The reviewed record of science sign in
Pith

arxiv: 2607.06283 · v1 · pith:UIMBIRKT · submitted 2026-07-07 · cs.AI

Task Decomposition-Guided Reranking for Adaptive Agent Skill Retrieval

Yanping Chen , Weijie Shi , Wen Yang , Jiajie Xu This is my paper

Reviewed by Pith T0 review T1 audit T2 compute T3 formal T4 kernel 2026-07-08 10:52 UTCglm-5.2pith:UIMBIRKTrecord.jsonopen to challenge →

classification cs.AI
keywords skill retrievaltask decompositionrerankingLLM agentsexecution graphadaptive skill selectiondirected acyclic graphcross-encoder
0
0 comments X

The pith

Decompose the task, rerank the skills, let the graph decide

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

SkillReranker is an inference-time framework that selects which reusable skills an LLM agent should load for a given task. The core idea is to decompose both the task and each candidate skill into structured state descriptions—preconditions and completion states—and then build a directed acyclic graph where task states are nodes and skills are edges. This graph reveals which skills advance the task from one state to the next. The framework then detects natural stage boundaries within the graph—points where one set of skills finishes and another set begins—and reranks skills within each stage using a cross-encoder that combines overall task relevance with local sub-task relevance. The number of selected skills is determined by the task's own structure rather than a fixed top-k cutoff. On ALFWorld and ScienceWorld across three LLM backbones, SkillReranker improves task success rates, reduces environment interaction steps, and lowers token consumption compared to LLM-as-selector, SkillRouter, and Graph of Skills baselines.

Core claim

The central mechanism is the execution graph: a directed acyclic graph whose nodes are intermediate task states (produced by decomposing the task instruction into ordered subtasks and their resulting states) and whose edges are candidate skills (produced by parsing each skill into a precondition state and a completion state). A skill becomes an edge from the task state matching its precondition to the task state matching its completion. Stage boundaries are identified at nodes where incoming skill strength and outgoing skill strength both exceed the strength of skills that straddle the node (Eq. 6). Within each stage, skills are scored by the geometric mean of task-level and stage-level rer-

What carries the argument

Execution graph (DAG with task states as nodes, skills as edges); split-point detection (In/Out/Cross score comparison at each node); stage-wise cross-encoder reranking (geometric mean of task-level and sub-task-level relevance)

If this is right

  • Agents that dynamically select the number of skills based on task structure rather than a fixed top-k can reduce context length and token costs while improving task success rates.
  • Modeling skills as state transitions (precondition → completion) rather than as flat text descriptions provides a structured signal that disambiguates semantically similar but functionally different skills.
  • The split-point detection mechanism could generalize to other retrieval settings where the query has internal structure—any pipeline where a complex request decomposes into stages could benefit from stage-aware reranking.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The framework's reliance on LLM-generated task decomposition and skill-state parsing creates a quality bottleneck: if the decomposition misorders subtasks or omits a critical intermediate state, the execution graph will be misaligned and skill selection will degrade. The paper acknowledges this but does not measure decomposition error rates or their downstream impact, which would be a natural next
  • The split-point condition (In > Cross and Out > Cross) is a heuristic based on intuition about skill flow bottlenecks. An alternative formulation—e.g., directly optimizing stage boundaries for downstream task performance—might yield tighter partitions, though at higher computational cost.
  • Since skill parsing is done offline and cached, the framework's scalability depends on the quality of that one-time parsing pass. A skill library with poorly documented skills would produce unreliable precondition/completion states, limiting the graph's accuracy regardless of the reranking mechanism.

Load-bearing premise

The framework assumes that an LLM can reliably decompose arbitrary task instructions into correct ordered subtasks and sub-states, and that skill documents can be reliably parsed into accurate precondition and completion states. If the decomposition is wrong—a missing or misordered intermediate state—the execution graph will be misaligned, stage boundaries will be incorrect, and the selected skills will not match the task's actual execution flow.

What would settle it

If LLM-based task decomposition and skill-state parsing produce errors at a rate above some threshold, the execution graph would be structurally misaligned with the true task flow, and the stage-wise reranking would select skills for the wrong stages—degrading performance below simpler flat reranking baselines.

Figures

Figures reproduced from arXiv: 2607.06283 by Jiajie Xu, Weijie Shi, Wen Yang, Yanping Chen.

