Pith. sign in

REVIEW 4 major objections 3 minor 49 references

Composable skill retrieval succeeds by graph traversal, not isolated ranking.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 08:47 UTC pith:7DTSROR7

load-bearing objection SkillTrace is a coherent integration of known pieces, but the dependency graph that drives its main gain is unvalidated and internally inconsistent; the 12.95-point ablation attributed to dependency traversal could just be undirected lexical expansion. the 4 major comments →

arxiv 2608.02356 v1 pith:7DTSROR7 submitted 2026-08-03 cs.AI

SkillTrace: Traversing a Query-Skill Graph for Composable LLM Agents

Yue Yao , Shengyuan Wang , Xin Chen , Minke Zhang , Jia He , Bingjun Luo , Tom Gedeon This is my paper
classification cs.AI
keywords skill retrievalLLM agentsquery decompositionskill dependency graphpersonalized PageRankcomposable skillsgraph traversalSkillsBench
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

SkillTrace argues that when an LLM agent must pull together several reusable skills, the standard approach—score every skill against the query and take the top few—misses the structure that makes a skill set executable. The paper's claim is that the job is better modelled as walking a three-layer graph: decompose the user query into atomic skill queries, match each atomic query one-to-one to a primary skill, then propagate through a library-wide dependency graph to pull in prerequisite and supporting skills. On SkillsBench the method reaches 53.17% task success and on ALFWorld 91.43%, outperforming independent retrieval and earlier graph-propagation baselines. A sympathetic reader would care because it gives a concrete mechanism for turning 'retrieve relevant skills' into 'retrieve the complete, executable skill bundle,' and it shows the gains hold across five different backbone language models.

Core claim

On its own terms, the central discovery is that jointly modelling query composition, query–skill correspondence, and inter-skill dependencies—and searching that structure in stages—yields better executable skill compositions than treating skills as independent candidates. SkillTrace decomposes the user query into a hierarchical tree whose leaves are atomic skill queries, solves a maximum-weight bipartite matching to assign each atomic query exactly one primary skill, and then runs a reverse-aware personalized PageRank over a directed skill-dependency graph, seeded by those primary skills, to add the top K supporting skills. The combination is reported to raise task success rate to 53.17% on

What carries the argument

The Query–Skill Graph, a three-part structure: a hierarchical query tree whose leaves are atomic skill queries; a weighted bipartite graph connecting atomic queries to candidate skills, resolved by maximum-weight matching; and a directed skill-dependency graph whose edges are declared with weight wD_ij = overlap(O_i, I_j), kept where the overlap meets threshold delta_D = 0.6. The workhorse is reverse-aware personalized PageRank: seed the graph with the primary skills and propagate relevance through both dependency directions so prerequisites and complementary skills are pulled into the final bundle S*_x = P_x ∪ TopK.

Load-bearing premise

The load-bearing premise is that overlap between a skill's declared output schema and another skill's declared input schema—kept at threshold 0.6—correctly represents real execution dependencies, so that propagating over this graph actually finds the skills needed to execute.

What would settle it

Compare SkillTrace's retrieved skill bundles against expert-annotated ground-truth required-skill sets on SkillsBench (or a fresh sample). If the dependency-augmented bundles are not more complete—higher recall of prerequisites with at least equal precision—than the top-K independently ranked skills, then the graph traversal is not what carries the gain. A cheaper check: rerun SkillTrace with the dependency graph randomly rewired (same degrees); if success stays near 53%, the edge semantics are irrelevant.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If SkillTrace is right, skill retrieval systems should stop scoring skills independently and instead recover prerequisite chains, because adding the top-ranked skills alone underperforms dependency-augmented bundles (the ablation removing the dependency subgraph drops success by 12.95 points).
  • Query decomposition into atomic skill queries is load-bearing: without it the success rate drops 6.68 points, which implies that a single-vector query match misses distinct requirements.
  • One-to-one bipartite matching prevents several atomic queries from collapsing onto the same skill; removing structured query–skill assignment costs 9.22 points.
  • The dependency graph is cheap to reuse: it is built offline once, and online retrieval is O(QN + Q^2N + K(N+E)), roughly linear in library size, so the mechanism scales to large skill libraries.
  • Generalization across five backbone models suggests the retrieval structure, not a particular LLM's reasoning strength, drives the improvements.

