-skill precedence relations and between-skill transitions transparent. Finally, on 120 paired synthetic Retail tasks, an executor combining the learned library before evaluation reduces online token use by 86.97% relative to a stepwise agent while meeting a prespecified non-inferiority margin for strict task success.

Primary area:
probabilistic methods (Bayesian methods, variational inference, sampling, UQ, etc.)
Keywords:
Partial-order learning, Bayesian inference, semi-Markov models, agent traces, skill discovery

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_QedVOpenReview ↗
Rating
3
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RPOS (Reusable Partial-Order Skills), a probabilistic model for learning reusable skill libraries from action-labeled agent traces.

The key contribution is jointly inferring skill boundaries, assignments, and shared partial-order precedence constraints using semi-Markov dynamic programming. The paper demonstrates that partial-order representations better capture real procedural flexibility than total-order models, enabling 86.97% reduction in token usage during downstream execution (120 paired retail tasks) while maintaining task success rates.

Validation spans synthetic data recovery, TaskBench workflow composition, and τ³-Retail real-world traces with qualitative skill structure analysis.

Critical strengths
  1. Novel partial-order representation with principled inference. The use of partial-order templates with exact semi-Markov dynamic programming is technically sound and well-motivated. Unlike total-order models, this approach naturally handles alternative orderings, repetitions, and repairs without adding cycles to the shared graph—a consequential architectural innovation.

  2. Compelling downstream results demonstrating practical value. The 86.97% token reduction on 120 paired synthetic retail tasks (Table 2: N=23.1k tokens vs. G=3.01k tokens per task) directly validates the utility of learned skill structure. The executor maintains strict task success (120/120) despite dramatic token savings, showing the model captures genuine procedural relationships rather than spurious correlations.

Critical weaknesses
  1. Severely limited evaluation scope with generalization uncertainty. Evaluation concentrates on τ³-Retail (436 traces, 24 CPA labels) and TaskBench synthetic workflows. The paper provides no evidence the approach generalizes to other agent trace datasets (robotics, web tasks, NLP workflows). Missing comparisons with recent partial-order and sequence discovery methods (e.g., Leemans & van Zelst 2025 POWL) on the same data—only SEQ-HSMM is directly compared.

  2. Unaddressed scalability and critical design choices lack validation. No runtime complexity analysis or convergence diagnostics (R̂, ESS) provided. The method assumes fixed K and action dictionary specified in advance, severely limiting applicability. Key assumptions (no self-transitions, validity as scalar, partially collapsed orderings) are never ablated on real data. RPOS actually underperforms SEQ-HSMM on τ³-Retail 20-trace final test (NLL 1.600 vs. 1.303), yet this critical failure is dismissed rather than investigated.

Questions for authors
  1. Why does RPOS achieve worse NLL (1.600) than SEQ-HSMM (1.303) on the τ³-Retail 20-trace final test? This contradicts the claim of superiority. Can you provide concrete examples of failure modes or data characteristics where total-order models outperform partial orders?

  2. How does the method scale with trace length and number of skills K? What is the computational complexity, and what are the wall-clock runtimes vs. baselines? The paper runs 4,000 MCMC sweeps but provides no convergence diagnostics (Gelman-Rubin R̂, ESS, mixing time).

  3. Can RPOS infer unknown K or automatically discover the number of skills? How sensitive is the model to misspecification of K and the action dictionary?

  4. Can you provide ablation studies on τ³-Retail validating key design choices: (a) the validity state in repeat modeling, (b) the partially collapsed ordering (Prop. 1), (c) hyperparameter sensitivity (β, λ_rep, λ_back)?

2 comments

We thank the reviewer for the detailed feedback. We address the main concerns below.

Q1. Retail predictive performance.

We acknowledge that RPOS underperforms sequence baselines on Retail. To clarify, NLL 1.303 belongs to the bigram baseline, while SEQ-HSMM achieves 1.305 and RPOS achieves 1.600 (K=5K=5).

Our primary contribution is joint inference of reusable partial-order structures, rather than universal superiority in sequence prediction. Sequence models capture observed local transitions, whereas RPOS aims to distinguish essential dependencies from incidental execution order. When local transitions are highly predictable, simpler sequence models may achieve better NLL. This is a possible explanation, not a verified diagnosis. We acknowledge the Retail predictive gap while noting that synthetic experiments evaluate structural recovery and TaskBench evaluates unseen valid orderings.

