Pith.

Artificial Intelligence · arXiv:2609.09153 v1 · submitted 8 Sep 2026 · 54 references

Procedural Graphs: Self-Evolving Execution Structures for LLM Agents

Yuxing Lu, Yicheng Chen, Shanchan Wu, Sercan Ö. Ar{i}k

The paper says

“Large language models are increasingly deployed as agents that plan over long horizons and act through external tools.”

more of the abstract

Most agents select actions through unconstrained generation over an accumulating history, leaving implicit the procedural knowledge of what to do, in what order, and under which conditions. As trajectories lengthen, agents can lose track of their objectives, invoke tools out of order, and repeat unproductive actions. We introduce the Procedural Graph: just as a knowledge graph organizes factual knowledge into (entity, relation, entity) triplets for what-is questions, a Procedural Graph organizes procedural knowledge into (procedure, relation, procedure) triplets for what-to-do questions. At each decision step, the framework localizes the agent's active node, and a guidance model translates the surrounding subgraph into step-level situational guidance that biases the solver's next action without dictating it. The graph is self-evolving: an LLM refiner contrasts failed trajectories with successful ones and edits the graph's topology and attributes, committing edits that preserve or improve held-out validation performance while retaining rejected ones to discourage repetition. Starting from a minimal skeleton, the loop builds graphs that match or surpass hand-designed ones. It can also repair a flawed expert prior. Across multiple datasets, task types, and LLMs, the Procedural Graph delivers consistent gains over memory-based baselines, and self-evolution further improves performance without manual engineering.

from the abstract

Pith says

Machine review · deepseek-v4-flash · 9 Sep 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. Solid empirical paper with a clever combination of procedural graphs and self-evolution, but the validation gate for self-evolution is underpowered and the LLM-judge evaluation for two benchmarks introduces reproducibility concerns. This line is wrong

3 major objections · 6 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 54 references say.

The paper says, at Section 5.4, self-evolution analysis

The review did not quote this passage.

Pith says: Objection 1 of 3 · major

The validation gate for EnterpriseArena uses only 20 episodes per split. The paper candidly states that 'with 20 episodes per split, individual accept/reject decisions turn on one or two episodes and should be read as a search trace rather than as significance tests' (Section 5.4).

more of objection 1

This is a load-bearing concern because the central claims about self-evolution—building effective graphs from scratch and repairing flawed priors—rest on these commit decisions being meaningful and reproducible. The observed test results (85% survival vs. 0% baseline) are suggestive, but the search may have selected the best candidate, and the test set is also only 20 episodes. The authors should provide additional analysis to demonstrate that the accepted graphs generalize beyond the particular 20-episode validation split: e.g., bootstrap confidence intervals for validation scores, a check that accepted graphs also improve on a larger held-out set (if available), or a discussion of the risk of overfitting to the small validation set. Without this, the self-evolution evidence is weakened.

The paper says, at Table 3, ablation design

The review did not quote this passage.

Pith says: Objection 2 of 3 · major

The ablation compares three PG configurations (full-graph raw injection; full-graph generative; subgraph generative) plus a no-graph baseline. All PG configurations use the same underlying graph (Appendix B.5), which is a graph that was presumably optimized via evolution or hand-designed for use with localized guidance.

more of objection 2

A graph designed for a specific graph-to-text prompt may not be optimal when injected as raw JSON. This could bias the comparison in favor of the localized variant. Please clarify whether the graph was held constant across configurations, and discuss how the design of the graph (e.g., attribute verbosity) might affect the raw-injection and full-graph generative baselines.

The paper says, at Section 5.3, Table 2, MultiChallenge test set size

The review did not quote this passage.

Pith says: Objection 3 of 3 · major

The construction study for MultiChallenge uses only 56 test samples (the 'fast split' from Appendix D.3). The differences between modes (e.g., Mode 3 at 92.86% vs. Mode 5 at 91.07%) are within the range of sampling variability. The hand-crafted expert graph (Mode 1) drops performance from 87.50% to 58.93%, which is a large and surprising decline.

more of objection 3

The authors should provide confidence intervals or statistical tests for the construction-mode comparisons, and briefly discuss whether the expert graph is a reasonable prior or was intentionally flawed to demonstrate recovery.

6 minor objections

The paper says, at Algorithm 1

Pith says: minor objection 1 of 6

The notation 'Tail_{L}^{max}' is unclear. Please write 'Tail with max tokens L_max' or use a more standard notation for truncating the trajectory concatenation.

