Pith.

Artificial Intelligence · arXiv:2610.32412 v1 · submitted 23 Oct 2026 · 26 references

LeafSkill: Efficient Leaf-Level Action Grounding for Recursive LLM Agents

Shen Liu, Ze Zhou, Zhenguo Xu, Shaopu Wang, Yike Gao, Chunlei Wang

The paper says

“Recursive large language model (LLM) agents solve long-horizon tasks through decomposition, execution, and environmental feedback. Yet after producing an atomic subtask, they often invoke the same large model again to translate a familiar instruction into an executable action.”

more of the abstract

We observe that high-level planning is task-specific, whereas leaf-level action grounding recurs across tasks and can be checked against environment-admitted actions. We introduce LeafSkill, an external reuse layer that retrieves semantically related grounding experience, binds action templates to current objects and admissible actions, and uses a lightweight 1.7B selector to choose a bound action or defer the unchanged subtask to the original agent. Skills remain outside the model parameters, allowing a frozen selector to use post-training additions. Under lexical and syntactic paraphrases, exact matching retrieves none of the reference actions, while semantic retrieval preserves 96.0% and 97.0% candidate coverage; the frozen selector adopts 97.34% of unseen skill types. On an independent confirmation set of 200 Robotouille tasks, LeafSkill completes 137 tasks versus 121 for ReCAP (68.5% vs. 60.5%, exact McNemar p=.0259). On the 106 tasks completed by both, it reduces Qwen3.5-27B tokens by 14.42% and calls by 14.08%. A separate held-out repeat yields a consistent 9.0-point success gain and 18.98% token reduction. Reusable leaf grounding can thus improve task completion while reducing host-model inference without replacing high-level planning.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 26 Oct 2026 · what was read was not recorded · not checked by a person

It can be wrong. Challenge this review or reply to it.

Holds up, with conditions. A credible, well-measured efficiency result for a genuinely narrower delegation boundary, with one real open edge in the ALFWorld adapter and a few secondary reporting gaps. This line is wrong

1 major objection · 4 minor

How often objections from reviews like this one are right: not yet measured

Are you an author? What you can do.

Not checked: figures, code, what the 26 references say.

The paper says, at Section 3, Environment Adapters and State-Bound Binding; Table 2

The review did not quote this passage.

Pith says: Objection 1 of 1 · major

The ALFWorld eligibility adapter is the gate for the paper's largest efficiency claim (41.32% host-token reduction), but the manuscript does not specify how the adapter separates an initial 'downward execution path' from recovery-driven movement. The text says the adapter excludes root, upward-return, and revisit phases, and that a node is not eligible 'when the node is revisited during recovery.'

more of objection 1

However, the host's recovery mechanism may return to a previously failed region of the task tree and descend again toward the failed leaf, and it is not stated whether such a re-descent counts as downward (admitted) or revisit (excluded). This ambiguity matters because the selector receives only the task node, local observation, admissible actions, and candidate descriptions, while the Table 3 guards enforce admissibility, not correctness. A legal-but-wrong action on an admitted recovery-context node can change whether a task is completed and, therefore, whether it remains in the jointly-solved intersection that defines the ALFWorld cost reduction and the success comparison. Table 9's input-scope diagnostic (93.3% vs. 90.7%) is reassuring about the local interface in general, but it is constructed from request-local ground truth and does not include recovery-context-dependent nodes, so it does not validate this adapter. I am not asking for a general solution: Section 5 already scopes LeafSkill to environments with an adapter. The revision should define the adapter's traversal state precisely, state the classification of recovery re-descents, and provide an adapter-level validation or sensitivity analysis, for example an audit of eligibility decisions on recovery paths or a comparison of success and cost under the two admission policies.

4 minor objections

The paper says, at Section 4, Table 7 and controller overhead

Pith says: minor objection 1 of 4

The controller overhead is reported in tokens and deliberately not summed with host tokens because the tokenizers differ. For the 'Efficient' framing, I would add a normalized cost comparison (measured throughput, GPU-seconds, or API-price per episode) so that a reader can see the net deployment effect; as written, the 6,900-9,952 controller tokens per episode are hard to compare with the host-token savings.