Q2. Computational complexity and convergence.

Appendix H.1 already provides complexity analysis. With precomputed block scores, the factorised semi-Markov forward pass reduces complexity from O(NJDK2)O(NJDK^2) to O(NJK2+NJDK)O(NJK^2+NJDK), excluding block-score construction and structural proposals.

The Retail experiments use four independent runs per capacity, each with 30,000 MCMC sweeps, not 4,000 (Appendix G.1).

Our MCMC procedure explores a high-dimensional, combinatorial posterior over latent skill structures and assignments, rather than seeking a single globally optimal DAG. In a potentially multimodal posterior, finite runs may not explore all structural modes. We therefore do not claim global posterior convergence. Proposition 1 establishes posterior invariance of the update kernel, not convergence of finite runs. Available R^\hat R, ESS, and between-run diagnostics are reported in the appendices, with incomplete exploration explicitly acknowledged.

Q3. Fixed number of skills and action dictionary.

Fixing KK and the CPA dictionary is a deliberate modelling choice that separates observable action definitions from latent structural inference. Skill boundaries, assignments, and partial orders remain unknown and are jointly inferred.

Importantly, KK controls the granularity of the learned representation: small KK may merge distinct procedures into overly broad skills, while large KK may produce fragmented or redundant skills. We evaluate K∈{4,5,6,8}K\in\{4,5,6,8\} on Retail, observing modest predictive variation, although this does not establish structural stability. Automatically selecting KK and adapting the vocabulary are valuable extensions beyond the present scope.

Q4. Ablations and modelling choices.

Our existing controlled experiments evaluate the principal methodological components:

  • Validity state: TaskBench includes a no-invalidation ablation (κ=0\kappa=0), directly testing repetition and repair.
  • Path marginalisation: Synthetic Full-Marg versus Full-Cond experiments evaluate its finite-run benefits, while Proposition 1 establishes posterior invariance.
  • Hyperparameters: Systematic Retail sensitivity analysis remains a limitation.
  • Self-transitions: Excluding immediate same-skill transitions is a modelling assumption; repetition is represented within skill instances.
  • Generalisability and POWL: We discuss POWL in Section 2 and acknowledge the need for broader cross-domain evaluation.

These targeted experiments assess the core methodological contributions in settings where their effects can be meaningfully isolated. Further Retail-specific ablations and broader comparisons would be valuable, but our primary objective is joint probabilistic inference of reusable partial-order skill structures, rather than exhaustive evaluation of every modelling variant. We will clarify the existing evidence and limitations in the revision.

Thank you again for your feedback. We have incorporated the clarifications discussed above into the revised uploaded manuscript:

  • Q1 — Retail prediction: Section 5.3 now explicitly acknowledges the predictive gap, gives the correct baseline values, and distinguishes predictive accuracy from structural inference (lines 415–419).

  • Q2 — Complexity and convergence: The main text now states the forward-pass complexity and clarifies the limits of finite-run MCMC conclusions (lines 290–294), with detailed costs in Appendix H. The Retail fitting budget is documented in lines 1608–1612, and its sampling diagnostics and interpretation appear in lines 1855–1860.

  • Q3 — Fixed K: We added discussion of K as a representation-granularity choice (Appendix G.1, line 1619 and lines 1652–1653) and explicitly state that automatic K selection is not implemented (Limitations, lines 503–507).

  • Q4 — Modelling choices and existing evidence: The revised discussion explains the validity-ablation trade-off (lines 357–366), the Full-Marg/Full-Cond comparison (lines 317–323), and the no-self-transition modelling restriction (lines 193–205). Broader-domain and annotation limitations are stated in lines 503–504. These revisions clarify the existing evidence; they do not represent additional Retail hyperparameter-sensitivity experiments.