The paper says, at Section 5.4

Pith says: minor objection 2 of 6

The paper states that the validation gate results should be read as a search trace, yet later reports a Fisher's exact p-value (2.6e-8) comparing the returned graph's test survival to the baseline's. Because the search selected the best candidate, the test p-value is conditional on search and may be optimistic. Please discuss this conditioning issue.

show 4 more

The paper says, at Appendix B.4

Pith says: minor objection 3 of 6

The relation vocabulary includes CONVERGES_TO, but its semantics are not explained. Provide a brief example illustrating how it differs from LEADS_TO.

The paper says, at Table 1

Pith says: minor objection 4 of 6

Please specify how the 95% confidence intervals are computed (e.g., Clopper-Pearson for proportions, bootstrap for rubric scores).

The paper says, at Figure 2 caption

Pith says: minor objection 5 of 6

The word 'offline' appears as 'o??line' due to a formatting issue.

The paper says, at Appendix B.3

Pith says: minor objection 6 of 6

The baseline descriptions are helpful, but it would be useful to state whether each baseline's hyperparameters (e.g., number of retrieved trajectories for RAP, forgetting rate for MemoryBank) were tuned on the same validation splits used by PG.

What would settle it

Re-run the EnterpriseArena self-evolution (10 rounds, same protocol) with a validation set of 100 episodes instead of 20, keeping training and test sets unchanged. Compare the final test survival rate and the evolution trajectory (which rounds accept/reject) with the reported results. If final test metrics drop substantially or the set of accepted graphs differs systematically (e.g., different rounds accepted), the original validation gate was too noisy to support the claim. If results are similar, the concern is mitigated.

from the skeptic's reading

How the review went

4 readings by deepseek-v4-flash · 9 Sep 2026

  1. Reader holds up, with conditions confidence moderate
  2. Skeptic agrees 24 s · 34,123 tokens read, 2,373 written
  3. Editor wrote the verdict above 13 s · 34,956 tokens read, 1,186 written
  4. Referee major revision: 3 major objections, 6 minor 46 s · 35,805 tokens read, 1,652 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

The paper introduces Procedural Graphs, a directed attributed graph that encodes procedural knowledge for LLM agents, with localized generative guidance at inference and a validation-gated self-evolution loop. The combination is genuinely new relative to KnowAgent, AutoGuide, AWM, and ToolNet. The evaluation is broad (7 benchmarks, 4 LLMs) and the ablations are informative—localized guidance beats full-graph injection, and self-evolution can repair a flawed expert prior. The EnterpriseArena analysis is the most convincing part: it shows real changes in agent behavior (anticipatory fundraising, fewer redundant tool calls) and includes honest qualitative traces. The paper is clearly written and the authors flag their own limitations in Section 5.4.

The main soft spot is the validation gate for self-evolution. With N=20 episodes per split, accept/reject decisions turn on one or two episodes, and the paper admits this. The stress-test note is correct that this weakens the evolution claims, though I would soften it slightly: the headline results (Table 1) use fixed graphs and are not affected. The self-evolution is a separate contribution, and the authors present it as a search trace, not as a significance test. Still, claiming self-evolution "builds effective graphs from scratch" would be stronger with larger splits or explicit power analysis.

Second concern: the LLM judge (Gemini 3.1 Pro) is used for HotpotQA and MultiChallenge accuracy. This is a real reproducibility issue—different judges give different verdicts. The paper does not validate the judge against ground truth or report agreement. Overlapping confidence intervals in Table 1 also mean the sign test (p=4.3e-4) is comparing point estimates without accounting for uncertainty.

Third: the paper would benefit from releasing code and artifacts. No code is provided, which limits reproducibility of the self-evolution loop.

This paper is for researchers working on LLM agent scaffolding or procedural memory. It deserves a serious referee—the core idea is sound and the empirical scope is substantial. I would accept a peer-review invitation. The flaws are addressable: larger validation splits, independent judge validation, and code release. The paper is honest on its own terms and does not oversell.

Verdict: Conditionally accept with major on the self-evolution statistics and LLM judge validation.

