arxiv: 2605.10057 · v2 · submitted 2026-05-11 · 💻 cs.AI · cs.MA

Recognition: no theorem link

STAR: Failure-Aware Markovian Routing for Multi-Agent Spatiotemporal Reasoning

Flora D. Salim, Hao Xue, Lihuan Li, Ruiyi Yang

Authors on Pith no claims yet

Pith reviewed 2026-05-13 03:42 UTC · model grok-4.3

classification 💻 cs.AI cs.MA

keywords failure-aware routingmulti-agent systemsspatiotemporal reasoningMarkovian routingexecution tracesrecovery transitionsLLM agentstool-augmented systems

0 comments

The pith

Failure-aware Markovian routing learns recovery transitions from unsuccessful execution traces for multi-agent spatiotemporal reasoning.

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

The paper presents STAR as a framework that turns routing decisions among specialist agents into an explicit state-conditioned Markovian policy. It combines expert-defined nominal routes with transitions learned from both successful and failed execution traces, so the router can react differently to distinct error types instead of treating all failures the same. Retaining unsuccessful traces expands the policy to cover error states that success-only training leaves unsupported. This produces measurable gains on spatiotemporal benchmarks, especially when queries force the system off the standard path, and the gains hold across eight different LLM backbones. A sympathetic reader would see this as a concrete way to make recovery in composed agent systems less ad hoc and more optimizable.

Core claim

STAR externalizes inter-agent control as a state-conditioned transition policy over the current agent, task type, and typed execution status. At its center is an agent routing matrix that merges expert-specified nominal routes with recovery transitions learned from execution traces. Conditioning on distinct failure states lets the router distinguish malformed outputs, missing dependencies, and tool-query mismatches rather than collapsing them into a generic retry. Retaining unsuccessful traces during training enlarges the support of the routing policy on error states, enabling recovery transitions that success-only training cannot represent. This yields improvements over baselines on three时空

What carries the argument

The agent routing matrix that merges expert nominal routes with failure-conditioned recovery transitions learned from execution traces, allowing differentiated responses to specific error types.

If this is right

Retaining unsuccessful traces enlarges the support of the routing policy on error states.
This enables recovery transitions that success-only training cannot represent.
Improvements appear across three spatiotemporal benchmarks and eight backbone LLMs.
The largest gains occur on queries whose execution deviates from the nominal routing path.
Typed failure-aware routing contributes more than specialist composition alone.

Where Pith is reading between the lines

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

The same matrix structure could be applied to other compositional tasks that involve heterogeneous agents or tools.
Explicit failure typing may reduce reliance on hand-crafted retry logic in broader LLM agent systems.
If failure categories prove stable, the approach could support incremental updates to the routing matrix as new error patterns appear.

Load-bearing premise

Distinct failure types can be reliably identified and classified from execution traces, and the learned recovery transitions will generalize beyond the training traces to new queries.

What would settle it

A direct comparison on the same benchmarks where a router trained only on successful traces matches or exceeds the performance of the failure-aware version on queries that require recovery from execution errors.

Figures

Figures reproduced from arXiv: 2605.10057 by Flora D. Salim, Hao Xue, Lihuan Li, Ruiyi Yang.

**Figure 1.** Figure 1: STAR architecture. Queries are parsed into a task profile, the failure-aware routing matrix selects specialists conditioned on the current agent, task type, execution status, and specialists execute through an extract-compute-deposit protocol over a shared blackboard before final fusion. This paper proposes STAR (Spatio-Temporal Agent Router), a failure-aware routing framework that externalizes inter-agent… view at source ↗

**Figure 2.** Figure 2: State-conditioned routing matrix slices for a representative task type. Each panel visualizes [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Failure-aware routing and execution feedback in [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

**Figure 4.** Figure 4: Routing example through the dual-system kernel. System 1 nominal routes (green arrows) [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗

**Figure 5.** Figure 5: Empirical precision–coverage trade-off of [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗

**Figure 6.** Figure 6: Learned transition matrix M averaged across task types, rendered as four status-conditional heatmaps (SUCCESS / FAIL / INFO_MISSING / BLOCKED). Rows index the from_agent, columns index the to_agent, and each cell is P(next | from,status). Each benchmark produces a qualitatively different SUCCESS map (because task taxonomies differ), but every benchmark exhibits the same structural property—error-state rows… view at source ↗

**Figure 7.** Figure 7: Top-3 successors per failure status, grouped by originating agent. For every [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗

read the original abstract

Compositional spatiotemporal reasoning often requires a system to invoke multiple heterogeneous specialists, such as geometric, temporal, topological, and trajectory agents. A central question is how such a system should route among specialists when execution does not simply succeed or fail, but fails in qualitatively different ways. Existing tool-augmented and multi-agent LLM systems typically leave this routing decision implicit in language generation, making recovery ad hoc, difficult to interpret, and hard to optimize. This paper presents STAR (Spatio-Temporal Agent Router), a failure-aware routing framework that externalizes inter-agent control as a state-conditioned transition policy over the current agent, task type, and typed execution status. At the center of STARis an agent routing matrix that combines expert-specified nominal routes with recovery transitions learned from execution traces. Because the matrix conditions on distinct failure states, the router can respond differently to malformed outputs, missing dependencies, and tool--query mismatches, rather than collapsing them into a generic retry signal. Specialists execute through a tool-grounded extract--compute--deposit protocol and write intermediate results to a shared blackboard for downstream fusion. Results prove that retaining unsuccessful traces during training enlarges the support of the routing policy on error states, enabling recovery transitions that success-only training cannot represent. Across three spatiotemporal benchmarks and eight backbone LLMs, STAR improves over multiple baselines with the clearest gains on queries whose execution deviates from the nominal routing path. Router-specific ablations and recovery analyses further show that typed failure-aware routing, rather than specialist composition alone, is a key factor for these improvements.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

STAR adds an explicit Markov routing matrix conditioned on typed failures and trained on unsuccessful traces, which helps multi-agent LLM routing on deviated paths in spatiotemporal tasks but leaves the experimental grounding thin.

read the letter

STAR turns routing between specialist agents into an explicit state-conditioned transition matrix that treats different failure modes separately and learns recovery transitions from traces that did not succeed. The main practical move is keeping unsuccessful executions in the training data so the policy can represent recoveries that success-only training misses. They combine this with hand-specified nominal routes and a shared blackboard for intermediate results, then test across three benchmarks and eight backbone LLMs, with the largest gains on queries that stray from the expected path. That framing addresses a real pain point in multi-agent LLM systems where recovery is usually left implicit and hard to optimize. The typed failures (malformed outputs, missing dependencies, tool mismatches) and the tool-grounded execution protocol are straightforward engineering choices that make the router more interpretable than pure language-based routing. The ablations on the router component and recovery analysis try to show that the failure awareness, not just the agent composition, drives the improvement. The soft spots sit in the evidence. The abstract states results and a proof about trace retention enlarging policy support, yet gives no specifics on how failure types are reliably partitioned, what the exact baselines are, or whether statistical tests back the gains. Without those, it is difficult to rule out that the reported improvements largely reflect the training distribution rather than new generalization. The stress-test point about novel error states is still open: if the learned transitions mainly interpolate within seen failure patterns, the claim of enlarged support would need stronger hold-out tests on out-of-distribution queries. This work is aimed at researchers building multi-agent systems for compositional reasoning tasks that mix geometry, time, and trajectories. Readers who need concrete mechanisms for error recovery in LLM agents will find the framework worth examining, even if they want tighter validation of the failure typing and generalization. It deserves a serious referee to examine the methods and experiments in detail.