Reviewer_bt4xOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper does process mining on agent traces. It first uses CPA to rewrite each agent run as a sequence of standardized step labels. A Bayesian model then induces an SOP-like flowchart from these labels, consisting of a fixed number of reusable sub-procedures, the partial order within each and the transitions between them. The paper calls these sub-procedures "skills", but they are neither the skill of agent research nor the skill of reinforcement learning. They are the paper's own construct, a latent partial order over action labels rather than an executable unit, which may confuse readers from those areas. It claims that, unlike prior work, it both segments traces into reusable pieces and recovers the partial order within each piece. The claims are tested on synthetic traces generated by the model, on TaskBench-derived workflows with known structure and on τ³-Retail agent traces. An executor built on the learned Retail library is also reported to use far fewer tokens than a stepwise agent on 120 synthetic Retail tasks.

Critical strengths
  1. The problem is a real one, and a partial order per skill is a fitting way to express the many valid ways of solving one task. On TaskBench, RPOS predicts unseen valid orderings better than the tested total-order and sequence models. I think this direction deserves attention.
  2. The model and its inference are careful. The validity state lets an acyclic partial order explain alternative orders, repetition and repair, and summing over segmentations when scoring a structure recovers the generating library on synthetic data.
Critical weaknesses
  1. The method is demonstrated in only one real setting, and there it does not predict well.
    • On synthetic data the model generates the traces itself, and TaskBench is simulated from known workflows, so good results there are easy to understand. Real agent traces come only from τ³-Retail (L483).
    • On Retail, RPOS predicts worse than simple sequence models. Its final-test NLL is 1.600 at K = 5, against 1.303 for a bigram (L399-400).
    • The library in Figs. 4 and 7 comes from one run, s102, selected after fitting (L391-392), so the only real-trace result is a reading of one chosen run.
    • The motivation concerns real agents, so this weighs most in my rating. Question 2 asks whether the displayed arrows recur across runs.
  2. The token savings are measured for the whole executor, so they do not show that the learned library helps an agent.
    • Besides the library, G uses 14 prepared API templates, business-stage prerequisites and policy checks (App. J.1, L462-463).
    • The 120 test tasks are new parameter draws from the 23 task families used during development (L420).
    • The executor is the paper's only test of whether the library helps an agent, which is what the motivation asks for. Question 3 would isolate the library's share.
Questions for authors
  1. [Clarification] Which agent produced the τ³-Retail traces, and do they come from more than one agent?
  2. [Requested analysis] For each arrow in Fig. 4, how often does it pass the 0.70 threshold in the other three K = 5 runs, after matching skills across runs? If most arrows recur, I would read Fig. 4 as a property of the traces rather than of one run.
  3. [Requested analysis] What are success and token use when G runs on the same 120 tasks with the library replaced by an antichain, keeping everything else fixed? Which decisions in G come from the library rather than from the templates and prerequisites? If success drops or token use rises, I would raise my rating.
  4. [Clarification] The paper calls these constructs skills, but they are neither the skill of agent research nor the skill of RL. Could you define the term once, early in the paper, and say there that a skill here is a partial-order template over labels rather than an executable unit?
4 comments

We thank the reviewer for the constructive feedback and specific suggestions. We have conducted additional analyses of structural stability and downstream execution to address these concerns.

Q1. Provenance of the τ³-Retail traces

We use publicly released τ³-Retail trajectories. The traces therefore originate from multiple runs of the same underlying agent model, rather than different agent models. We verified the source provenance against the original manifest and SHA-256 hashes. Separately, a 160-trace CPA relabelling pass used Claude Code (claude-opus-5) as an annotator applying our frozen 24-label CPA dictionary.

Q2. Cross-run stability of the learned precedence relations Thank you for this suggestion. We matched skills across the four independent K=5 runs using their occurrence assignments and examined the 15 displayed precedence relations. Of these, 5 exceed the 0.70 threshold in all three other runs, 2 in two other runs, and 8 in one other run. These results indicate that several local dependencies recur across independent fits, while the complete inferred structures remain sensitive to the fitted run.

Q3. Contribution of the learned library to downstream execution

We appreciate this important question and would like to clarify the central contribution of our work. RPOS primarily addresses the inference of useful, reusable dependency structures from unsegmented agent traces, rather than the design of an execution system. Previous work, including LLMCompiler (Kim et al., 2024) and ReWOO (Xu et al., 2023), motivates the efficiency benefits of structured execution and reducing repeated LLM reasoning. Building on this motivation, RPOS tackles the more challenging inference problem of jointly discovering skill boundaries, assignments, and shared partial-order relationships while accommodating alternative action orderings, repetition, and repair. Our synthetic and TaskBench experiments directly evaluate these structural inference properties.