Figure 1
Figure 1. Figure 1: Task-matching difficulty caused by semantic simi [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed framework. (1) Structural Parsing decomposes the task into sub-tasks and states (online) and parses each skill into precondition/completion states (offline). (2) Execution Graph Construction recalls top-K candidate skills and matches their precondition/completion states to task states, turning each skill into an edge of a state-matching graph. (3) Adaptive Skill Selection detects s… view at source ↗
Figure 3
Figure 3. Figure 3: Average token consumption of different methods on Alfworld and ScienceWorld (lower is better). Each group along [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Two examples of skill retrieval by our method on [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
read the original abstract

Skill usage can significantly enhance the ability of modern agent systems to complete complex tasks. However, the growing scale of skill libraries makes accurate skill selection increasingly challenging. In real-world scenarios, ambiguous semantic matching often arises between a specific task requirement and multiple generic yet semantically similar candidate skills. Moreover, existing methods tend to overlook the dynamic influence of task difficulty and skill applicability when selecting the optimal target skill set. To address these issues, we propose SkillReranker, an inference-time reranking framework for adaptive skill selection. Specifically, we first perform semantic decomposition on both the task and skill sides, yielding informative subtask and execution-state descriptions as well as transition-state descriptions that characterize each skill's functionality. These descriptions are then used to construct a directed acyclic execution graph, where intermediate task states are modeled as nodes and candidate skills as edges, thereby establishing a structured task-skill correspondence. On this basis, SkillReranker determines whether each state node satisfies the split condition to identify subtask intervals. For each task interval, we employ a cross-encoder to perform comprehensive scoring over candidate skills and select the most suitable ones to form the final target skill set. Experiments on ALFWorld and ScienceWorld with three backbone LLMs show that SkillReranker effectively improves task performance, reduces environment interaction steps, and lowers token consumption compared with existing skill selection baselines.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 6 minor

Summary. The paper proposes SkillReranker, an inference-time reranking framework for adaptive skill selection in LLM-based agents. The method decomposes task instructions into ordered subtasks and intermediate states, parses skills into precondition/completion states, constructs a directed acyclic execution graph (task states as nodes, skills as edges), detects stage split points, and performs stage-wise cross-encoder scoring to select an adaptive number of skills. Experiments on ALFWorld and ScienceWorld with three backbone LLMs show improvements in reward, environment steps, and token consumption over three baselines (LLM-as-selector, SkillRouter, Graph of Skills). Ablations support the contribution of structured parsing, graph edges, and split-point detection.

Significance. The paper addresses a practical problem: selecting an adaptive number of skills from a large library rather than using a fixed top-k. The execution graph formulation, where skills are edges between task states, is a clean and intuitive design. The framework produces falsifiable predictions (adaptive skill counts per task, reported in Appendix A) and the ablation study isolates each component. The token consumption analysis is a useful addition. However, the significance of the results is tempered by the baseline comparison methodology (see Major Comment 1).