Where Pith is reading between the lines

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

  • The schema-overlap rule for dependency edges is the part most likely to be improvable; a natural extension is to learn the dependency graph from execution traces or tool-invocation logs, which would likely strengthen PPR seeding.
  • The reverse-PPR traversal can be read as recovering prerequisite chains not present in the query; a testable extension is to verify on tasks whose official solutions name hidden prerequisites whether SkillTrace retrieves exactly those skills.
  • Because the method returns a bundle independent of the generator, the same graph could be used to audit or explain an agent's composition: the retrieved subgraph is an inspection artifact, not just a prompt.
  • The decomposition step is delegated to the backbone LLM; a plausible improvement is to make atomic queries themselves retrievable objects, so the query tree could be cached and reused across similar tasks.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 3 minor

Summary. SkillTrace proposes a three-level graph formulation for composable LLM skill retrieval: a hierarchical query tree that decomposes a user query into atomic skill queries, a bipartite graph that matches atomic queries to candidate skills via maximum-weight matching, and a skill dependency graph over which personalized PageRank propagates relevance from primary skills to supporting skills. The final retrieved bundle is the primary skills plus the top-K propagated neighbors. The paper reports state-of-the-art results on SkillsBench (53.17% success) and ALFWorld (91.43%), attributes large ablation gains to dependency-aware propagation, and shows consistent improvements over the GoS baseline across five backbone LLMs.

Significance. The paper addresses a timely and important problem: retrieval for composable skills should return a coherent, executable set rather than a ranked list of individually similar items. The three-level graph formulation is natural and the presentation includes useful analyses: cross-model generalization, per-category breakdowns, complexity bounds, and an ablation of each component. If the dependency graph genuinely encodes execution prerequisites, the framework would be a meaningful advance over independent ranking and prior propagation-based methods. However, two load-bearing aspects are not yet firmly established: the dependency graph construction is internally inconsistent in direction and unvalidated, and the main empirical claims rest on small, unreplicated performance differences. The contribution is plausible and worth revising, but the current evidence does not fully support the stated state-of-the-art claim.

major comments (4)
  1. [Method — Skill dependency subgraph, Eq. (7)–(8)] The direction of the dependency edge is described inconsistently. The text first says a directed edge (s_i, s_j) means 'executing skill s_i requires ... skill s_j' (consumer→provider), but the immediately following sentence and Eq. (7), wD_ij = overlap(O_i, I_j), define it as provider→consumer. Moreover, Eq. (8) symmetrizes the adjacency as A_D + (A_D)^T, so the edge direction is discarded entirely. Thus the method as implemented does not test directed prerequisite traversal; it performs an undirected, lexicon-driven expansion. Consequently, the 12.95-pp drop in the w/o SDS ablation (Fig. 5a) cannot be attributed specifically to recovering prerequisites. Please fix the direction convention, run the directed variant (or report symmetric vs. directed results), and reinterpret the ablation accordingly.
  2. [Eq. (7) and Appendix hyperparameters] The overlap function in Eq. (7) is under-specified: 'Overlap function measures whether input and output are matched' is a placeholder. More importantly, no validation is provided that schema-overlap edges correspond to true execution dependencies. Given that the SDS component drives most of the reported gain, the paper should at least define the overlap computation precisely and provide a sanity check — e.g., precision of predicted dependency edges against task traces or human judgments, or an ablation replacing the edge set with random/similarity-based edges of matched degree. Without this, the improvement may reflect adding lexically related skills rather than completing an executable dependency chain.
  3. [Table 1 and Experiment Setup] The headline success-rate differences are small relative to the stochasticity of LLM execution, and the paper reports no repeated runs, error bars, or significance tests. On SkillsBench, the improvements over the strongest baseline are one task in Core and one in Extreme; on ALFWorld the overall gain is 2/140 episodes (91.43% vs. 90.00%), and on Medium/Hard SkillTrace merely ties the best baseline. Given that each task is attempted once, these deltas may be noise. The authors should provide multiple seeds or a bootstrap over episodes, or temper the 'state-of-the-art' claim accordingly.
  4. [Appendix — Hyperparameters and experimental configuration] Key free parameters — dependency threshold δ_D=0.6, PPR restart α=0.2, supporting-skill budget K, and linking candidates per skill (8) — are fixed without sensitivity analysis. Because the SDS ablation is the largest contributor to the gain, the results should be shown to be robust to δ_D and K at minimum. Otherwise it is unclear whether the reported performance reflects a tuned configuration rather than the method's intrinsic behavior.
minor comments (3)
  1. [Method — Eq. (5) and matching paragraph] Eq. (5) defines a per-query argmax, while the next paragraph selects primary skills via maximum-weight bipartite matching. These are different selection rules; please clarify which one is actually used and whether Eq. (5) is simply illustrative.
  2. [Figure 5 (c), caption] 'on this yet' appears to be a typo for 'on this set.'
  3. [Experiment Setup — ALFWorld stratification] The Easy/Medium/Hard stratification is derived from SkillTrace's own query decomposition and reused for all methods. This may introduce method-specific bias in the category analysis; please clarify whether the decomposition is independent of the retrieval method or report a sensitivity check with a fixed decomposition.