The paper says, at Section 4, Table 6

Pith says: minor objection 2 of 4

The McNemar p-values are reported per run without multiplicity adjustment. Since Group C is the independent confirmatory set, I suggest labeling it as the pre-specified test and describing the repeat-level B and ALFWorld p-values as descriptive, which would prevent readers from over-weighting the B R2 p=.025.

show 2 more

The paper says, at Section 4, Growth study and Table 10

Pith says: minor objection 3 of 4

Table 10's efficiency comparison uses only the 88 tasks solved by ReCAP and both library variants, and the 105-common-success comparison is another selected subset. The paper does report the success counts, but a sentence acknowledging the selection and reporting the overlap counts (or a union-of-successes sensitivity analysis) would make the growth-savings claim easier to interpret.

The paper says, at Global

Pith says: minor objection 4 of 4

Please add a code/data availability statement. The paper does not say whether the library construction, adapter implementations, and evaluation harness will be released; given the adapter concern above, releasing the eligibility-trace logs would substantially help verification.

What would settle it

Instrument the ALFWorld adapter to tag every admitted eligible node with its phase: initial downward descent vs. post-recovery re-descent (the unspecified case). Then, on all valid_unseen runs, re-run the full retrieve-bind-select pipeline on every admitted node under two observation conditions: local-only (task node, local observation, admissible set) vs. local-plus-root-goal-and-recovery-context appended to the observation. Count nodes where the controller's selected action differs between conditions. If the divergence rate is zero or negligible on tasks solved by both systems, the boundary holds and the jointly-solved claims are safe; a nonzero divergence on any admitted node demonstrates that the adapter admits nodes whose grounding depends on non-local context, which would force re-examining the ALFWorld efficiency figure and the success claim.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 26 Oct 2026

  1. Reader holds up, with conditions confidence high
  2. Skeptic agrees 35 s · 14,367 tokens read, 4,344 written
  3. Editor wrote the verdict above 16 s · 15,434 tokens read, 1,992 written
  4. Referee major revision: 1 major objection, 4 minor 2 min 14 s · 16,220 tokens read, 1,446 written

Not run on this paper: the simulated authors' defence.

Pith did not check this review's objections against the paper.

The review in full

The review in brief, machine-written

LeafSkill is worth taking seriously. The core idea is genuinely narrower than prior skill-reuse work: delegate only the final grounding choice over currently admissible actions, with deterministic binding, request-local candidate IDs, and explicit abstention. The evaluation is better executed than most agent papers: disjoint splits, nine cross-run pairings, bootstrap CIs on the independent set, paired-success cost alongside full-set success, and ablations that separate retrieval, binding, and selection. On those terms, the 23.1% Robotouille / 41.3% ALFWorld token reductions are credible as measurements, and the 14.4% independent result with p=0.0259 is the right headline.

The soft spots are real but not fatal. First, the ALFWorld eligibility adapter is the gate for every delegated decision and is not independently validated. The paper says the adapter admits only non-root nodes on the downward execution path and excludes revisit phases (Table 2), and scopes the method to environments with adapters (b179). What is unresolved is ReCAP's recovery behavior: when the planner re-descends toward a failed leaf, does that count as downward path or revisit? If it is admitted, recovery-context-dependent nodes can be mis-delegated; the deterministic guards prevent illegal actions but not wrong ones. The 93.3% local versus 90.7% plus-planner diagnostic is reassuring but uses request-local ground truth and does not test recovery-dependent nodes. This affects the ALFWorld efficiency and success claims because tasks that fail from mis-delegation drop out of the jointly-solved intersection. I would not call it a load-bearing flaw: the Robotouille results use a native atomic-leaf flag and stand on their own, and the authors are explicit about the adapter being a scope condition (b32, b179). But the re-descent edge should be specified and tested before the system is treated as deployment guidance.