Referee Report

3 major / 2 minor

Summary. The paper introduces STAR, a failure-aware Markovian routing framework for multi-agent spatiotemporal reasoning. It externalizes inter-agent control as a state-conditioned transition policy over the current agent, task type, and typed execution status (e.g., malformed outputs, missing dependencies, tool-query mismatches). The routing matrix combines expert-specified nominal routes with recovery transitions learned from execution traces that include unsuccessful cases. The central claim is that retaining unsuccessful traces enlarges the support of the routing policy on error states, enabling recovery transitions unavailable under success-only training. Empirical results across three spatiotemporal benchmarks and eight backbone LLMs show improvements over baselines, with largest gains on queries whose execution deviates from the nominal path; ablations attribute gains to typed failure-aware routing.

Significance. If the central claims hold after addressing methodological gaps, the work offers a concrete mechanism for making routing decisions explicit and optimizable in multi-agent LLM systems, moving beyond implicit language-based recovery. Learning recovery transitions from typed failure traces is a promising direction for robustness in compositional tool-use settings. The multi-benchmark, multi-LLM evaluation provides a reasonable testbed, and the distinction between nominal and recovery transitions could influence future agent architectures if the generalization properties are demonstrated.

major comments (3)

[Abstract] Abstract: The statement that 'Results prove that retaining unsuccessful traces during training enlarges the support of the routing policy on error states' is presented as a core result, yet the manuscript provides no mathematical definition of the routing matrix, no derivation of policy support, and no proof or formal argument showing enlargement. This is load-bearing for the central theoretical claim.
[Experimental Evaluation] Experimental sections: No description is given of the failure-typing procedure (how malformed outputs, missing dependencies, and tool-query mismatches are reliably partitioned into distinct, non-overlapping categories), baseline definitions, statistical testing, or hold-out protocols that introduce novel failure categories or out-of-distribution queries. Without these, it is impossible to determine whether reported gains reflect genuine enlargement of support or interpolation within the training failure distribution.
[Ablation Studies] Ablation and recovery analyses: The claim that 'typed failure-aware routing, rather than specialist composition alone, is a key factor' rests on router-specific ablations, but the manuscript does not report effect sizes, controls for training-data distribution, or comparisons that isolate the contribution of typed failure conditioning from other factors such as trace volume or blackboard usage.

minor comments (2)

[Abstract] Abstract contains the typo 'STARis' (should be 'STAR is').
[Abstract] Abstract uses inconsistent hyphenation ('tool--query' with double dash); standardize to 'tool-query'.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed review. The comments highlight important areas for clarification in the presentation of the routing matrix, experimental protocols, and ablation controls. We address each point below and will revise the manuscript accordingly to strengthen the methodological transparency while preserving the core empirical contributions.

read point-by-point responses

Referee: [Abstract] Abstract: The statement that 'Results prove that retaining unsuccessful traces during training enlarges the support of the routing policy on error states' is presented as a core result, yet the manuscript provides no mathematical definition of the routing matrix, no derivation of policy support, and no proof or formal argument showing enlargement. This is load-bearing for the central theoretical claim.

Authors: We agree that the abstract's use of 'prove' is imprecise for an empirical result and that a formal definition of the routing matrix and policy support should be provided in the main text. The routing matrix is defined in Section 3.2 as the state-conditioned transition matrix M(s, a, f) combining expert-specified nominal routes with learned recovery transitions, where s encodes (current_agent, task_type), a is the next agent, and f is the typed failure status. Policy support refers to the set of reachable (state, action) pairs under the induced Markov chain. Enlargement is demonstrated empirically by the additional recovery transitions present when unsuccessful traces are retained (Figure 4). We did not include a formal derivation or proof, as the work focuses on the algorithmic framework and empirical validation rather than theoretical bounds on support size. We will revise the abstract to 'empirically demonstrate,' add an explicit mathematical definition of the matrix and support in Section 3, and include a brief structural argument based on the expanded state space. revision: yes
Referee: [Experimental Evaluation] Experimental sections: No description is given of the failure-typing procedure (how malformed outputs, missing dependencies, and tool-query mismatches are reliably partitioned into distinct, non-overlapping categories), baseline definitions, statistical testing, or hold-out protocols that introduce novel failure categories or out-of-distribution queries. Without these, it is impossible to determine whether reported gains reflect genuine enlargement of support or interpolation within the training failure distribution.