The strongest objection

The self-evolution loop (Section 3.3) commits graph edits based on validation performance on a set of only 20 episodes for EnterpriseArena (Section 5.4). The paper itself states that 'with 20 episodes per split, individual accept/reject decisions turn on one or two episodes and should be read as a search trace rather than as significance tests.' This is a load-bearing concern because the claim that self-evolution builds effective graphs from scratch or repairs flawed priors (Section 5.3, Table 2; Section 5.4, Figure 4) relies on these commit decisions being meaningful and reproducible. With N=20, the validation score variance is high; a single lucky or unlucky episode can flip an accept/reject decision. Moreover, the same small validation set is reused across rounds, creating risk of overfitting to noise rather than genuine improvement. The paper presents the evolution results as direct evidence that the method constructs effective graphs, but if the gate is unreliable, the observed improvements could be artifacts of random variation or inadvertent selection on the validation set. The main comparison (Table 1) uses fixed graphs and is not affected by this concern, but the self-evolution contribution, a major part of the paper's novelty, is weakened.

The reader identified the validation gate's stability as the weakest assumption. I agree and elaborate why it is load-bearing: the self-evolution claim directly depends on commit decisions made with N=20. The paper's own caveat (Section 5.4) confirms the noise, yet the evolution results are presented as evidence of method effectiveness. The central claim about PG outperforming baselines (Table 1) does not rely on evolution, so the overall verdict remains CONDITIONAL. No adjustment is needed because the reader already flagged this and assigned the appropriate cautionary verdict.

The referee's report

If the results hold, PG provides a structured, editable procedural memory for LLM agents that combines localized guidance with automated refinement from execution feedback. This addresses a recognized gap in CoALA-style cognitive architectures (procedural memory is often implicit or scattered). The empirical coverage—seven benchmarks, four LLMs, seven baselines—is strong, and the ablation study (Table 3) convincingly shows that localized generative guidance outperforms full-graph injection. The self-evolution loop, while the most novel component, is also the most fragile; if its limitations are addressed, the framework could advance the state of the art in agent self-improvement.

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

The paper tries to establish that the procedural knowledge an LLM agent needs — what to do, in what order, and under which conditions — can be made explicit as a directed, attributed graph of (procedure, relation, procedure) triplets, and that this graph can steer the agent's next action without dictating it. It further claims the graph can be improved automatically: an LLM refiner proposes structural edits from execution feedback, and a validation gate commits only edits that preserve or improve held-out performance. If these claims hold, agents would not need hand-designed workflows or large episodic memories; a minimal skeleton could grow into effective procedural guidance from experience alone. The evidence is that the full pipeline ranks first or joint-first in 21 of 24 model–benchmark settings across four LLM families and seven benchmarks, and that evolution from a bare start-to-end skeleton matches or beats hand-crafted expert graphs while repairing a flawed expert prior (MultiChallenge recovers from 58.93% to 92.86%).

What carries the argument. The load-bearing object is the Procedural Graph, a directed attributed graph $G=(V,R,E,\Phi)$ whose edges are triplets $(u,r,v)$ stating that node $v$ is admissible after node $u$ under relation $r$, with each edge carrying three textual attributes: condition (when the transition applies), guidance (how to proceed), and pitfalls (what to avoid). Three mechanisms work on this object: localization with two-hop subgraph retrieval at inference time, so the guidance model reads transitions in their topological context; generative guidance, which verbalizes the subgraph into step-level advice appended to the solver's prompt, biasing without commanding; and validation-gated self-evolution with rejection memory, in which an LLM refiner proposes node and edge edits from contrastive trajectory analysis, a held-out validation score decides acceptance, and rejected candidates are logged to prevent repeated proposals. The validation gate is what keeps the loop trustworthy: structurally valid candidates are committed only when validation performance does not decrease.

Core claim. The central claim is that an LLM agent benefits from an external, inspectable, editable representation of admissible transitions — the Procedural Graph — instead of relying on free-form generation over a growing trajectory. At each decision step, the framework matches the agent's last action to the active graph node, extracts the surrounding two-hop subgraph, and has a guidance model turn the edge attributes (condition, guidance, pitfalls) into situational advice that biases the solver's next step. Offline, an LLM refiner contrasts failed with successful trajectories and proposes additions, deletions, and attribute revisions; a candidate is committed only if it matches or beats the current graph on a held-out validation set, and rejected edits are stored to discourage repeats. The authors show this pipeline outperforms the strongest memory-based baseline in 19 of 24 settings (two ties, three losses; one-sided sign test $p = 4.3 \times 10^{-4}$). The most direct demonstration is EnterpriseArena, a 132-month financial simulator with hidden crises: evolution from a start-to-end skeleton discovers the backbone of auditing cash, forecasting runway, and fundraising before liquidity runs out, lifting validation survival from 0% to 90% across ten rounds and test survival from 0% to 85%.

