Pith. sign in

REVIEW 4 major objections 4 minor 53 references

Planning once with enforced waits and on-demand replanning, SyncPlan claims LLM-based multi-agent coordination can be both adaptive and fast, beating repeated-invocation baselines in task success while using under 0.05% of their wall-clock

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

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

A plan-execute-correct framework with explicit wait primitives and a learned plan-staleness detector enables LLM-based multi-agent coordination that is both faster and more adaptive than repeated-invocation baselines.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection SyncPlan's explicit synchronization primitives are a genuine step forward for LLM coordination, but the learned staleness detector's coverage blind spot and Overcooked train/eval overlap undercut the adaptivity headline. the 4 major comments →

arxiv 2608.01652 v1 pith:JCSKPUKQ submitted 2026-08-03 cs.RO cs.AI

SyncPlan: Long-Horizon LLM Coordination with Explicit Synchronization and Adaptive Correction

Shen You , Xiaoming Zhu , Weining Weng , Hefei Mei , Weixuan Wang , Zhongshen Li , Zeji LI , Ye-Wen Wang
show 12 more authors
Zijun Liao Juchao Zhuo Yang Wei Fuhao Qiu Siqin Li Zhenjie Lian Danei Gong Junkai Ji Xiangtao Li Qiuzhen Lin Liang Wang Ka-Chun Wong
This is my paper
classification cs.RO cs.AI
keywords multi-agent coordinationLLM planningsynchronization primitivesplan staleness detectionreinforcement learningOvercookedHonor of Kingslong-horizon planning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

SyncPlan claims that the perceived trade-off between efficiency and adaptivity in LLM-based multi-agent coordination is not fundamental. Instead of re-invoking the LLM during execution, the framework produces the whole per-agent action plan in one call and keeps it alive through three mechanisms: wait primitives that turn coordination dependencies into machine-checkable conditions, a deadlock detector, and a lightweight learned Plan Staleness Detector that triggers replanning only when the remaining plan has actually been invalidated. Because replanning is event-driven rather than periodic, coordination adapts to environmental changes without paying for constant LLM inference: on Overcooked dynamic settings SyncPlan reaches 88.3-100% task achievement versus 60-96% for LLM baselines, with under 0.05% of their runtime, and on Honor of Kings it reaches 86.3% versus 68.75% for the best baseline with a 26x runtime reduction. The paper's central assertion is that when to correct matters more than how often, and that a cheap synchronous execution layer plus selective correction makes one-shot planning viable for real-time environments.

Core claim

The central claim is that a single LLM planning call can carry long-horizon multi-agent coordination if two runtime mechanisms are added: explicit synchronization primitives that make the plan's dependencies executable, and a fast Plan Staleness Detector that decides, from the remaining action chains and current entity states, when the plan is no longer valid. When the detector or the wait-graph cycle check flags a problem, the coordinator is re-invoked only at that moment. Optimized by SFT and planning-oriented RL with a composite reward over task progress and execution outcomes, this design reaches 93-100% task achievement on Overcooked dynamic settings (88.3% on the hardest Ring layout) a

What carries the argument

The load-bearing piece is the Plan Staleness Detector (PSD): a small binary classifier fed by three tensors — the chain-entity incidence matrix Ct marking which entities the unexecuted plan steps reference, the normalized entity-state matrix Xt, and the state-difference matrix Δt. Two bilinear pathways embed the chain structure and the state changes separately, and their interaction is mapped through an MLP to a staleness probability Preplan(t); a replan is triggered when it exceeds a threshold. Around it sits the execution layer: per-agent chain pointers advance only when actions complete, Wait_agents and Wait_entity primitives block an agent until a teammate milestone or environment condit

Load-bearing premise

The Plan Staleness Detector only monitors entities that a hand-written relevance rule marks in Ct; any plan-invalidating change on an unmarked entity passes unseen and the agents keep executing a stale plan — and the paper never shows that its relevance rules cover all disturbance types that actually occur.

What would settle it

Run SyncPlan with an injected event that invalidates the plan through an entity the relevance rule ignores — for example, a sudden enemy threat appearing just outside the fixed radius in Honor of Kings, or an Overcooked pot that changes state while no remaining step references it — and record whether the PSD triggers a replan and whether task success holds. If the system proceeds with the stale chain, the adaptivity claim is falsified; measuring PSD recall on such out-of-prior events would settle it directly.