major comments (3)
  1. §5.1 (Evaluation paragraph): The baseline comparison methodology is a load-bearing concern for the central claim of consistent improvement. The paper states that baseline results are averaged over top-1 and top-2 skill settings, while SkillReranker selects ~1.3 skills per task. Averaging top-1 and top-2 performance is not equivalent to evaluating at 1.3 skills. If top-2 performance is systematically worse than top-1 (a phenomenon the paper itself acknowledges in §1, citing Liu et al. 2024 and Han et al. 2026 regarding context dilution and decision burden), then (perf_top1 + perf_top2)/2 < perf_top1, making baselines appear weaker than they are at their best operating point. The paper does not report top-1 and top-2 results separately, so it is impossible to verify whether baselines at top-1 are competitive with SkillReranker. This directly threatens the headline claim. The authors should
  2. §4.3, Eq. (5): The split-point detection heuristic (In(i) > Cross(i) and Out(i) > Cross(i)) is stated as an axiom without justification. The distance-based decay factor 1+(i−tgt_k)/(n−1) is introduced without explanation of why this particular form is chosen or whether it is robust to varying numbers of subtasks. Since the number of stages (and hence the number of selected skills) is entirely determined by this heuristic, its behavior is load-bearing for the adaptive selection claim. The paper should provide sensitivity analysis or at least report the distribution of detected split points across tasks to demonstrate that the heuristic produces reasonable stage partitions.
  3. §4.1: The framework relies on LLM-based task decomposition and skill-state parsing, acknowledged as a limitation in §7. However, the paper does not measure decomposition error rates or their downstream impact on graph construction and skill selection. The ablation in Table 2 (w/o Parsing) shows that removing structured parsing causes large performance drops, but this does not isolate the effect of incorrect decompositions. Since the execution graph, split-point detection, and stage-wise scoring all depend on correct decomposition, the paper should report decomposition quality (e.g., human or LLM-judge evaluation of subtask/state accuracy on a sample) and ideally analyze how decomposition errors propagate to skill selection errors.
minor comments (6)
  1. Table 1: The caption states 'Best in bold, second-best underlined,' but no underlines are visible in the rendered table. Please verify formatting.
  2. §5.2: The paper mentions 'SkillReranker shows particularly pronounced gains with DeepSeek-v4-Flash and GPT-5.4-Mini' but does not explain why gains are smaller with Qwen3.6-27B, where SkillReranker loses on ScienceWorld-seen reward and ScienceWorld-unseen steps. A brief discussion would help.
  3. Figure 3: The y-axis ranges differ between the two subfigures (DeepSeek: 40k–80k; GPT-5.4-Mini: 40k–100k), making visual comparison across models difficult. Consider using consistent ranges or separate axes with clear labels.
  4. §3: The skill library size N=67,884 is mentioned, but it is unclear how many skills are relevant to ALFWorld and ScienceWorld specifically. Reporting the number of candidate skills per task (before and after top-K filtering) would help readers assess the selection difficulty.
  5. References: Several references appear to be from 2026, which is unusual for a July 2026 submission. Please verify that all arXiv preprint identifiers and dates are correct.
  6. Appendix A, Table 3: For Qwen3.6-27B ScienceWorld-unseen, SkillReranker reports 76763 tokens, which is higher than Graph of Skills (50360) and LLM-as-selector (55616). This appears inconsistent with the claim in §5.2 that 'SkillReranker achieves the lowest average token consumption across all settings.' Please clarify or correct.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for a careful and constructive report. The three major comments all identify legitimate gaps that we will address in revision. Comment 1 (baseline comparison methodology) is the most consequential: we agree that averaging top-1 and top-2 baseline results is not equivalent to evaluating at ~1.3 skills, and we will report top-1 and top-2 results separately so the comparison is transparent. Comment 2 (split-point heuristic justification) and Comment 3 (decomposition quality measurement) both request analyses we can and will provide. We do not believe any comment is unanswerable.

read point-by-point responses

  1. Referee: §5.1: Baseline comparison methodology — averaging top-1 and top-2 baseline results is not equivalent to evaluating at ~1.3 skills. If top-2 is systematically worse than top-1 due to context dilution, then (perf_top1 + perf_top2)/2 < perf_top1, making baselines appear weaker than they are at their best operating point. The paper does not report top-1 and top-2 separately, so it is impossible to verify whether baselines at top-1 are competitive with SkillReranker.

    Authors: The referee is correct. Averaging top-1 and top-2 baseline performance is not equivalent to evaluating baselines at 1.3 skills, and if top-2 performance is systematically worse than top-1 (as the context dilution literature we cite suggests), then the average understates the baselines' best operating point. This is a genuine methodological weakness in the current manuscript. We will revise the paper to report top-1 and top-2 baseline results separately in all experimental settings, so that readers can directly compare SkillReranker against each baseline's best configuration. We will also add a column reporting the best baseline result (max of top-1 and top-2) alongside our method's result, ensuring the comparison is fair and transparent. If, as we expect, SkillReranker remains competitive or superior against each baseline's best individual setting, this will strengthen rather than weaken our claims; if not, we will honestly report where baselines at their best operating point match or exceed our method. We acknowledge that the current presentation does not allow the reader to verify this, and the revision will fix this. revision: yes

  2. Referee: §4.3, Eq. (5): The split-point detection heuristic (In(i) > Cross(i) and Out(i) > Cross(i)) is stated as an axiom without justification. The distance-based decay factor 1+(i−tgt_k)/(n−1) is introduced without explanation of why this particular form is chosen or whether it is robust to varying numbers of subtasks. Since the number of stages (and hence the number of selected skills) is entirely determined by this heuristic, its behavior is load-bearing for the adaptive selection claim. The paper should provide sensitivity analysis or at least report the distribution of detected split points across tasks.