The premise it rests on. The load-bearing premise is that the validation signal is trustworthy enough to steer the search: with 20 episodes per split, an accept or reject can turn on one or two episodes, and on benchmark sets where an LLM judge scores answers, both the reported accuracy and the acceptance gate depend on that judge's verdicts; the paper itself cautions that individual decisions should be read as a search trace rather than as significance tests.

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

Procedural knowledge can be learned from execution experience without updating model weights: the graph lives outside the weights and is edited between training batches, so an agent's know-how becomes a persistent artifact rather than a side effect of prompting.

Expert priors that actively hurt performance can be repaired: on MultiChallenge, a hand-crafted graph dropped success from 87.50% to 58.93%, and iterative evolution (Mode 3) recovered to 92.86%.

Localization is the efficient design: generative guidance from the two-hop neighborhood beats both raw full-graph injection and full-graph generative guidance, raising ALFWorld success by about 9 points while cutting tokens by 70.9% relative to full-graph generation.

Long-horizon behavior in EnterpriseArena changes qualitatively, not just in counts: guided agents initiate fundraising during stable months, respecting the one-to-six-month delivery lag, while unguided agents request too late or violate the single-pending-request constraint.

Guidance costs tokens even when it saves steps, so further gains should come from reusing guidance across steps or generating it selectively, as the authors note in their conclusion.

A direct test of the gate's role: scale the 20-episode validation sets upward and measure how many round-by-round accept/reject decisions flip; the loop's robustness should scale gracefully if the gate is reading real signal rather than noise.

Because the graph is decoupled from the solver's weights, procedures learned with one LLM could plausibly transfer to a cheaper or different LLM, turning the graph into a portable execution skill; the paper raises cross-solver transfer as open future work rather than a demonstrated result.

The rejection memory is effectively a cache of negative procedural knowledge; weighting it by how close a rejected candidate came to acceptance, instead of merely concatenating it, could speed convergence of the evolution search.

Putting procedures outside the weights makes agent behavior partially inspectable and steerable through human-readable edits to conditions and pitfalls, a property that matters for auditing and correcting deployed agents; the paper notes the representation is inspectable but does not develop that angle.

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 · Review history (7 versions)

The paper

Abstract

Large language models are increasingly deployed as agents that plan over long horizons and act through external tools. Most agents select actions through unconstrained generation over an accumulating history, leaving implicit the procedural knowledge of what to do, in what order, and under which conditions. As trajectories lengthen, agents can lose track of their objectives, invoke tools out of order, and repeat unproductive actions. We introduce the Procedural Graph: just as a knowledge graph organizes factual knowledge into (entity, relation, entity) triplets for what-is questions, a Procedural Graph organizes procedural knowledge into (procedure, relation, procedure) triplets for what-to-do questions. At each decision step, the framework localizes the agent's active node, and a guidance model translates the surrounding subgraph into step-level situational guidance that biases the solver's next action without dictating it. The graph is self-evolving: an LLM refiner contrasts failed trajectories with successful ones and edits the graph's topology and attributes, committing edits that preserve or improve held-out validation performance while retaining rejected ones to discourage repetition. Starting from a minimal skeleton, the loop builds graphs that match or surpass hand-designed ones. It can also repair a flawed expert prior. Across multiple datasets, task types, and LLMs, the Procedural Graph delivers consistent gains over memory-based baselines, and self-evolution further improves performance without manual engineering.

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 · 2026-09-23

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

“Across multiple datasets, task types, and LLMs, the Procedural Graph delivers consistent gains over memory-based baselines, and self-evolution further improves performance without manual engineering.”