Second, total deployment cost is not in the headline: controller tokens are reported separately (b137), which is sensible given differing tokenizers, but an end-to-end cost model that includes controller calls would strengthen the efficiency claim. Third, the success gain on Robotouille C is a single test with p=0.0259, so it should be read as suggestive rather than robust.

No code or artifacts are released, which I would want before citing the numbers as reproducible. Even so, the central claim holds up as an empirical claim on these benchmarks: repeated leaf grounding can be delegated to a constrained local selector without replacing planning or recovery. This paper deserves a serious referee and likely acceptance after revision.

Recommendation: send it to peer review; ask for the adapter behavior on recovery re-descents, an end-to-end cost comparison, and release of code and data.

The strongest objection

The strongest claim's efficiency figures (23.1% Robotouille, 41.3% ALFWorld, 14.4% independent) and the Robotouille success gain are all measured on tasks solved by both systems, and both depend on the delegation boundary being safe: a mis-delegated leaf produces a legal-but-wrong action (the deterministic guards in Table 3 prevent illegal actions, but not wrong ones), which can drop a task out of the jointly-solved intersection and thereby distort both the token-reduction and success numbers. The paper explicitly scopes this boundary to the environment adapter (b32, b179) and acknowledges ALFWorld has no native atomic-leaf flag, substituting a heuristic: non-root nodes on the 'downward execution path,' excluding root, upward-return, and revisit phases (Section 3, Table 2). The sharp edge is that ReCAP's recovery mechanism re-descends toward failed leaves; whether that re-descent counts as 'downward path' (admitted) or 'revisit' (excluded) is unspecified, and this is precisely the node class whose correct grounding may depend on recovery state or root goal. The input-scope diagnostic (Table 9, 93.3% local vs 90.7% plus-planner) is reassuring but is constructed with request-local ground truth and does not include recovery-context-dependent nodes, so it does not validate the adapter. Since the adapter is the gate on every delegated decision and is neither independently validated nor covered by the diagnostic, the efficiency/success claims inherit its correctness risk. This matches the reader's weakest_assumption; it is a partially acknowledged limitation (b179, Table 2) whose residual open problem—the re-descent phase ambiguity and any consequent mis-admission—remains unresolved.

The reader identified the environment adapter as the load-bearing component, and my read converges on the same point. The strongest claim is scoped precisely ('host tokens on jointly solved tasks'), so the headline is accurate as worded, and controller tokens are reported separately (Table 7) rather than concealed—an acknowledged scoping choice, not a hidden flaw. The genuinely unvalidated link is the eligibility adapter, which decides when a leaf task may be delegated. On Robotouille the adapter is native and low-risk; on ALFWorld it is a hand-written heuristic whose treatment of recovery re-descents is ambiguous in the text. A wrong admission yields a legal-but-wrong action that can remove a task from the jointly-solved intersection, corrupting both the token-reduction and success claims simultaneously. The paper's own ablation evidence (Table 9 input-scope diagnostic) supports the local interface only for request-locally-groundable nodes and does not cover recovery-context-dependent ones, so the boundary is not independently validated. This does not overturn the central claim—the measured reductions are real on the measured subset—but it justifies the CONDITIONAL verdict, with the concrete test above able to settle whether the boundary actually degrades the claims. No theatrical or personal critique is warranted; the authors state their scope and report controller overhead transparently, which deserves credit.

The referee's report

If the results hold, LeafSkill is a well-scoped and useful contribution: it identifies a measurable source of host-model waste (leaf grounding is 22.9-44.5% of host tokens in the source traces), constrains delegation to a finite admissible action set, and preserves planning and recovery in the host. The evaluation is stronger than typical for this area: disjoint construction/evaluation splits; three repeats crossed in all nine pairings; bootstrap CIs on the independent set; full-set success reported alongside paired-success cost; controlled diagnostics for paraphrase coverage and input scope; and ablations that separate retrieval, binding, and selection. The Robotouille success gain is significant on an independent set, and the paper is appropriately cautious about ALFWorld success. The main unresolved risk is the ALFWorld eligibility adapter, which is the gate for the largest efficiency claim and is not independently validated. I do not see circularity in the held-out evaluation, since Group C and valid_unseen are disjoint from all training and adaptation data.