We agree that the original 86.97% reduction in online token use reflects the complete execution system, including learned skills, tool grounding, prepared templates, business prerequisites, and runtime control, rather than an isolated effect of learned precedence relations.

An important distinction is that learned partial-order relations are not executable procedures on their own. They specify dependencies between canonical procedural-action labels, but execution necessarily requires action grounding, task interpretation, state management, and runtime control. Thus, a standalone executor based solely on learned relations is not well-defined without these additional components. Nevertheless, the marginal effect of learned edges can be assessed within a fixed executor. The original executor could not isolate a downstream benefit from learned precedence relations.

To investigate the value of the learned representation in a setting where it can actively guide execution, we developed a new shared executor supporting multi-action planning, deterministic execution, and bounded repair. Both conditions used identical tools, common prompt templates, safety checks, controller logic, and execution budgets. Full K5 used the learned skill structures, within-skill precedence relations, and skill transition probabilities ((H,\pi,P)), whereas Flat-24 retained the same 24 CPA labels without learned procedural memory.

On 95 goal-only Retail tasks across 18 families, we obtained:

MetricFull K5Flat-24
Strict success68/95 (71.58%)59/95 (62.11%)
Mean tokens/task9,534.5211,292.49
Mean model calls/task3.4004.032

Full K5 used 15.57% fewer online tokens, corresponding to a paired reduction of 1,757.98 tokens per task. Its observed strict success rate was 9.47 percentage points higher. We also observed 107 action selections affected by learned precedence filtering, confirming that the revised executor actively uses the inferred structure.

We emphasise that the central scientific contribution of RPOS is the joint inference of reusable partial-order skill structures, not a particular execution policy. Efficient execution motivates this inference problem and illustrates its potential applications, but the proposed probabilistic framework is independent of any specific executor. We will revise the manuscript to clarify this distinction and the evidential role of each experiment.

Q4. Definition of a skill.

We agree that the terminology should be clearer. In RPOS, a skill is a reusable partial-order template over canonical procedural-action labels, rather than an executable program or a reinforcement-learning option. Different instances of the same skill may follow different valid action orderings, while instance-level validity states account for repetition and repair. We will introduce this definition explicitly in the revised introduction.

Thank you for the response and the extra experiments. I appreciate your clarifications on the trace source, the skill definition, and the fact that the original executor G cannot isolate the learned precedence relations. However, my main concerns remain. Most of the edges in Figure 4 still do not recur consistently across runs, and the new Flat-24 baseline drops skill grouping and transitions along with within-skill edges rather than isolating the partial order itself. In addition, no revised PDF has been uploaded to include these clarifications or tone down the original executor claims, so I will keep my score at 2.

Dear Reviewer,

Thank you for your follow-up. We have uploaded the revised PDF after summarising all suggestions. We would first clarify our central contribution. The segmentation is learned together with the relationships, rather than supplied in advance. Therefore, for your two further concerns:

1. Cross-run consistency.

Because segmentation is inferred, the same observed action order can span different skill instances across runs. Figures 9–10 and Table 15 in Appendix G.2 document how these boundaries change the scope of a within-skill prerequisite. This provides a concrete mechanism for some cross-run differences, without establishing identical complete graphs. Figure 9 now reports continuous precedence frequencies, and there are two states from the four seeds, different MCMC runs may explore different combinations of these latent variables. In particular, different segmentations and skill assignments may lead to different inferred partial-order structures.

Figure 10 illustrates this on the same observed action sequence: the predecessor and successor belong to one instance in one fitted decomposition and separate instances in another (p. 34, lines 1808–1812 and 1823–1833). The observed temporal order is unchanged, while the scope of a within-skill prerequisite changes. These findings identify recurring local dependencies and document how segmentation interacts with the remaining structural variation; they do not establish identical libraries across runs. The main-text interpretation appears in Section 5.3, lines 401–408

2. Isolating learned order.

The original comparison with Flat-24 was designed to evaluate the learned library (The whole statistical inference part) as a whole, including the inferred skill decomposition, transitions, and within-skill partial-order relations. To address your concern, we have now added an H=0H=0 antichain intervention, holding all non-HH parameters and executor components fixed. On 95 development-exposed tasks, Full K5 achieves 83/95 strict successes, compared with 55/95 for the antichain condition, while using 17.67% fewer online tokens (Table 20; main-text lines 482–485).