Authors: We will add a dedicated subsection in the experimental setup detailing the failure-typing procedure: failures are classified automatically from execution logs using rule-based detectors with priority ordering to ensure mutual exclusivity (malformed outputs via parse errors, missing dependencies via graph resolution failures, tool-query mismatches via embedding similarity threshold). Baselines are the standard LLM router, success-only router, and untyped failure router, all using the same backbone LLMs. Statistical testing consists of mean and standard deviation over three random seeds with paired t-tests for significance (p-values reported in supplementary material). The hold-out protocol is an 80/20 query split within each benchmark; failure types are drawn from the same distribution as training traces, with no novel OOD failure categories introduced. We will clarify that the gains reflect improved recovery within the observed failure distribution rather than extrapolation to unseen failure types, and add pseudocode for the typing rules. revision: yes
Referee: [Ablation Studies] Ablation and recovery analyses: The claim that 'typed failure-aware routing, rather than specialist composition alone, is a key factor' rests on router-specific ablations, but the manuscript does not report effect sizes, controls for training-data distribution, or comparisons that isolate the contribution of typed failure conditioning from other factors such as trace volume or blackboard usage.

Authors: The router ablations (Table 3) compare the full typed failure-aware router against untyped failure (all failures collapsed to one category) and success-only variants, with trace volume and blackboard usage held constant across conditions. We report accuracy as mean ± standard deviation; raw deltas between conditions can be derived from the table. To further isolate typed conditioning, all variants are trained on identical trace sets, differing only in the failure-state representation. We will expand the ablation section to explicitly report effect sizes (e.g., accuracy deltas and Cohen's d), add a supplementary control varying trace volume while fixing typing, and clarify that blackboard usage is shared. A complete factorial isolation of every factor would require additional runs; we will perform and report the most critical controls within space limits. revision: partial

Circularity Check

0 steps flagged

No significant circularity; empirical validation is self-contained

full rationale

The paper's central claim—that retaining unsuccessful traces enlarges policy support on error states—is presented as an empirical result from training and benchmarking on spatiotemporal tasks, not as a mathematical derivation or first-principles prediction. No equations, self-citations, or ansatzes are invoked in the abstract or described structure that would reduce the reported gains to a definitional equivalence or fitted input by construction. The framework is validated across multiple benchmarks and LLMs with ablations, making the derivation chain independent of its own inputs.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The framework rests on the ability to distinguish failure types and to learn useful recovery transitions from traces; no explicit free parameters or invented entities are named in the abstract.

free parameters (1)

recovery transition parameters
Learned from execution traces that include failures, as stated in the abstract.

axioms (1)

domain assumption Typed failure states provide distinguishable and actionable signals for routing decisions
Invoked when the router is said to respond differently to malformed outputs, missing dependencies, and tool-query mismatches.

pith-pipeline@v0.9.0 · 5587 in / 1307 out tokens · 101007 ms · 2026-05-13T03:42:04.871171+00:00 · methodology

Review history (2 revisions) →

discussion (0)

Reference graph

Works this paper leans on

32 extracted references · 32 canonical work pages · 5 internal anchors

[1]

Can large language models be good path planners? a benchmark and investigation on spatial-temporal reasoning.arXiv preprint arXiv:2310.03249, 2023

Mohamed Aghzal, Erion Plaku, and Ziyu Yao. Can large language models be good path planners? a benchmark and investigation on spatial-temporal reasoning.arXiv preprint arXiv:2310.03249, 2023

work page arXiv 2023
[2]