Pith's reading of the paper (a machine's summary, not the authors')

Recursive LLM agents decompose goals into atomic subtasks but then often call the same large host model just to translate each subtask into one executable action; the paper measures that this leaf-level action grounding consumes 22.9–44.5% of host-model tokens. LeafSkill is an external reuse layer that stores these translations as canonical task–action templates, retrieves them semantically, binds them to the environment's currently admissible actions, and lets a frozen 1.7B controller either execute a bound action or abstain, returning the unchanged subtask to the host. With the ReCAP planner and a Qwen3.5-27B host untouched, the paper reports host-token reductions of 23.1% on Robotouille and 41.3% on ALFWorld on jointly solved tasks, plus an independent Robotouille set where LeafSkill completes 137 of 200 tasks versus 121 for the baseline (p = .0259). The claim matters because it isolates a recurring, verifiable operation — the final choice inside a finite set of legal actions — and shows it can be offloaded to a small model without replacing high-level planning.

What carries the argument. The load-bearing object is the retrieve–bind–select pipeline, which converts a leaf task into one element of the environment's current admissible set $A_t$ while keeping the host planner out of the loop. Retrieval (via the BGE-small-en-v1.5 embedding model) ranks canonical task keys and returns the top-$k$ above a similarity threshold; binding canonicalizes entities ($\texttt{onion1} \to \texttt{onion#}$), instantiates typed template slots from the current observation, and discards every command that is not an exact member of $A_t$; the selector, a frozen Qwen3-1.7B model, chooses among request-local candidate identifiers or $\bot$ (abstention), and every deterministic guard — empty retrieval, empty binding, unknown identifier, or a failed final admissibility check — returns control to the host. The three stages carry distinct load: exact-string reuse of the same task–action pairs changes tokens by only 0.20%, semantic retrieval supplies the paraphrase coverage (96.0% lexical and 97.0% syntactic under controlled rewrites, versus 0% for exact matching), binding enforces executability, and learned selection decides when reuse is appropriate. Request-local identifiers, reshuffled at every call, are what allow compatible skills to be added after training without retraining the selector.

Core claim. The paper's central claim is that leaf-level action grounding is task-recurring rather than task-specific: once the planner has produced a leaf task, the remaining decision is a narrow mapping from a local instruction to one currently admissible environment action, and that mapping can be reused across tasks and delegated safely. To establish this, the paper builds a retrieve–bind–select layer whose frozen library holds canonical-task/action-template pairs (called skills), whose retriever proposes top-k semantically related templates, whose deterministic binder instantiates typed slots with current-state entities and keeps only commands that exactly match an element of the admissible action set, and whose lightweight selector chooses among the resulting request-local candidates or abstains, with abstention preserving host state. The evidence is a sequence of contrasts: exact-string reuse changes host tokens by only 0.20% while semantic retrieval keeps 96–97% of reference actions covered under paraphrases that defeat exact matching; the frozen selector adopts 97.3% of held-out compatible skill types; and growing the library from 446 to 588 skills after training deepens fixed-set token savings from 18.3% to 25.6%. The authors conclude that state-bound action reuse is a complement to recursive LLM planning, not a replacement for it.

The premise it rests on. The load-bearing premise is that the environment adapter reliably identifies which task nodes are safely delegable leaves, because the ALFWorld adapter is a hand-written rule admitting only non-root nodes on the downward execution path; the authors acknowledge that new environments require corresponding adapters, but a node whose correct grounding depends on the root goal, recovery state, or ancestor context could, if wrongly admitted, lead the lightweight selector to pick an admissible-but-wrong action that no guard catches.

If the paper is right (Pith's inferences, not the authors')

On tasks solved by both systems, host-token use drops 23.1% on Robotouille and 41.3% on ALFWorld with no change to the ReCAP planner or the Qwen3.5-27B host, and observed success is numerically higher in both environments rather than traded away for savings.