We also clarify the interpretation of this intervention. In this adapter, HH distinguishes the skill-conditioned action models. Removing HH makes these models coincide, so the retained π\pi and PP no longer distinguish action predictions, even though their values remain fixed. The comparison therefore measures HH’s contribution through the fixed execution architecture, including its role in differentiating skills for composition (Table 20 caption; lines 2094–2097). The cohort construction and development exposure are documented in Appendix J.3, lines 2138–2145.

3. Revised PDF and original execution claim.

The revised abstract attributes the 86.97% reduction to the integrated executor (lines 024–026). Section 5.4 identifies its prepared components and explicitly states the scope of this system-level comparison (lines 464–480). The new antichain experiment is reported separately.

For completeness, we also address your original four questions and where we put into the uploaded paper:

  • Q1 — Trace source: Qwen3.6-35B-A3B agent model, clarified in lines 385–389.
  • Q2 — Arrow recurrence: Of the 15 relations, 5 pass the 0.70 threshold in all three other runs, 2 in two runs, and 8 in one. Continuous frequencies and segmentation diagnostics appear in Appendix G.2, Figures 9–10.
  • Q3 — Answered above
  • Q4 — Skill definition:** The introduction now defines a skill as a reusable procedural structure represented by a partial order over CPA labels (lines 040–047). It specifies procedural relationships; it is not itself an executable unit.

We would appreciate your reconsideration based on the revised manuscript and the added matched ablation.

Thank you for uploading the revised PDF and adding the H = 0 ablation. Although setting H = 0 in this adapter also collapses skill differentiation alongside within-skill ordering and real-trace stability across runs remains limited, the revised manuscript properly scopes the original executor claims, defines skills early, and documents the new experiments clearly. Given these revisions, I am raising my score to 3.

Reviewer_i1xcOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

Overall, I feel that the paper is a good work with complete experiments, mathematical proofs, and analysis. The paper introduces the "Reusable Partial-Order Skills" (RPOS) model, which can jointly infer reusable skills, the boundaries of skill instances, and the internal sequential dependencies among actions from agent trajectory sequences labeled with action tags.

  1. By incorporating a partial-order structure, the model captures essential dependencies while allowing independent actions to be executed in varying sequences.
  2. The authors use a semi-Markov recursion approach, summing over all possible paths when evaluating structural configurations. Given a set of global parameters, skill paths can be sampled precisely from the conditional distribution.
  3. Experiments using synthetic data and the TaskBench simulation environment validate the model’s capability in structural recovery and its ability to predict valid, previously unseen action sequences. In a set of 120 synthetic retail tasks, based on the learned skill library, the agent can reduce the number of tokens required during the online phase by 86.97% compared to an agent executing actions step-by-step.
Critical strengths

The core strength of this paper lies in its joint approach to "skill segmentation" and “the learning of internal sequential dependencies”. For example, Figure 1 shows actions B and C can be executed in any order while remaining within the same skill template. Figure 5 illustrates how the model can represent repeated actions and subsequent corrective processes without altering the underlying skill templates.

Furthermore, when evaluating candidate structures, the model accounts for all possible segmentations and skill assignments. Comparative experiments on synthetic data clearly demonstrate that this "path-marginal” approach improves the accuracy of structural recovery.

Critical weaknesses

My primary concern is the writing and structure of the paper; I found it difficult to get the key points. The writing is straight forward, and I am not very excited about the proposed method. It is also hard for me to get the key conclusion, as the paper addresses three distinct questions: whether skill structures can be learned, whether action prediction can be improved, and whether execution costs can be reduced. The evidence supporting these arguments is derived from distinct experimental setups. Consequently, the current experimental results do not clearly isolate the specific downstream benefits attributable to the learned partially ordered skill relationships.