How the main claim is stated: asserted 1.00

What the abstract itself states, as probabilities: claims first 0.07 · states limitation 0.13 · human participants 0.05 · controlled comparison 0.92 · randomized assignment 0.02 · states quantitative result 0.05 · compares with prior results 0.97 · 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
Large language models are increasingly deployed as agents that plan over long horizons and act through external tools. background 1.00 0.04 0.02 asserted 0.99
Most agents select actions through unconstrained generation over an accumulating history, leaving implicit the procedural knowledge of what to do, in what order, and under which conditions. background 0.62 0.08 0.02 asserted 1.00
As trajectories lengthen, agents can lose track of their objectives, invoke tools out of order, and repeat unproductive actions. background 0.83 0.09 0.02 hedged possibility 0.79
We introduce the Procedural Graph: just as a knowledge graph organizes factual knowledge into (entity, relation, entity) triplets for what-is questions, a Procedural Graph organizes procedural knowledge into (procedure, relation, procedure) triplets for what-to-do questions. contribution 0.90 0.58 0.07 asserted 0.99
At each decision step, the framework localizes the agent's active node, and a guidance model translates the surrounding subgraph into step-level situational guidance that biases the solver's next action without dictating it. method 1.00 0.74 0.04 asserted 1.00
The graph is self-evolving: an LLM refiner contrasts failed trajectories with successful ones and edits the graph's topology and attributes, committing edits that preserve or improve held-out validation performance while retaining rejected ones to discourage repetition. method 0.98 0.91 0.08 asserted 1.00
Starting from a minimal skeleton, the loop builds graphs that match or surpass hand-designed ones. result 0.89 0.91 0.12 asserted 0.90
It can also repair a flawed expert prior. result 0.49 0.77 0.04 asserted 0.80
Across multiple datasets, task types, and LLMs, the Procedural Graph delivers consistent gains over memory-based baselines, and self-evolution further improves performance without manual engineering. result 1.00 0.95 0.85 asserted 1.00

What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.93

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, 2026-09-23: 1 request, 6,986 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