The ablation ladder assigns each stage a distinct job: without semantic retrieval, exact reuse changes tokens by only 0.20%; semantic Top-5 retrieval lifts paraphrase coverage from 0% to 96–97% under lexical and syntactic rewrites; and the controller alone falls back correctly in only 7.5% of adversarial all-negative cases, implying that deterministic binding, not learned abstention, is the operat

A frozen selector can exploit compatible skills added after training: expanding the library from 446 to 588 skills deepens fixed-set host-token savings from 18.3% to 25.6%, with 97.3% adoption of held-out keys, so post-training library growth needs no selector retraining.

The delegation boundary transfers across environments with target adaptation: a Robotouille-initialized selector adapted on ALFWorld valid_seen reaches 91.79% mean success with 41.32% host-token reduction, and the measured controller overhead of 6,900–9,952 tokens per episode is reported separately from host tokens.

Because the selector cannot generate free-form commands and the environment interface only executes actions present in the admissible set, every executed action is legal by construction, which is what lets the paper report efficiency gains without a separate safety audit of the host's planning path.

Editorial inference: if the request-local candidate trick scales to much larger libraries, it sketches a general recipe for agent infrastructure — frozen small controllers plus deterministic environment guards as a stable substrate, with the expensive model reserved for genuinely novel reasoning.

Editorial inference: the paper's audit method (measuring the share of host tokens spent on leaf grounding) is itself transferable, and running the same measurement on other recursive or hierarchical agent frameworks is a cheap way to predict where a LeafSkill-style layer would pay off.

Editorial inference: the clean split between learned selection of 'which candidate' and deterministic binding of 'what is admissible' invites a stress test the paper does not run — shrinking or freezing the controller further should erode savings but not success, while weakening binding should erode both, which would confirm that the guards, not the 1.7B model, carry the safety.

Editorial inference: the stated scope of discrete admissible-action interfaces points to the natural next domain, such as robot manipulation with continuous or parametric actions, where the binding step would have to change from exact-string matching to constraint satisfaction.

Check this review

Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.

How often reviews like this one are right

Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.

How often objections from reviews like this one are right: not yet measured

Every value above, in the review's public record (JSON)

Reviews by people

This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.

Request a human review

How Pith checks papers · The full audit of this paper

The paper

Abstract

Recursive large language model (LLM) agents solve long-horizon tasks through decomposition, execution, and environmental feedback. Yet after producing an atomic subtask, they often invoke the same large model again to translate a familiar instruction into an executable action. We observe that high-level planning is task-specific, whereas leaf-level action grounding recurs across tasks and can be checked against environment-admitted actions. We introduce LeafSkill, an external reuse layer that retrieves semantically related grounding experience, binds action templates to current objects and admissible actions, and uses a lightweight 1.7B selector to choose a bound action or defer the unchanged subtask to the original agent. Skills remain outside the model parameters, allowing a frozen selector to use post-training additions. Under lexical and syntactic paraphrases, exact matching retrieves none of the reference actions, while semantic retrieval preserves 96.0% and 97.0% candidate coverage; the frozen selector adopts 97.34% of unseen skill types. On an independent confirmation set of 200 Robotouille tasks, LeafSkill completes 137 tasks versus 121 for ReCAP (68.5% vs. 60.5%, exact McNemar p=.0259). On the 106 tasks completed by both, it reduces Qwen3.5-27B tokens by 14.42% and calls by 14.08%. A separate held-out repeat yields a consistent 9.0-point success gain and 18.98% token reduction. Reusable leaf grounding can thus improve task completion while reducing host-model inference without replacing high-level planning.

The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.

abstract compiled · 2026-10-26

The authors' main claim, as the classifier picked it out 1.00

“Reusable leaf grounding can thus improve task completion while reducing host-model inference without replacing high-level planning.”