Questions for authors
  1. Would it be necessary to conduct an ablation study? For example, keeping all other components constant, to specifically evaluate the contribution of the "learned sequential ordering relationships" to the reported 86.97% reduction in online token consumption?

  2. The resolution of Figure 8 is poor.

3 comments

We thank the reviewer for this constructive feedback. Figure 8 has been updated and will uploaded as a new pdf later. And about other key arguments:

Central contribution and experimental coherence. Our primary contribution is a probabilistic framework for jointly inferring reusable skill boundaries, assignments, and shared partial-order structures from unsegmented agent traces. The three experimental settings provide complementary evidence for this central contribution:

  1. Synthetic experiments evaluate whether the proposed inference procedure can recover latent skill structure when ground truth is available.
  2. TaskBench experiments evaluate whether the learned partial-order representation can capture previously unseen valid action orderings, compared with total-order and sequence-based alternatives.
  3. Retail experiments examine the interpretability and predictive limitations of inferred skill libraries on agent-generated traces, together with their integration into a downstream execution system.

Regarding the 86.97% reduction in online token use, this is a result for the complete execution system. The inferred partial orders provide a basis for organising workflows, but turning them into executable procedures also requires tool grounding, business constraints, and implementation choices. Removing edges from an already constructed executor therefore does not fully address the role of the learned representation in building that system.

Therefore, we have also conducted an anti-chain experiment under a new Executor, see response for bt4x below. To evaluate the learned representation more directly, we developed a shared executor supporting multi-action planning, deterministic tool execution, and bounded repair. Both conditions used the same tools, base prompt template, safety checks, controller, and execution budgets, differing in procedural memory. Full K5 used 15.57% fewer online tokens, Its observed strict success rate was 9.47 percentage points higher.

Actually, our focus is on learning reusable partial-order skill structures, rather than developing an optimal executor to fully use the partial order. We will revise the manuscript to distinguish the structural inference contribution, predictive evaluation, and downstream application more clearly.

Thanks for your response. I maintain my rating.

Thank you for your feedback and suggestions. We have revised the manuscript and uploaded the newest PDF:

  1. Contribution and experimental coherence: We clarify that RPOS jointly infers skill boundaries, assignments, and within-skill partial orders (lines 040–053 and 071–075). The complementary roles of synthetic, TaskBench, and Retail experiments are explained in lines 299–303.

  2. Figure 8: We have redrawn the figure with a shared CPA row order, skill columns, and explicit normalisation and usage percentages. The updated caption appears in lines 1702–1705, with interpretation in lines 410–412.

  3. Execution claim: The 86.97% token reduction is explicitly attributed to the integrated executor, including its prepared components (Abstract, lines 024–026; Section 5.4, lines 464–480).

  4. Additional ablation: We added an H=0H=0 comparison holding all non-HH parameters and executor components fixed. On 95 development-exposed tasks, Full K5 achieves 83/95 versus 55/95 strict successes, with 17.67% fewer tokens (lines 482–485; Table 20). The intervention and its interpretation are detailed in Appendix J.2, lines 2084–2097.

Comments on the paper

1 comment

To all reviewers, thank you again for your reviews. We've posted detailed responses to each review. In brief, we have:

  • Run a executor comparison (Reviewers i1xc and bt4x). We tested whether the learned library helps on its own. Using a shared executor, we compared the full learned K=5 library with a flat 24-label baseline on 95 Retail tasks. Everything else was held fixed. The learned library had higher strict success (71.58% vs 62.11%) and used 15.57% fewer tokens per task.

  • Clarified the main contribution and the skill definition (Reviewers i1xc and bt4x). Each experiment now has a clearly stated role. We define a skill as a reusable partial-order template, not an executable unit.

  • Corrected and clarified results (Reviewer QedV). On Retail, the 1.303 NLL belongs to the bigram baseline; SEQ-HSMM scores 1.305, However the main goal is to get a explicit structure from agent traces. We also pointed to the complexity analysis (App. H.1), and described the existing K sensitivity checks and ablations.

We'd be grateful if you could look at our responses and tell us whether they address your concerns. If anything is still unclear, we're happy to provide more explanation or analysis before the discussion period ends.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.