References (54)
  1. [1] M. Besta, N. Blach, A. Kubicek, R. Gerstenberger, M. Podstawski, L. Gianinazzi, J. Gajda, T. Lehmann, H. Niewiadomski, P. Nyczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. In Proceedings of the AAAI conference on artificial intelligence, ... on Pith
  2. [2] K. Deshpande, V. Sirdeshmukh, J. B. Mols, L. Jin, E.-Y. Hernandez-Cardona, D. Lee, J. Kritz, W. E. Primack, S. Yue, and C. Xing. MultiChallenge : A realistic multi-turn conversation evaluation benchmark challenging to frontier LLMs . In Findings of the Association for Computat... on Pith
  3. [3] Y. Du, F. Wei, and H. Zhang. AnyTool : Self-reflective, hierarchical agents for large-scale API calls. arXiv preprint arXiv:2402.04253, 2024 on PithPith's reviewarXiv
  4. [4] D. Edge, H. Trinh, N. Cheng, J. Bradley, A. Chao, A. Mody, S. Truitt, D. Metropolitansky, R. O. Ness, and J. Larson. From local to global: A graph RAG approach to query-focused summarization. arXiv preprint arXiv:2404.16130, 2024 on PithPith's reviewarXiv
  5. [5] R. Fang, Y. Liang, X. Wang, J. Wu, S. Qiao, P. Xie, F. Huang, H. Chen, and N. Zhang. Memp : Exploring agent procedural memory. arXiv preprint arXiv:2508.06433, 2025 on PithPith's reviewarXiv
  6. [6] Y. Fu, D.-K. Kim, J. Kim, S. Sohn, L. Logeswaran, K. Bae, and H. Lee. AutoGuide : Automated generation and selection of context-aware guidelines for large language model agents. Advances in Neural Information Processing Systems, 37: 0 119919--119948, 2024 on Pith
  7. [7] L. Gao, Y. Wang, M. Peng, J. Tang, Y. Shang, M. Sun, and J. Su. Tool graph retriever: Exploring dependency graph-based tool retrieval for large language models. arXiv preprint arXiv:2508.05152, 2025 on PithPith's reviewarXiv
  8. [8] Y. Han, Y. Wang, L. Qian, H. Li, Y. Cao, Y. He, X. Peng, N. Shen, Y. Xu, Y. Chen, et al. Can LLM agents be CFOs ? benchmarking long-horizon resource allocation in an uncertain enterprise environment. arXiv preprint arXiv:2603.23638, 2026 on PithPith's reviewarXiv
  9. [9] Y. Huang, J. Shi, Y. Li, C. Fan, S. Wu, Q. Zhang, Y. Liu, P. Zhou, Y. Wan, N. Gong, et al. MetaTool benchmark for large language models: Deciding whether to use tools and which to use. In International Conference on Learning Representations, volume 2024, pages 42978--43007, 2024 on Pith
  10. [10] Y. Jiang, H. Zhou, L. GU, A. Han, and T. Li. NaviAgent : Bilevel planning on tool navigation graph for large-scale orchestration. arXiv preprint arXiv:2506.19500, 2025 on PithPith's reviewarXiv
  11. [11] T. Kagaya, T. J. Yuan, Y. Lou, J. Karlekar, S. Pranata, A. Kinose, K. Oguri, F. Wick, and Y. You. RAP : Retrieval-augmented planning with contextual memory for multimodal LLM agents. In NeurIPS 2024 Workshop on Open-World Agents, 2024 on Pith
  12. [12] M. Li, Y. Zhao, B. Yu, F. Song, H. Li, H. Yu, Z. Li, F. Huang, and Y. Li. API-Bank : A comprehensive benchmark for tool-augmented LLMs . In Proceedings of the 2023 conference on empirical methods in natural language processing, pages 3102--3116, 2023 on Pith
  13. [13] X. Liu, Z. Peng, X. Yi, X. Xie, L. Xiang, Y. Liu, and D. Xu. ToolNet : Connecting large language models with massive tools via tool graph. arXiv preprint arXiv:2403.00839, 2024 a on PithPith's reviewarXiv
  14. [14] Z. Liu, Z. Lai, Z. Gao, E. Cui, Z. Li, X. Zhu, L. Lu, Q. Chen, Y. Qiao, J. Dai, et al. ControlLLM : Augment language models with tools by searching on graphs. In European Conference on Computer Vision, pages 89--105. Springer, 2024 b on Pith
  15. [15] E. Lumer, P. H. Basavaraju, M. Mason, J. A. Burke, and V. K. Subbiah. Graph RAG -tool fusion. arXiv preprint arXiv:2502.07223, 2025 on PithPith's reviewarXiv
  16. [16] Z. Nie, R. Shen, X. Yu, B. Yin, J. Zhang, and X. Hu. SkillGraph : Self-evolving multi-agent collaboration with multimodal graph topology. arXiv preprint arXiv:2604.17503, 2026 on PithPith's reviewarXiv
  17. [17] J. S. Park, J. O'Brien, C. J. Cai, M. R. Morris, P. Liang, and M. S. Bernstein. Generative agents: Interactive simulacra of human behavior. In Proceedings of the 36th annual acm symposium on user interface software and technology, pages 1--22, 2023 on Pith