How the main claim is stated: hedged possibility 0.78

What the abstract itself states, as probabilities: claims first 0.06 · states limitation 0.25 · human participants 0.04 · controlled comparison 0.95 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.98 · mentions code or data availability 0.03

The abstract, sentence by sentence

This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.

SentenceRoleThe authors' own workStates the main conclusionStrength
Recursive large language model (LLM) agents solve long-horizon tasks through decomposition, execution, and environmental feedback. background 1.00 0.06 0.02 asserted 1.00
Yet after producing an atomic subtask, they often invoke the same large model again to translate a familiar instruction into an executable action. background 0.51 0.14 0.02 asserted 0.89
We observe that high-level planning is task-specific, whereas leaf-level action grounding recurs across tasks and can be checked against environment-admitted actions. gap 0.32 0.77 0.14 asserted 1.00
We introduce LeafSkill, an external reuse layer that retrieves semantically related grounding experience, binds action templates to current objects and admissible actions, and uses a lightweight 1.7B selector to choose a bound action or defer the unchanged subtask to the original agent. method 0.59 0.88 0.10 asserted 1.00
Skills remain outside the model parameters, allowing a frozen selector to use post-training additions. method 0.93 0.60 0.04 asserted 1.00
Under lexical and syntactic paraphrases, exact matching retrieves none of the reference actions, while semantic retrieval preserves 96.0% and 97.0% candidate coverage; the frozen selector adopts 97.34% of unseen skill types. result 1.00 0.93 0.06 asserted 1.00
On an independent confirmation set of 200 Robotouille tasks, LeafSkill completes 137 tasks versus 121 for ReCAP (68.5% vs. 60.5%, exact McNemar p=.0259). result 1.00 0.95 0.09 asserted 1.00
On the 106 tasks completed by both, it reduces Qwen3.5-27B tokens by 14.42% and calls by 14.08%. result 1.00 0.95 0.06 asserted 1.00
A separate held-out repeat yields a consistent 9.0-point success gain and 18.98% token reduction. result 1.00 0.94 0.07 asserted 1.00
Reusable leaf grounding can thus improve task completion while reducing host-model inference without replacing high-level planning. interpretation 1.00 0.72 0.92 hedged possibility 0.78

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.99

What was examined, and what was not

Sources read

  • arXiv listing abstract v1

Not examined

  • full text
  • figures and tables
  • references and citing sentences
  • relations to other papers

Levels run

  • P abstract

Not run

  • S sections
  • T body sentences
  • C citations
  • R cross-paper relations

How the labels were made

Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-26: 1 request, 7,739 input tokens.

Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON

Figures (4)
Figure 1 of the paper
Figure 1. Host-token composition of ReCAP trajectories used
Figure 2 of the paper
Figure 2. Overview of LeafSkill. At an eligible leaf task, retrieval recalls templates, binding instantiates admissible actions, and
Figure 3 of the paper
Figure 3. (95% CI [2.29, 19.49]) and calls by 10.36% relative to A. Group-C success is 131/200 with A+B and 137/200 with A (p = .5118). In the controlled selector diagnostic, inserting 124 held-out compatible keys yields 97.3% adoption and 98.9% retention. Together, these results show that a frozen selector can use compatible additions at read time. Controller Transfer and Adaptation
Figure 4 of the paper
Figure 4. Robotouille cross-run paired-success host-token

Figures are the authors', reproduced from the paper.