    Authors: We agree that the split-point heuristic is load-bearing for the adaptive selection claim and that the current manuscript does not adequately justify the design choices or demonstrate that the heuristic produces reasonable partitions in practice. We will address this in two ways. First, we will add a distributional analysis of detected split points across all tasks in both benchmarks, reporting how many stages are detected per task, where split points fall relative to subtask boundaries, and how this correlates with task complexity. This will appear as a new table or figure in the revised paper. Second, we will add a sensitivity analysis varying the decay factor form (e.g., linear decay as currently formulated vs. no decay vs. quadratic decay) to show whether the split-point detection is robust to this choice. We chose linear decay for simplicity and because it provides a smooth, monotonic decrease in influence with distance from the node; the revised paper will state this rationale explicitly and report whether the results are sensitive to the specific form. revision: yes

  3. Referee: §4.1: The framework relies on LLM-based task decomposition and skill-state parsing, acknowledged as a limitation in §7. However, the paper does not measure decomposition error rates or their downstream impact on graph construction and skill selection. The ablation (w/o Parsing) shows large drops but does not isolate the effect of incorrect decompositions. The paper should report decomposition quality (e.g., human or LLM-judge evaluation of subtask/state accuracy on a sample) and ideally analyze how decomposition errors propagate to skill selection errors.

    Authors: This is a fair point. The w/o Parsing ablation demonstrates that structured parsing matters, but it does not isolate the effect of decomposition errors specifically. We will add a decomposition quality evaluation on a random sample of tasks (approximately 50 tasks per benchmark) using LLM-judge evaluation to assess subtask accuracy, state description accuracy, and ordering correctness. We will report the accuracy rates and include representative failure cases. For the error propagation analysis, we will identify tasks where decomposition errors occurred and compare the skill selection outcomes on those tasks against tasks with correct decompositions, reporting whether decomposition errors lead to downstream skill selection errors and performance degradation. We acknowledge that a fully controlled study (e.g., injecting synthetic decomposition errors) would be stronger but is beyond the scope of what we can complete in revision; the observational analysis on naturally occurring errors will still provide meaningful evidence about robustness. revision: partial

Circularity Check

0 steps flagged

No circularity found: the derivation chain uses off-the-shelf pre-trained models and parameter-free scoring equations, with no fitted constants or self-citation load-bearing steps.

full rationale

The paper's derivation chain is self-contained and does not exhibit circularity. The three main components are: (1) Structured Parsing (§4.1) uses an expert LLM (DeepSeek-v4-Flash) to decompose tasks into subtasks/states and skills into precondition/completion states — this is an LLM inference step, not a parameter fit. (2) Execution Graph Construction (§4.2, Eqs. 3–4) uses a pre-trained cross-encoder (Qwen3-Reranker-0.6B) off-the-shelf to compute relevance scores r(·,·) for aligning skill states to task states; no parameters are fitted by the authors. (3) Adaptive Skill Selection (§4.3, Eqs. 5–7) computes edge weights w_k = σ(ρ_k), split-point conditions (Eq. 6), and stage-fit scores (Eq. 7) — all parameter-free given the cross-encoder outputs. The log-summation in Eq. 7 is a fixed geometric-mean combination, not a tuned weighting. No self-citations are load-bearing: the reference list contains no prior work by the present authors (Chen, Shi, Yang, Xu). The skeptic's concern about averaging top-1 and top-2 baseline results is an experimental fairness issue (correctness risk), not a circularity issue — it does not make any 'prediction' equivalent to its inputs by construction. The method's outputs (selected skills) are computed from cross-encoder scores over candidate skills and decomposed task/skill states, none of which are fitted to the evaluation benchmarks.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The framework introduces no new entities (particles, forces, dimensions, etc.). It operates entirely within the existing paradigm of LLM-based agents with skill libraries. The free parameters (K=30, max steps=30) are standard experimental settings. The ad-hoc axioms (split-point heuristic, scoring formula) are the main theoretical contributions but lack independent validation.