  18. [18] S. G. Patil, T. Zhang, X. Wang, and J. E. Gonzalez. Gorilla : Large language model connected with massive APIs . Advances in Neural Information Processing Systems, 37: 0 126544--126565, 2024 on Pith
  19. [19] S. G. Patil, H. Mao, F. Yan, C. C.-J. Ji, V. Suresh, I. Stoica, and J. E. Gonzalez. The Berkeley function calling leaderboard ( BFCL ): From tool use to agentic evaluation of large language models. In Forty-second International Conference on Machine Learning, 2025 on Pith
  20. [20] T. Patwardhan, R. Dias, E. Proehl, G. Kim, M. Wang, O. Watkins, S. Posada Fishman, M. Aljubeh, P. Thacker, L. Fauconnet, N. S. Kim, et al. GDPval : Evaluating AI model performance on real-world economically valuable tasks. SuperIntelligence-Robotics-Safety & Alignment, 2 0 (4), 2025 on Pith
show 34 more
  1. [21] A. Prasad, A. Koller, M. Hartmann, P. Clark, A. Sabharwal, M. Bansal, and T. Khot. ADaPT : As-needed decomposition and planning with language models. In Findings of the Association for Computational Linguistics: NAACL 2024, pages 4226--4252, 2024 on Pith
  2. [22] Y. Qin, S. Liang, Y. Ye, K. Zhu, L. Yan, Y. Lu, Y. Lin, X. Cong, X. Tang, B. Qian, et al. ToolLLM : Facilitating large language models to master 16000+ real-world APIs . In International Conference on Learning Representations, volume 2024, pages 9695--9717, 2024 on Pith
  3. [23] C. Qu, S. Dai, X. Wei, H. Cai, S. Wang, D. Yin, J. Xu, and J.-R. Wen. Towards completeness-oriented tool retrieval for large language models. In Proceedings of the 33rd ACM International Conference on Information and Knowledge Management, pages 1930--1940, 2024 on Pith
  4. [24] P. Rasmussen, P. Paliychuk, T. Beauvais, J. Ryan, and D. Chalef. Zep : a temporal knowledge graph architecture for agent memory. arXiv preprint arXiv:2501.13956, 2025 on PithPith's reviewarXiv
  5. [25] T. Schick, J. Dwivedi-Yu, R. Dess \` , R. Raileanu, M. Lomeli, E. Hambro, L. Zettlemoyer, N. Cancedda, and T. Scialom. Toolformer : Language models can teach themselves to use tools. Advances in neural information processing systems, 36: 0 68539--68551, 2023 on Pith
  6. [26] Y. Shen, K. Song, X. Tan, D. Li, W. Lu, and Y. Zhuang. HuggingGPT : Solving AI tasks with ChatGPT and its friends in Hugging Face . Advances in Neural Information Processing Systems, 36: 0 38154--38180, 2023 on Pith
  7. [27] Y. Shen, K. Song, X. Tan, W. Zhang, K. Ren, S. Yuan, W. Lu, D. Li, and Y. Zhuang. TaskBench : Benchmarking large language models for task automation. Advances in Neural Information Processing Systems, 37: 0 4540--4574, 2024 on Pith
  8. [28] T. Shi, S. Chen, B. Jiang, L. Song, L. Yang, and J. Zhao. Experiential reinforcement learning. arXiv preprint arXiv:2602.13949, 2026 on PitharXiv
  9. [29] N. Shinn, F. Cassano, A. Gopinath, K. Narasimhan, and S. Yao. Reflexion : Language agents with verbal reinforcement learning. Advances in neural information processing systems, 36: 0 8634--8652, 2023 on Pith
  10. [30] M. Shridhar, X. Yuan, M.-A. Cote, Y. Bisk, A. Trischler, and M. Hausknecht. ALFWorld : Aligning text and embodied environments for interactive learning. In International Conference on Learning Representations, 2021 on Pith
  11. [31] T. Sumers, S. Yao, K. R. Narasimhan, and T. L. Griffiths. Cognitive architectures for language agents. Transactions on Machine Learning Research, 2023 on Pith
  12. [32] Z. Sun, Z. Liu, Y. Zang, Y. Cao, X. Dong, T. Wu, D. Lin, and J. Wang. SE-Agent : Self-evolving computer use agent with autonomous learning from experience. arXiv preprint arXiv:2508.04700, 2025 on PithPith's reviewarXiv
  13. [33] G. Wang, Y. Xie, Y. Jiang, A. Mandlekar, C. Xiao, Y. Zhu, L. Fan, and A. Anandkumar. Voyager : An open-ended embodied agent with large language models. arXiv preprint arXiv:2305.16291, 2023 a on PithPith's reviewarXiv
  14. [34] L. Wang, W. Xu, Y. Lan, Z. Hu, Y. Lan, R. K.-W. Lee, and E.-P. Lim. Plan-and-solve prompting: Improving zero-shot chain-of-thought reasoning by large language models. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: long pap... on Pith
  15. [35] R. Wang, X. Han, L. Ji, S. Wang, T. Baldwin, and H. Li. ToolGen : Unified tool retrieval and calling via generation. In International Conference on Learning Representations, volume 2025, pages 73473--73498, 2025 a on Pith