Circularity Check

0 steps flagged

No significant circularity: SkillTrace is evaluated against external benchmarks and baselines; graph-construction concerns are threats to validity, not circular reasoning.

full rationale

The central claim is benchmark performance on SkillsBench and ALFWorld, measured by task verifiers outside the method's own equations. The formal objective in Eq. (1), S*_x = argmax R(x,S'), is a task-success objective, and SkillTrace does not fit any parameter to the reported success rates; hyperparameters such as δ_D=0.6, α=0.2, and K are stated in the appendix and applied uniformly. The dependency graph D_S is constructed from lexical overlap of input/output schemas (Eq. 7), and the ablation attributes a 12.95-pp gain to the SDS component; this is a possible validity concern if schema overlap does not capture true prerequisites, and the paper's direction language is inconsistent (one passage says (s_i,s_j) means s_i requires s_j, another says the output of s_i is consumed by s_j, and Eq. 8 symmetrizes the adjacency). However, these issues do not reduce the derivation to its inputs: the method's retrieved bundles are evaluated by external downstream task success, not by the overlap scores used to build the graph. Citations to prior work are baselines (GoS, SkillDAG, SkillsBench) rather than load-bearing self-citations or uniqueness theorems, and the paper itself flags its runtime correlation as 'descriptive evidence rather than a statistically conclusive correlation test.' The derivation chain is therefore self-contained against external benchmarks, and no circular step meeting the quoted-evidence standard is present.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 1 invented entities

The central claim rests on four modeling choices: the reliability of LLM query decomposition, the schema-overlap proxy for dependencies, PPR as a selection mechanism, and the one-to-one matching constraint. The first three are unvalidated assumptions that could significantly change the outcome; the numeric free parameters (δ_D, α, K, linking candidates) are set by hand without sensitivity analysis.

free parameters (4)
  • dependency threshold δ_D = 0.6
    Prunes dependency edges in Eq. (7). No sensitivity analysis is reported; the choice directly changes the skill graph and the propagated skill set.
  • PPR restart probability α = 0.2
    Balances seed relevance versus neighborhood propagation in Eq. (8). No sensitivity analysis is reported.
  • supporting-skill budget K = not fully specified (ALFWorld capped at 6; SkillsBench K not stated)
    Controls how many skills are added by Eq. (9). The missing value makes exact reproduction of the SkillsBench experiment impossible.
  • linking candidates per skill = 8
    Limits dependency graph density in the appendix. No sensitivity analysis is reported.
axioms (4)
  • domain assumption Every complex query can be decomposed into atomic skill queries, each fulfilled by one primary skill (Eq. 3, leaf constraints).
    The hierarchical prompt assumes a clean one-to-one mapping between subtask requirements and individual skills; malformed decompositions would propagate errors downstream.
  • domain assumption Lexical overlap of declared input/output schemas detects real execution dependencies (Eq. 7).
    The paper never validates this proxy against actual dependency ground truth, and the edge direction is described inconsistently in the text around Eq. (7).
  • domain assumption Personalized PageRank over a symmetrized dependency graph ranks supporting skills by utility (Eq. 8).
    The method assumes stationary relevance scores are a good surrogate for executable skill composition; no theorem or analysis connects PPR ranks to task success.
  • domain assumption One-to-one bipartite matching is the right way to assign primary skills (Eq. 5).
    The constraint that each skill is matched to at most one atomic query may discard cases where the same skill should serve multiple sub-requirements.
invented entities (1)
  • None no independent evidence
    purpose: No new physical, causal, or ontological entities are introduced.
    The 'Query-Skill Graph' is a data structure over existing skills and queries, not a new postulated entity requiring independent evidence.