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

If this is right

  • If the central claim holds, real-time LLM coordination does not require repeated planning: a single invocation plus cheap staleness monitoring is enough to stay adaptive.
  • Coordination dependencies expressed as wait predicates become verifiable at runtime, so LLM plans no longer have to be trusted as natural language.
  • Event-driven replanning dominates fixed-frequency replanning: periodic re-invocation at 4 or 8 frames per second pays an order-of-magnitude runtime penalty for equal or worse task success.
  • Coordinator training matters as much as the framework: RL with execution-level feedback lifts task success (e.g., 79.0% to 86.3% on Honor of Kings) by aligning plans with the executor's real timing.
  • Below roughly one billion parameters, coordinators cannot reliably emit the structured action chains, placing a hard floor on model capacity for this style of planning.

Where Pith is reading between the lines

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

  • Editorial inference: the PSD's relevance prior is the untested boundary — a change affecting an entity the rule-based Ct does not mark (e.g., an enemy appearing just outside the fixed radius) would be invisible, and the reported high recall is measured on perturbations that the rule already covers.
  • Editorial inference: if a cheap detector can decide when a plan is invalid, a natural next step the paper does not take is local chain repair — patching only the invalidated segment of the remaining plan rather than full replanning, which would cut latency further.
  • Editorial inference: the plan-execute-correct division should transfer to other embodied or robotic settings, but only where an author can hand-write the relevance rule; the paper demonstrates two environments, not a general recipe.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. SyncPlan proposes a plan-execute-correct framework for long-horizon LLM coordination. A centralized LLM coordinator generates per-agent action chains in a single planning call; a runtime executor enforces explicit Wait_agents/Wait_entity synchronization primitives and performs cycle-based deadlock detection; a lightweight Plan Staleness Detector (PSD) decides when the remaining plan is invalid and triggers replanning. The coordinator is optimized via SFT and GRPO with a composite reward. Experiments on Overcooked (three layouts, standard and dynamically perturbed settings) and Honor of Kings (5v5 Commander mode) report higher task achievement rates than LLM baselines at substantially lower wall-clock runtime, with ablations isolating the contributions of the PSD, SFT, and RL.

Significance. The central hypothesis—that a one-shot joint plan plus explicit synchronization plus a lightweight staleness detector can match or exceed the adaptivity of repeated LLM replanning at a fraction of the latency—is well-motivated and practically important. The paper's strengths include a broad evaluation across two environments, a detailed ablation of PSD and periodic replanning, a PSD architecture search, RL training curves, model-scaling analysis, and unusually complete appendices with reward coefficients, thresholds, hyperparameters, and case-study traces. However, the load-bearing component of the adaptivity claim is the PSD, and the current evidence for its general validity is incomplete: the Overcooked PSD is trained on the same scripted event types used at evaluation, and the hand-built C_t relevance mask is not validated for coverage of plan-invalidating changes outside that prior.

major comments (4)
  1. [§4.4 / Appendix A.6 / Eq. (6)] The PSD's staleness signal is structurally blind to changes on entities not marked by the hand-built C_t. Since H_C in Eq. (4) is built from C_t, any entity j with no nonzero entry in C_t has a zero row in H_C and contributes nothing to H_C^T H_X in Eq. (6). Appendix A.6/Fig. 10 defines C_t by fixed-radius proximity (HoK) or by the current step template (Overcooked). No experiment checks whether plan-invalidating changes ever occur on unmarked entities. As a concrete counterexample, a pot-content change in Overcooked is invisible if the current step does not mention the pot. Because 'trigger replanning only when needed' is the core adaptivity claim, this coverage gap must be tested—e.g., by injecting perturbations on unmarked entities and measuring PSD recall and downstream plan success.
  2. [Appendix A.6 / B.4] The Overcooked PSD is trained on labels derived from the same scripted event injection protocol used in evaluation: positive frames are exactly the injected-event frames (A.6), and the dynamic evaluation uses the same three event types (Table 8, B.4). Table 11's high PSD recall is therefore an in-distribution measure. The dynamic-task advantage over w/o PSD may partly reflect the detector having been tuned to these three event types rather than detecting general plan invalidation. The authors should either evaluate on a held-out event type (e.g., an enemy rotation or a resource change not in O1–O3) or explicitly bound the adaptivity claim to the three tested perturbation categories.
  3. [Table 1] Table 1 reports only point estimates and permutation-test stars; no error bars, confidence intervals, or trial counts are shown for the Overcooked main comparison. Given that the HoK ablations in Tables 2 and 4 report standard deviations of 3–10 TAR points, several contrasted differences in Table 1 (e.g., Dynamic Ring 88.3% vs 76.5% for w/o SFT) could be within noise. Report per-seed results and confidence intervals, especially for the claimed SOTA comparisons.
  4. [§5.1 vs Appendix A.7] The provenance of the SFT corpus is inconsistent. §5.1 states Overcooked SFT data comes from Gemini model distillation, while Appendix A.7 describes constructing SFT QA pairs by replaying 'prompt-only Coordinator' trajectories and treating each chain-update event as a supervision boundary. These are different data sources and imply different claims about independence from the method itself. If the SFT corpus is partly self-generated by the prompt-only SyncPlan, the SFT warm-start is iterative self-imitation rather than expert demonstration, which should be stated and analyzed (e.g., with a distillation-only SFT baseline).
minor comments (4)
  1. [Equation (1)] The notation 'C:S × E →0,1' should read '→{0,1}'; several set/arrow symbols are malformed throughout the problem formulation.
  2. [§5.1 / Table 1] Deadlock Rate is listed as a reported metric, but Table 1 has no Deadlock Rate column; deadlock information appears only in training curves (Fig. 5/14). Either add the column or remove the metric from the setup.
  3. [Abstract / §5.2] The 'less than 0.05% of wall-clock runtime' figure is computed for a specific Overcooked comparison (A-ToM vs Ours Full) and should be contextualized in the main text; as written in the abstract it reads like a universal bound.
  4. [Appendix A.8] The semantic penalty equation is incomplete: 'Psem(A) = λ sem' is missing its operand. Table 7 gives the value, but the equation should be fixed.