Graph of thoughts: Solving elaborate problems with large language models

Maciej Besta, Nils Blach, Ales Kubicek, Robert Gerstenberger, Michal Podstawski, Lukas Gianinazzi, Joanna Gajda, Tomasz Lehmann, Hubert Niewiadomski, Piotr Nyczyk, et al. Graph of thoughts: Solving elaborate problems with large language models. InProceedings of the AAAI conference on artificial intelligence, volume 38, pages 17682–17690, 2024

work page 2024
[3]

V-star: Bench- marking video-llms on video spatio-temporal reasoning.arXiv preprint arXiv:2503.11495, 2025

Zixu Cheng, Jian Hu, Ziquan Liu, Chenyang Si, Wei Li, and Shaogang Gong. V-star: Bench- marking video-llms on video spatio-temporal reasoning.arXiv preprint arXiv:2503.11495, 2025

work page arXiv 2025
[4]

A., Tihanyi, N., and Debbah, M

Mohamed Amine Ferrag, Norbert Tihanyi, and Merouane Debbah. From llm reasoning to autonomous ai agents: A comprehensive review.arXiv preprint arXiv:2504.19678, 2025

work page arXiv 2025
[5]

Tremu: Towards neuro-symbolic temporal reasoning for llm-agents with memory in multi-session dialogues

Yubin Ge, Salvatore Romeo, Jason Cai, Raphael Shu, Yassine Benajiba, Monica Sunkara, and Yi Zhang. Tremu: Towards neuro-symbolic temporal reasoning for llm-agents with memory in multi-session dialogues. InFindings of the Association for Computational Linguistics: ACL 2025, pages 18974–18988, 2025

work page 2025
[6]

Llm-based intent processing and network optimization using attention-based hierarchical reinforcement learning

Md Arafat Habib, Pedro Enrique Iturria Rivera, Yigit Ozcan, Medhat Elsayed, Majid Bavand, Raimundus Gaigalas, and Melike Erol-Kantarci. Llm-based intent processing and network optimization using attention-based hierarchical reinforcement learning. In2025 IEEE Wireless Communications and Networking Conference (WCNC), pages 1–6. IEEE, 2025

work page 2025
[7]

Exploring advanced llm multi-agent systems based on blackboard architecture.arXiv preprint arXiv:2507.01701, 2025

Bochen Han and Songmao Zhang. Exploring advanced llm multi-agent systems based on blackboard architecture.arXiv preprint arXiv:2507.01701, 2025

work page arXiv 2025
[8]

Metagpt: Meta programming for a multi-agent collaborative framework

Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Jinlin Wang, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, et al. Metagpt: Meta programming for a multi-agent collaborative framework. InThe twelfth international conference on learning representations, 2023

work page 2023
[9]

Stbench: Assessing the ability of large language models in spatio-temporal analysis

Wenbin Li, Di Yao, Ruibo Zhao, Wenjie Chen, Zijie Xu, Chengxue Luo, Chang Gong, Quanliang Jing, Haining Tan, and Jingping Bi. Stbench: Assessing the ability of large language models in spatio-temporal analysis. InCompanion Proceedings of the ACM on Web Conference 2025, pages 749–752, 2025

work page 2025
[10]

A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges.Vicinagearth, 1(1):9, 2024

Xinyi Li, Sai Wang, Siqi Zeng, Yu Wu, and Yi Yang. A survey on llm-based multi-agent systems: workflow, infrastructure, and challenges.Vicinagearth, 1(1):9, 2024

work page 2024
[11]

Zechen Li, Baiyu Chen, Hao Xue, and Flora D. Salim. Zara: Training-free motion time-series reasoning via evidence-grounded llm agents.arXiv preprint arXiv:2508.04038, 2026

work page internal anchor Pith review Pith/arXiv arXiv 2026
[12]