free parameters (2)
  • K (candidate pool size) = 30
    The number of candidate skills retrieved for reranking. Set by the authors without systematic justification for this specific value.
  • Maximum environment steps = 30
    The maximum number of environment interaction steps allowed per task. A standard benchmark setting but still a parameter choice.
axioms (5)
  • domain assumption An expert LLM can reliably decompose arbitrary task instructions into correct ordered subtasks and corresponding sub-states.
    §4.1: The entire framework depends on the quality of LLM-based task parsing. No error analysis of decomposition quality is provided.
  • domain assumption An expert LLM can reliably extract accurate precondition and completion states from skill documents.
    §4.1: Skill parsing is performed offline by an LLM. The accuracy of this parsing is not measured or validated.
  • ad hoc to paper The split-point detection heuristic (In(i) > Cross(i) and Out(i) > Cross(i)) correctly identifies natural stage boundaries in task execution.
    §4.3, Eq. 5-6: The decay formulas and threshold condition are introduced without theoretical justification or empirical validation of boundary detection accuracy.
  • domain assumption Cross-encoder relevance scores between task states and skill states accurately reflect functional applicability of skills.
    §4.2-4.3: The method uses cross-encoder scores as the basis for graph construction and skill selection, assuming these scores capture genuine functional matching rather than surface-level semantic similarity.
  • ad hoc to paper The geometric mean of task-level and stage-level relevance probabilities (Eq. 7) is the appropriate combination strategy for skill scoring.
    §4.3, Eq. 7: The log-summation formula is presented as equivalent to a geometric mean, but no comparison with alternative combination strategies is provided.

pith-pipeline@v1.1.0-glm · 16059 in / 3189 out tokens · 374260 ms · 2026-07-08T10:52:40.999891+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

38 extracted references · 38 canonical work pages · 15 internal anchors

  1. [1]

    Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education

    Clancey, William J. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education. Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83)

    work page

  2. [2]

    Classification Problem Solving

    Clancey, William J. Classification Problem Solving. Proceedings of the Fourth National Conference on Artificial Intelligence

    work page

  3. [3]

    , title =

    Robinson, Arthur L. , title =. 1980 , doi =. https://science.sciencemag.org/content/208/4447/1019.full.pdf , journal =

    work page 1980

  4. [4]

    New Ways to Make Microcircuits Smaller---Duplicate Entry

    Robinson, Arthur L. New Ways to Make Microcircuits Smaller---Duplicate Entry. Science

    work page

  5. [5]

    and Rennels, Glenn R

    Hasling, Diane Warner and Clancey, William J. and Rennels, Glenn R. and Test, Thomas. Strategic Explanations in Consultation---Duplicate. The International Journal of Man-Machine Studies

    work page

  6. [6]

    Poligon: A System for Parallel Problem Solving

    Rice, James. Poligon: A System for Parallel Problem Solving

    work page

  7. [7]

    Transfer of Rule-Based Expertise through a Tutorial Dialogue

    Clancey, William J. Transfer of Rule-Based Expertise through a Tutorial Dialogue

    work page

  8. [8]

    The Engineering of Qualitative Models

    Clancey, William J. The Engineering of Qualitative Models

    work page

  9. [9]

    2023 , eprint=

    Attention Is All You Need , author=. 2023 , eprint=

    work page 2023

  10. [10]

    Pluto: The 'Other' Red Planet

    NASA. Pluto: The 'Other' Red Planet

    work page

  11. [11]

    Voyager: An Open-Ended Embodied Agent with Large Language Models

    Voyager: An open-ended embodied agent with large language models , author=. arXiv preprint arXiv:2305.16291 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  12. [12]

    SkillsBench: Benchmarking How Well Agent Skills Work Across Diverse Tasks

    SkillsBench: Benchmarking how well agent skills work across diverse tasks , author=. arXiv preprint arXiv:2602.12670 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  13. [13]

    How Well Do Agentic Skills Work in the Wild: Benchmarking LLM Skill Usage in Realistic Settings

    How well do agentic skills work in the wild: Benchmarking llm skill usage in realistic settings , author=. arXiv preprint arXiv:2604.04323 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  14. [14]

    arXiv preprint arXiv:2603.28815 , year=

    Skilltester: Benchmarking utility and security of agent skills , author=. arXiv preprint arXiv:2603.28815 , year=

    work page arXiv

  15. [15]