pith-pipeline@v1.3.0-daily-deepseek · 12288 in / 11231 out tokens · 115278 ms · 2026-08-04T08:47:02.674421+00:00 · methodology

0 comments
read the original abstract

Large language model agents increasingly solve complex tasks by composing reusable skills from a library. To address this, the key challenge is not merely to retrieve individually relevant skills, but to identify a complete and executable skill composition. In this paper, we argue that this problem can be solved in a graph with three levels: compositional relations among skill queries, similarity between queries and candidates in the skill library, and the dependencies among the selected candidates. We introduce SkillTrace, which organizes the user query into a semantic hierarchy, matches skill queries and candidates, and propagates over the skill dependencies. Experiments on SkillsBench and ALFWorld demonstrate that SkillTrace achieves state-of-the-art performance, reaching a success rate of 53.17% on SkillsBench and 91.43% on ALFWorld. SkillTrace also delivers consistent improvements across different backbone language models, demonstrating the generality and robustness of graph-based skill retrieval.

Figures

Figures reproduced from arXiv: 2608.02356 by Bingjun Luo, Jia He, Minke Zhang, Shengyuan Wang, Tom Gedeon, Xin Chen, Yue Yao.

Figure 1
Figure 1. Figure 1: Motivation. We argue that independent skill re￾trieval may overlook skill dependencies, while SkillTrace connects atomic queries and skills through a query–skill graph, enabling dependency-aware traversal to retrieve a composable skill set. query should be decomposed into atomic skill queries while preserving their compositional relations. Second, matching skill queries to candidate skills requires both se… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of SkillTrace with an intuitive example. SkillTrace decomposes a complex query into atomic skill queries, performs bipartite query–skill matching to identify seed skills, and traverses the skill subgraph with personalized PageRank and selection constraints to retrieve a composable skill bundle. retrieve the supporting skills required for execution. We or￾ganize these relations into a Query-Skill G… view at source ↗
Figure 4
Figure 4. Figure 4: Category-wise performance of SkillTrace with five [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Additional analysis. (a) Overall success rates of the full SkillTrace and its variant ablations without the Hierarchical Query Tree (w/o HQT), Query-Skill Bipartite Graph (w/o QSBG), or Skill Dependency Subgraph (w/o SDS). (b) Numbers of successfully completed tasks on the Core, Extreme, and Extended subsets under each ablation setting. (c) Execution time and success rate of SkillTrace across five backbone… view at source ↗
Figure 6
Figure 6. Figure 6: Benchmark subset statistics used in the main ex [PITH_FULL_IMAGE:figures/full_fig_p009_6.png] view at source ↗

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

49 extracted references · 2 canonical work pages · 1 internal anchor

  1. [1]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Liu, Shubo and Zhang, Hongsheng and Qi, Yihong and Wang, Peng and Zhang, Yanning and Wu, Qi , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  2. [2]

    2026 , month = jul, howpublished =

    2026

  3. [3]

    arXiv preprint arXiv:2605.05726 , year=

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

    Pith/arXiv arXiv

  4. [4]

    arXiv preprint arXiv:2604.05333 , year=

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

    Pith/arXiv arXiv

  5. [5]

    arXiv preprint arXiv:2606.03056 , year=

    SkillDAG: Self-Evolving Typed Skill Graphs for LLM Skill Selection at Scale , author=. arXiv preprint arXiv:2606.03056 , year=

    Pith/arXiv arXiv

  6. [6]

    2: Pushing the frontier of open large language models , author=

    Deepseek-v3. 2: Pushing the frontier of open large language models , author=. arXiv preprint arXiv:2512.02556 , year=