Chameleon: Plug-and-play compositional reasoning with large language models.Advances in Neural Information Processing Systems, 36:43447–43478, 2023

Pan Lu, Baolin Peng, Hao Cheng, Michel Galley, Kai-Wei Chang, Ying Nian Wu, Song-Chun Zhu, and Jianfeng Gao. Chameleon: Plug-and-play compositional reasoning with large language models.Advances in Neural Information Processing Systems, 36:43447–43478, 2023

work page 2023
[13]

Self-refine: Iterative refinement with self-feedback.Advances in neural information processing systems, 36:46534–46594, 2023

Aman Madaan, Niket Tandon, Prakhar Gupta, Skyler Hallinan, Luyu Gao, Sarah Wiegreffe, Uri Alon, Nouha Dziri, Shrimai Prabhumoye, Yiming Yang, et al. Self-refine: Iterative refinement with self-feedback.Advances in neural information processing systems, 36:46534–46594, 2023. 10

work page 2023
[14]

Omnirouter: Budget and performance controllable multi-llm routing.ACM SIGKDD Explorations Newsletter, 27(2):107–116, 2025

Kai Mei, Wujiang Xu, Minghao Guo, Shuhang Lin, and Yongfeng Zhang. Omnirouter: Budget and performance controllable multi-llm routing.ACM SIGKDD Explorations Newsletter, 27(2):107–116, 2025

work page 2025
[15]

STReasoner: Empowering LLMs for Spatio-Temporal Reasoning in Time Series via Spatial-Aware Reinforcement Learning

Juntong Ni, Shiyu Wang, Ming Jin, Qi He, and Wei Jin. Streasoner: Empowering llms for spatio-temporal reasoning in time series via spatial-aware reinforcement learning.arXiv preprint arXiv:2601.03248, 2026

work page internal anchor Pith review Pith/arXiv arXiv 2026
[16]

TaskWeaver: A code-first agent framework,

Bo Qiao, Liqun Li, Xu Zhang, Shilin He, Yu Kang, Chaoyun Zhang, Fangkai Yang, Hang Dong, Jue Zhang, Lu Wang, et al. Taskweaver: A code-first agent framework.arXiv preprint arXiv:2311.17541, 2023

work page arXiv 2023
[17]

Benchmarking spatiotemporal reasoning in llms and reasoning models: Capabilities and challenges.arXiv preprint arXiv:2505.11618, 2025

Pengrui Quan, Brian Wang, Kang Yang, Liying Han, and Mani Srivastava. Benchmarking spatiotemporal reasoning in llms and reasoning models: Capabilities and challenges.arXiv preprint arXiv:2505.11618, 2025

work page arXiv 2025
[18]

CoRR , volume =

Matthew Renze and Erhan Guven. Self-reflection in llm agents: Effects on problem-solving performance.arXiv preprint arXiv:2405.06682, 2024

work page arXiv 2024
[19]

Llm-based multi-agent blackboard system for information discovery in data science.arXiv preprint arXiv:2510.01285, 2025

Alireza Salemi, Mihir Parmar, Palash Goyal, Yiwen Song, Jinsung Yoon, Hamed Zamani, Tomas Pfister, and Hamid Palangi. Llm-based multi-agent blackboard system for information discovery in data science.arXiv preprint arXiv:2510.01285, 2025

work page arXiv 2025
[20]

Toolformer: Language models can teach themselves to use tools.Advances in neural information processing systems, 36:68539– 68551, 2023

Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. Toolformer: Language models can teach themselves to use tools.Advances in neural information processing systems, 36:68539– 68551, 2023

work page 2023
[21]

Outrageously Large Neural Networks: The Sparsely-Gated Mixture-of-Experts Layer

Noam Shazeer, Azalia Mirhoseini, Krzysztof Maziarz, Andy Davis, Quoc Le, Geoffrey Hinton, and Jeff Dean. Outrageously large neural networks: The sparsely-gated mixture-of-experts layer.arXiv preprint arXiv:1701.06538, 2017

work page internal anchor Pith review Pith/arXiv arXiv 2017
[22]