    2026 , eprint=

    SkillFlow: Scalable and Efficient Agent Skill Retrieval System , author=. 2026 , eprint=

    work page 2026

  16. [16]

    arXiv preprint arXiv:2603.22455 , year=

    Skillrouter: Skill routing for llm agents at scale , author=. arXiv preprint arXiv:2603.22455 , year=

    work page arXiv

  17. [17]

    Skill Retrieval Augmentation for Agentic AI

    Skill retrieval augmentation for agentic AI , author=. arXiv preprint arXiv:2604.24594 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  18. [18]

    SkillsInjector: Dynamic Skill Context Construction for LLM Agents

    SkillsInjector: Dynamic Skill Context Construction for LLM Agents , author=. arXiv preprint arXiv:2605.29794 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  19. [19]

    SkillReducer: Optimizing LLM Agent Skills for Token Efficiency

    Skillreducer: Optimizing llm agent skills for token efficiency , author=. arXiv preprint arXiv:2603.29919 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2602.19672 , year=

    Skillorchestra: Learning to route agents via skill transfer , author=. arXiv preprint arXiv:2602.19672 , year=

    work page arXiv

  21. [21]

    SoK: Agentic Skills -- Beyond Tool Use in LLM Agents

    SoK: Agentic Skills--Beyond Tool Use in LLM Agents , author=. arXiv preprint arXiv:2602.20867 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2603.04448 , year=

    Skillnet: Create, evaluate, and connect ai skills , author=. arXiv preprint arXiv:2603.04448 , year=

    work page arXiv

  23. [23]

    From Skill Text to Skill Structure: The Scheduling-Structural-Logical Representation for Agent Skills

    From skill text to skill structure: The scheduling-structural-logical representation for agent skills , author=. arXiv preprint arXiv:2604.24026 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Graph-of-Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills

    Graph of Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills , author=. arXiv preprint arXiv:2604.05333 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  25. [25]

    GraSP: Graph-Structured Skill Compositions for LLM Agents

    GraSP: Graph-structured skill compositions for LLM agents , author=. arXiv preprint arXiv:2604.17870 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  26. [26]

    arXiv preprint arXiv:2603.02176 , year=

    Organizing, orchestrating, and benchmarking agent skills at ecosystem scale , author=. arXiv preprint arXiv:2603.02176 , year=

    work page arXiv

  27. [27]

    Learning Generalizable Skills from Offline Multi-Task Data for Multi-Agent Cooperation

    Learning generalizable skills from offline multi-task data for multi-agent cooperation , author=. arXiv preprint arXiv:2503.21200 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  28. [28]

    Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Reinforcement learning for self-improving agent with skill library , author=. Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    work page

  29. [29]

    Agent Skills for Large Language Models: Architecture, Acquisition, Security, and the Path Forward

    Agent skills for large language models: Architecture, acquisition, security, and the path forward , author=. arXiv preprint arXiv:2602.12430 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  30. [30]

    arXiv preprint arXiv:2603.15401 , year=

    SWE-Skills-Bench: Do Agent Skills Actually Help in Real-World Software Engineering? , author=. arXiv preprint arXiv:2603.15401 , year=

    work page arXiv

  31. [31]

    Transactions of the association for computational linguistics , volume=

    Lost in the middle: How language models use long contexts , author=. Transactions of the association for computational linguistics , volume=

    work page

  32. [32]

    Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    SMART: Self-aware agent for tool overuse mitigation , author=. Findings of the Association for Computational Linguistics: ACL 2025 , pages=

    work page 2025

  33. [33]

    SkillRet: A Large-Scale Benchmark for Skill Retrieval in LLM Agents

    SkillRet: A large-scale benchmark for skill retrieval in LLM agents , author=. arXiv preprint arXiv:2605.05726 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  34. [34]

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models

    Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models , author=. arXiv preprint arXiv:2506.05176 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  35. [35]

    ALFWorld: Aligning Text and Embodied Environments for Interactive Learning

    Alfworld: Aligning text and embodied environments for interactive learning , author=. arXiv preprint arXiv:2010.03768 , year=

    work page internal anchor Pith review Pith/arXiv arXiv 2010

  36. [36]

    Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    Scienceworld: Is your agent smarter than a 5th grader? , author=. Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing , pages=

    work page 2022

  37. [37]

    2022 , date =

    Chase, Harrison , title =. 2022 , date =

    work page 2022

  38. [38]

    2026 , howpublished=

    MiMo-V2.5-Pro , author=. 2026 , howpublished=

    work page 2026