  16. [36] X. Wang, Y. Chen, L. Yuan, Y. Zhang, Y. Li, H. Peng, and H. Ji. Executable code actions elicit better LLM agents. arXiv preprint arXiv:2402.01030, 2024 a on PithPith's reviewarXiv
  17. [37] Z. Wang, D. Fried, and G. Neubig. TroVE : Inducing verifiable and efficient toolboxes for solving programmatic tasks. arXiv preprint arXiv:2401.12869, 2024 b on PithPith's reviewarXiv
  18. [38] Z. Wang, Q. Wu, X. Zhang, C. Zhang, W. Yao, F. E. Faisal, B. Peng, S. Qin, S. Nath, Q. Lin, et al. WebXSkill : Skill learning for autonomous web agents. arXiv preprint arXiv:2604.13318, 2026 on PithPith's reviewarXiv
  19. [39] Z. Z. Wang, J. Mao, D. Fried, and G. Neubig. Agent workflow memory. In International Conference on Machine Learning, pages 63897--63911. PMLR, 2025 b on Pith
  20. [40] B. T. Willard and R. Louf. Efficient guided generation for large language models. arXiv preprint arXiv:2307.09702, 2023 on PithPith's reviewarXiv
  21. [41] R. Wu, X. Wang, J. Mei, P. Cai, D. Fu, C. Yang, L. Wen, X. Yang, Y. Shen, Y. Wang, et al. EvolveR : Self-evolving LLM agents through an experience-driven lifecycle. arXiv preprint arXiv:2510.16079, 2025 on PithPith's reviewarXiv
  22. [42] X. Wu, Y. Shen, C. Shan, K. Song, S. Wang, B. Zhang, J. Feng, H. Cheng, W. Chen, Y. Xiong, et al. Can graph learning improve planning in LLM -based agents? Advances in Neural Information Processing Systems, 37: 0 5338--5383, 2024 on Pith
  23. [43] R. Xiao, W. Ma, K. Wang, Y. Wu, J. Zhao, H. Wang, F. Huang, and Y. Li. FlowBench : Revisiting and benchmarking workflow-guided planning for LLM -based agents. In Findings of the Association for Computational Linguistics: EMNLP 2024, pages 10883--10900, 2024 on Pith
  24. [44] W. Xu, Z. Liang, K. Mei, H. Gao, J. Tan, and Y. Zhang. A-Mem : Agentic memory for LLM agents. Advances in Neural Information Processing Systems, 38: 0 17577--17604, 2026 on Pith
  25. [45] Z. Yang, P. Qi, S. Zhang, Y. Bengio, W. Cohen, R. Salakhutdinov, and C. D. Manning. HotpotQA : A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 conference on empirical methods in natural language processing, pages 2369--2380, 2018 on Pith
  26. [46] S. Yao, D. Yu, J. Zhao, I. Shafran, T. Griffiths, Y. Cao, and K. Narasimhan. Tree of thoughts: Deliberate problem solving with large language models. Advances in neural information processing systems, 36: 0 11809--11822, 2023 a on Pith
  27. [47] S. Yao, J. Zhao, D. Yu, N. Du, I. Shafran, K. Narasimhan, and Y. Cao. ReAct : Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023 b on Pith
  28. [48] S. Yao, N. Shinn, P. Razavi, and K. Narasimhan. tau-bench: A benchmark for tool-agent-user interaction in real-world domains. arXiv preprint arXiv:2406.12045, 2024 on PithPith's reviewarXiv
  29. [49] J. Zhang, J. Xiang, Z. Yu, F. Teng, X. Chen, J. Chen, M. Zhuge, X. Cheng, S. Hong, J. Wang, et al. AFlow : Automating agentic workflow generation. In International Conference on Learning Representations, volume 2025, pages 34040--34077, 2025 on Pith
  30. [50] K. Zhang, H. Chen, L. Li, and W. Wang. Don't fine-tune, decode: Syntax error-free tool use via constrained decoding. arXiv preprint arXiv:2310.07075, 2023 on PithPith's reviewarXiv
  31. [51] A. Zhao, D. Huang, Q. Xu, M. Lin, Y.-J. Liu, and G. Huang. ExpeL : LLM agents are experiential learners. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 38, pages 19632--19642, 2024 on Pith
  32. [52] B. Zheng, M. Y. Fatemi, X. Jin, Z. Z. Wang, A. Gandhi, Y. Song, Y. Gu, J. Srinivasa, G. Liu, G. Neubig, et al. SkillWeaver : Web agents can self-improve by discovering and honing skills. arXiv preprint arXiv:2504.07079, 2025 on PithPith's reviewarXiv
  33. [53] W. Zhong, L. Guo, Q. Gao, H. Ye, and Y. Wang. MemoryBank : Enhancing large language models with long-term memory. In Proceedings of the AAAI conference on artificial intelligence, volume 38, pages 19724--19731, 2024. Issue 17 on Pith
  34. [54] Y. Zhu, S. Qiao, Y. Ou, S. Deng, S. Lyu, Y. Shen, L. Liang, J. Gu, H. Chen, and N. Zhang. KnowAgent : Knowledge-augmented planning for LLM -based agents. In Findings of the Association for Computational Linguistics: NAACL 2025, pages 3709--3732, 2025 on Pith
Cited by 4 Pith papers

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