Exploring multi-modal data with tool-augmented llm agents for precise causal discovery

ChengAo Shen, Zhengzhang Chen, Dongsheng Luo, Dongkuan Xu, Haifeng Chen, and Jingchao Ni. Exploring multi-modal data with tool-augmented llm agents for precise causal discovery. In Findings of the Association for Computational Linguistics: ACL 2025, pages 636–660, 2025

work page 2025
[23]

Reflexion: Language agents with verbal reinforcement learning.Advances in neural information processing systems, 36:8634–8652, 2023

Noah Shinn, Federico Cassano, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. Reflexion: Language agents with verbal reinforcement learning.Advances in neural information processing systems, 36:8634–8652, 2023

work page 2023
[24]

Learning options in reinforcement learning

Martin Stolle and Doina Precup. Learning options in reinforcement learning. InInternational Symposium on abstraction, reformulation, and approximation, pages 212–223. Springer, 2002

work page 2002
[25]

Multi-Agent Collaboration Mechanisms: A Survey of LLMs

Khanh-Tung Tran, Dung Dao, Minh-Duong Nguyen, Quoc-Viet Pham, Barry O’Sullivan, and Hoang D Nguyen. Multi-agent collaboration mechanisms: A survey of llms.arXiv preprint arXiv:2501.06322, 2025

work page internal anchor Pith review arXiv 2025
[26]

Beyond react: A planner-centric framework for complex tool- augmented llm reasoning

Xiaolong Wei, Yuehu Dong, Xingliang Wang, Xingyu Zhang, Zhejun Zhao, Dongdong Shen, Long Xia, and Dawei Yin. Beyond react: A planner-centric framework for complex tool- augmented llm reasoning. InProceedings of the AAAI Conference on Artificial Intelligence, volume 40, pages 33845–33853, 2026

work page 2026
[27]

Autogen: Enabling next-gen llm applications via multi-agent conversations

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, et al. Autogen: Enabling next-gen llm applications via multi-agent conversations. InFirst conference on language modeling, 2024

work page 2024
[28]

Large language models can learn temporal reasoning

Siheng Xiong, Ali Payani, Ramana Kompella, and Faramarz Fekri. Large language models can learn temporal reasoning. InProceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 10452–10470, 2024

work page 2024
[29]

Ruiyi Yang, Hao Xue, Imran Razzak, Hakim Hacid, and Flora D. Salim. Reloop: Recur- sive retrieval with multi-hop reasoner and planners for heterogeneous qa.arXiv preprint arXiv:2510.20505, 2025. 11

work page internal anchor Pith review Pith/arXiv arXiv 2025
[30]

Tree of thoughts: Deliberate problem solving with large language models.Ad- vances in neural information processing systems, 36:11809–11822, 2023

Shunyu Yao, Dian Yu, Jeffrey Zhao, Izhak Shafran, Tom Griffiths, Yuan Cao, and Karthik Narasimhan. Tree of thoughts: Deliberate problem solving with large language models.Ad- vances in neural information processing systems, 36:11809–11822, 2023

work page 2023
[31]

React: Synergizing reasoning and acting in language models

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik R Narasimhan, and Yuan Cao. React: Synergizing reasoning and acting in language models. InThe eleventh international conference on learning representations, 2022

work page 2022
[32]

which parameters?

Yao Yao, Zuchao Li, and Hai Zhao. Got: Effective graph-of-thought reasoning in language models. InFindings of the Association for Computational Linguistics: NAACL 2024, pages 2901–2921, 2024. 12 A Theoretical Proofs and Structural Properties This appendix collects proofs, structural properties, and auxiliary analysis for STAR. Theorem 1 is stated in the m...

work page 2024