Circularity Check

2 steps flagged

Overcooked PSD is trained on the same injected event protocol used in dynamic evaluation, and the SFT corpus is bootstrapped from the authors' own prompt-only coordinator; the core framework still has independent support from HoK.

specific steps
  1. fitted input called prediction [Appendix A.6 (Overcooked: event-frame labeling) and Section B.4 (Dynamic Event Taxonomy and Trigger Protocols)]
    "Overcooked: event-frame labeling. We reuse the dynamic-event injection protocol (Section B.4): the frame at which a dynamic event is injected is labelled yt = 1, since the event is designed to invalidate at least one field of the active chain. ... All three events are triggered from a seeded RNG so that runs across methods and seeds see the exact same sequence."

    The PSD's positive supervision is defined by the same dynamic-event injection protocol used to create the Overcooked dynamic evaluation settings. The detector is therefore trained to fire on the exact perturbation types (held-item drop, pot scatter, agent displacement) that the evaluation injects. The adaptive-correction gain reported in Table 1 and the PSD ablation in Table 3 thus measures recognition of training-distribution events, not a demonstrated ability to detect arbitrary plan-invalidating changes. This is a fitted input called prediction: the 'prediction' of when to replan is statistically forced by the training labels.

  2. other [Appendix A.7 (SFT Data Augmentation, Construction protocol)]
    "We replay SyncPlan trajectories collected from the prompt-only Coordinator and treat each chain-update event as a supervision boundary."

    The SFT warm-start is trained on trajectories generated by the same SyncPlan prompt-only coordinator that serves as the w/o-SFT baseline. The 'expert/distilled' supervision is therefore the method's own output, making the SFT gain a form of self-imitation rather than evidence from independent expert data. This does not by itself invalidate the framework, but it means the SFT improvement is a bootstrapping effect, not an externally grounded prediction.

full rationale

SyncPlan's central contribution is a plan-execute-correct loop with explicit wait primitives, deadlock detection, and a lightweight learned staleness detector. This architecture is genuine and is not, by construction, equivalent to its inputs. The Honor of Kings evaluation is injection-free: HoK dynamics arise from uncontrolled bot behavior, and the PSD is trained offline on synthetic perturbations (health reduction, enemy repositioning) rather than on the exact evaluation frames. That provides independent, external support for the framework's adaptivity claim. However, the Overcooked dynamic evaluation is significantly weakened by a training/evaluation overlap: the PSD's positive labels are generated by reusing the same dynamic-event injection protocol (held-item drop, pot-content scatter, agent displacement) that defines the dynamic evaluation settings. Consequently, the reported 8+ TAR improvement from the PSD in the ablation (Table 3) and the strong dynamic-setting results (Table 1) partly reflect in-distribution detection of the exact events on which the detector was trained, not demonstrated generalization to novel plan-invalidating changes. Additionally, the SFT corpus is bootstrapped from the authors' own prompt-only coordinator, so the SFT warm-start is a self-imitation/distillation step rather than independent expert supervision. These issues are real but do not reduce the whole derivation to its inputs: the HoK results, the synchronization primitives, and the RL reward design stand on their own. No load-bearing self-citation chain or imported uniqueness theorem was found. Score 4 reflects partial circularity in the Overcooked adaptive-correction evidence and the self-bootstrapped SFT data, while acknowledging the independent HoK evaluation and the non-circular core architecture.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The paper introduces algorithmic constructs: wait primitives, a PSD classifier, and a composite reward. No new physical entities are postulated. The load-bearing adaptive mechanism rests on hand-set reward coefficients, a PSD threshold selected by grid search, and heuristic entity-relevance rules whose coverage is not exhaustively validated.