    Pith/arXiv arXiv

  7. [7]

    2026 , month =

    Gemini 3.1 Flash-Lite Model Card , author =. 2026 , month =

    2026

  8. [8]

    5: Visual Agentic Intelligence , author=

    Kimi K2. 5: Visual Agentic Intelligence , author=. arXiv preprint arXiv:2602.02276 , year=

    Pith/arXiv arXiv

  9. [9]

    Agent Skills: Specification and Documentation for Agent Skills , year =

  10. [10]

    arXiv preprint arXiv:2602.12670 , year=

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

    Pith/arXiv arXiv

  11. [11]

    The twelfth international conference on learning representations , year=

    Toolllm: Facilitating large language models to master 16000+ real-world apis , author=. The twelfth international conference on learning representations , year=

  12. [12]

    Advances in Neural Information Processing Systems , volume=

    Gorilla: Large language model connected with massive apis , author=. Advances in Neural Information Processing Systems , volume=

  13. [13]

    Advances in neural information processing systems , volume=

    Retrieval-augmented generation for knowledge-intensive nlp tasks , author=. Advances in neural information processing systems , volume=

  14. [14]

    arXiv preprint arXiv:2010.03768 , year=

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

    Pith/arXiv arXiv 2010

  15. [15]

    Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    Dense passage retrieval for open-domain question answering , author=. Proceedings of the 2020 conference on empirical methods in natural language processing (EMNLP) , pages=

    2020

  16. [16]

    arXiv preprint arXiv:2305.16291 , year=

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

    Pith/arXiv arXiv

  17. [17]

    arXiv preprint arXiv:2603.02176 , year=

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

    arXiv

  18. [18]

    arXiv preprint arXiv:2603.04448 , year=

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

    arXiv

  19. [19]

    arXiv preprint arXiv:2605.12039 , year=

    SkillGraph: Skill-Augmented Reinforcement Learning for Agents via Evolving Skill Graphs , author=. arXiv preprint arXiv:2605.12039 , year=

    Pith/arXiv arXiv

  20. [20]

    arXiv preprint arXiv:2603.22455 , volume=

    Skillrouter: Retrieve-and-rerank skill selection for llm agents at scale , author=. arXiv preprint arXiv:2603.22455 , volume=

    Pith/arXiv arXiv

  21. [21]

    arXiv preprint arXiv:2302.07842 , year=

    Augmented language models: a survey , author=. arXiv preprint arXiv:2302.07842 , year=

    Pith/arXiv arXiv

  22. [22]

    arXiv preprint arXiv:2604.04323 , year=

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

    Pith/arXiv arXiv

  23. [23]

    arXiv preprint arXiv:2604.24594 , year=

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

    Pith/arXiv arXiv

  24. [24]

    International Conference on Learning Representations , volume=

    Toolllm: Facilitating large language models to master 16000+ real-world apis , author=. International Conference on Learning Representations , volume=

  25. [25]

    URL https://arxiv

    Gorilla: Large language model connected with massive apis, 2023 , author=. URL https://arxiv. org/abs/2305.15334 , year=

    Pith/arXiv arXiv 2023

  26. [26]

    Advances in neural information processing systems , volume=

    Toolformer: Language models can teach themselves to use tools , author=. Advances in neural information processing systems , volume=

  27. [27]

    and others , title =

    Gao, Y. and others , title =. 2025 , month = feb, eprint =. doi:10.48550/arXiv.2502.18041 , url =

    doi:10.48550/arxiv.2502.18041 2025

  28. [28]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

    Zhang, Weichen and others , title =. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =

  29. [29]

    and others , title =

    Xu, H. and others , title =. Pattern Recognition , volume =. 2026 , month = sep, doi =

    2026

  30. [30]

    and Huai, T

    Li, T. and Huai, T. and Li, Z. and Gao, Y. and Li, H. and Zheng, X. , title =. Proceedings of the IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) , year =

  31. [31]

    and others , title =

    Cai, H. and others , title =. 2026 , month = may, eprint =. doi:10.48550/arXiv.2601.03707 , url =

    Pith/arXiv arXiv doi:10.48550/arxiv.2601.03707 2026

  32. [32]

    Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

    Ning, Yuwei and Zhao, Ganlong and Qin, Yipeng and Liu, Si and Liu, Yang and Lin, Liang and Li, Guanbin , title =. Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) , year =

  33. [33]

    ImagineUAV: Aerial Vision-Language Navigation via World-Action Modeling and Kinodynamic Planning

    Liu, X. and Huang, J. and Xia, S. and Liu, B. and Cui, J. and Yang, J. , title =. 2026 , month = jun, eprint =. doi:10.48550/arXiv.2606.01205 , url =

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2606.01205 2026

  34. [34]

    The International Journal of Robotics Research , volume =

    Karaman, Sertac and Frazzoli, Emilio , title =. The International Journal of Robotics Research , volume =. 2011 , month = jun, doi =

    2011

  35. [35]

    Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , year =

    Mellinger, Daniel and Kumar, Vijay , title =. Proceedings of the IEEE International Conference on Robotics and Automation (ICRA) , year =

  36. [36]

    IEEE Robotics and Automation Letters , volume =

    Zhou, Xin and Wang, Zhepei and Ye, Hongkai and Xu, Chao and Gao, Fei , title =. IEEE Robotics and Automation Letters , volume =. 2021 , month = apr, doi =

    2021

  37. [37]

    and Zhan, Y

    Jiang, Y. and Zhan, Y. and Gupta, H. V. and Borde, C. and Geng, J. , title =. 2026 , month = mar, eprint =. doi:10.48550/arXiv.2504.04289 , url =

    doi:10.48550/arxiv.2504.04289 2026

  38. [38]

    and Wang, G

    Ryou, G. and Wang, G. and Karaman, S. , title =. The International Journal of Robotics Research , volume =. 2026 , month = apr, doi =

    2026

  39. [39]

    Venkateswara Rao, D. M. K. K. and Habibi, H. and Voos, H. , title =. Aerospace Science and Technology , volume =. 2026 , month = jul, doi =

    2026

  40. [40]

    and others , title =

    Qin, C. and others , title =. 2026 , month = mar, eprint =. doi:10.48550/arXiv.2603.04305 , url =

    work page doi:10.48550/arxiv.2603.04305 2026

  41. [41]

    and Truong, A

    Perille, D. and Truong, A. and Xiao, X. and Stone, P. , title =. Proceedings of the IEEE International Symposium on Safety, Security, and Rescue Robotics (SSRR) , year =

  42. [42]

    and others , title =

    Nair, A. and others , title =. Proceedings of the IEEE International Symposium on Safety, Security, and Rescue Robotics (SSRR) , year =

  43. [43]

    and Hauser, K

    Stratton, A. and Hauser, K. and Mavrogiannis, C. , title =. IEEE Robotics and Automation Letters , volume =. 2025 , month = jan, doi =

    2025

  44. [44]

    and Zheng, Y

    Yu, R. and Zheng, Y. and Qu, X. , title =. Transportation Research Part C: Emerging Technologies , volume =. 2021 , month = jun, doi =

    2021

  45. [45]

    and Li, B

    Peng, L. and Li, B. and Yu, W. and Yang, K. and Shao, W. and Wang, H. , title =. IEEE Transactions on Intelligent Transportation Systems , volume =. 2024 , month = feb, doi =

    2024

  46. [46]

    and others , title =

    Feng, T. and others , title =. IEEE Transactions on Pattern Analysis and Machine Intelligence , volume =. 2023 , month = jul, doi =

    2023

  47. [47]

    Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

    Beche, Radu and Nedevschi, Sergiu , title =. Proceedings of the IEEE/CVF International Conference on Computer Vision (ICCV) , year =

  48. [48]

    and Li, L

    Guo, X. and Li, L. and Yan, T. and Wang, L. and Qian, Y. , title =. Journal of Visual Communication and Image Representation , volume =. 2026 , month = jan, doi =

    2026

  49. [49]

    IEEE Transactions on Knowledge and Data Engineering , volume =

    Yang, Mingji and Wang, Hanzhi and Wei, Zhewei and Wang, Sibo and Wen, Ji-Rong , title =. IEEE Transactions on Knowledge and Data Engineering , volume =. 2024 , doi =

    2024