References (26)
  1. [1] Chain-of-Thought Prompting Elicits Reasoning in Large Language Models , author=. Advances in Neural Information Processing Systems , volume= on Pith
  2. [2] Yao, Shunyu and Zhao, Jeffrey and Yu, Dian and Du, Nan and Shafran, Izhak and Narasimhan, Karthik and Cao, Yuan , booktitle= on Pith
  3. [3] Tree of Thoughts: Deliberate Problem Solving with Large Language Models , author=. Advances in Neural Information Processing Systems , volume= on Pith
  4. [4] Zhang, Zhenyu and Chen, Tianyi and Xu, Weiran and Pentland, Alex and Pei, Jiaxin , booktitle=. 2025 , url= on Pith
  5. [5] Hu, Mengkang and Chen, Tianxing and Chen, Qiguang and Mu, Yao and Shao, Wenqi and Luo, Ping , booktitle= on Pith
  6. [6] Reflexion: Language Agents with Verbal Reinforcement Learning , author=. Advances in Neural Information Processing Systems , volume= on Pith
  7. [7] Zhao, Andrew and Huang, Daniel and Xu, Quentin and Lin, Matthieu and Liu, Yong-Jin and Huang, Gao , booktitle= on Pith
  8. [8] Voyager: An Open-Ended Embodied Agent with Large Language Models , author=. Transactions on Machine Learning Research , year= on Pith
  9. [9] Kim, Minsoo and Bursztyn, Victor and Koh, Eunyee and Guo, Shunan and Hwang, Seung-won , booktitle= on Pith
  10. [10] Agent Workflow Memory , author=. Proceedings of the 42nd International Conference on Machine Learning , pages=. 2025 , publisher= on Pith
  11. [11] Ferraz, Thomas Palmeira and Deffayet, Romain and Nikoulina, Vassilina and D. Retrieval-Augmented. arXiv preprint arXiv:2603.18272 , year= on PithPith's reviewarXiv
  12. [12] Xia, Tianle and Hu, Lingxiang and Sun, Yiding and Xu, Ming and Xu, Lan and Wang, Siying and Xu, Wei and Jiang, Jie , journal=. 2026 , url= on Pith
  13. [13] Graph of Skills: Dependency-Aware Structural Retrieval for Massive Agent Skills , author=. arXiv preprint arXiv:2604.05333 , year= on PithPith's reviewarXiv
  14. [14] Lewis, Patrick and Perez, Ethan and Piktus, Aleksandra and Petroni, Fabio and Karpukhin, Vladimir and Goyal, Naman and K. Retrieval-Augmented Generation for Knowledge-Intensive. Advances in Neural Information Processing Systems , volume= on Pith
  15. [15] Xiao, Shitao and Liu, Zheng and Zhang, Peitian and Muennighoff, Niklas and Lian, Defu and Nie, Jian-Yun , journal= on Pith
  16. [16] Do As I Can, Not As I Say: Grounding Language in Robotic Affordances , author=. Conference on Robot Learning , year= on Pith
  17. [17] Constrained Natural Language Action Planning for Resilient Embodied Systems , author=. arXiv preprint arXiv:2510.06357 , year= on PitharXiv
  18. [18] Banerjee, Debangshu and Suresh, Tarun and Ugare, Shubham and Misailovic, Sasa and Singh, Gagandeep , journal=. 2025 , url= on Pith
  19. [19] Gonzalez-Pumariega, Gonzalo and Leong, Su Yean and Sunkara, Neha and Choudhury, Sanjiban , booktitle=. Robotouille: An Asynchronous Planning Benchmark for. 2025 , url= on Pith
  20. [20] Cho, Hongcheol and Kang, Ryangkyung and Kim, Youngeun , journal= on Pith
show 6 more
  1. [21] Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and others , journal= on Pith
  2. [22] 2026 , howpublished= on Pith
  3. [23] Hu, Edward J. and Shen, Yelong and Wallis, Phillip and Allen-Zhu, Zeyuan and Li, Yuanzhi and Wang, Shean and Wang, Lu and Chen, Weizhu , booktitle= on Pith
  4. [24] Shridhar, Mohit and Yuan, Xingdi and C. International Conference on Learning Representations , year= on Pith
  5. [25] Chen, Lingjiao and Zaharia, Matei and Zou, James , journal=. 2024 , url= on Pith
  6. [26] Ong, Isaac and Almahairi, Amjad and Wu, Vincent and Chiang, Wei-Lin and Wu, Tianhao and Gonzalez, Joseph E. and Kadous, M. Waleed and Stoica, Ion , booktitle=. 2025 , url= on Pith

Challenges and discussion

Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.

No comments yet.

Sign in to comment