free parameters (5)
  • PSD replan threshold η = 0.7
    Grid-searched over {0.3, 0.5, 0.7, 0.9} on a held-out validation set (A.6); directly controls when replanning fires, trading recall against overhead.
  • GRPO success bonus α = 0.45
    Hand-set reward coefficient in Eq. (7) and Table 7; shapes the RL objective by rewarding full task completion.
  • GRPO penalty coefficients λsyn, λsem, λdlk, λtmo = 0.1, 0.3, 1.0, 0.0
    Hand-set penalty schedule in Table 7; the paper states this numerical ordering is the one used in training, but provides no sensitivity analysis.
  • Progress reward weights wfrac, weff, wlen and length range = 0.15, 0.15, 0.25; [100, 500] tokens
    Hand-set weights and length regularization range in A.8; these determine the dense progress signal that drives RL.
  • HoK entity-relevance radius for C_t = unspecified
    The rule-based incidence matrix C_t includes entities within a fixed radius of the self-agent or target (A.6, Figure 10); the radius value is not reported, so a load-bearing relevance prior is under-specified.
axioms (4)
  • domain assumption Low-level agents reliably execute emitted action primitives and the runtime executor faithfully enforces wait predicates.
    The whole plan-execute-correct loop depends on the executor mapping high-level steps to atomic environment actions (Sec 4.3, A.5).
  • domain assumption Any plan-invalidating environmental change is observable in the normalized entity-state tensors X_t and Δ_t.
    PSD inputs (Sec 4.4) only cover F entity properties; changes outside these features are invisible to the detector, so the claim of adaptive correction is bounded by feature coverage.
  • ad hoc to paper The rule-based chain-entity incidence matrix C_t marks every entity whose change could invalidate a remaining chain step.
    C_t is constructed by spatial proximity in HoK and action-object binding in Overcooked (A.6, Figure 10); coverage is asserted but not validated against counterexamples, and unmarked entity changes go undetected.
  • domain assumption The five terminal outcomes (Success, SyntaxError, SemanticError, Deadlock, Timeout) are mutually exclusive and exhaustive for chain execution.
    The RL reward in Eq. (7) and A.8 relies on this partition; if execution can terminate in other modes, reward attribution is misspecified.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of SyncPlan: Long-Horizon LLM Coordination with Explicit Synchronization and Adaptive Correction." pith.science (2026). https://pith.science/paper/JCSKPUKQ

@misc{pith2026260801652,
  author       = {Pith},
  title        = {Pith review of: SyncPlan: Long-Horizon LLM Coordination with Explicit Synchronization and Adaptive Correction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JCSKPUKQ}},
  note         = {Machine review of arXiv:2608.01652}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

LLM-based multi-agent coordination faces a fundamental trade-off between efficiency and adaptivity in dynamic environments. Existing approaches typically rely on repeated LLM invocations or multi-round communication to adapt decisions during execution, introducing substantial latency and making coordination vulnerable to asynchronous progress and environmental changes. Conversely, one-shot planning reduces coordination overhead but produces open-loop plans that can quickly become stale or fail when actions depend on other agents and the environment. We introduce SyncPlan, a plan-execute-correct framework for long-horizon coordination through explicit synchronization and adaptive correction. Given the state and team-level task, a centralized LLM coordinator generates per-agent action chains in a single planning call. During execution, explicit wait primitives and deadlock detection enforce inter-agent and agent-environment dependencies, while a lightweight Plan Staleness Detector continuously assesses the remaining plan and triggers replanning when environmental changes invalidate its assumptions. We further optimize the coordinator through SFT and planning-oriented RL with dense task progress and outcome-level execution feedback. Experiments on the public Overcooked benchmark and the complex Honor of Kings environment show that SyncPlan achieves state-of-the-art task success rates while using less than 0.05% of the wall-clock runtime compared with existing LLM-based coordinators. Code and datasets will be made publicly available.

Figures

Figures reproduced from arXiv: 2608.01652 by Danei Gong, Fuhao Qiu, Hefei Mei, Juchao Zhuo, Junkai Ji, Ka-Chun Wong, Liang Wang, Qiuzhen Lin, Shen You, Siqin Li, Weining Weng, Weixuan Wang, Xiangtao Li, Xiaoming Zhu, Yang Wei, Ye-Wen Wang, Zeji Li, Zhenjie Lian, Zhongshen Li, Zijun Liao.

Figure 1
Figure 1. Figure 1: Core challenges of multi-agent coordination: real￾time constraints, plan staleness, and coordination conflicts. 2022; Kuba et al. 2022) require extensive task-specific in￾teraction and careful reward engineering, while complex coordination preferences are difficult to encode as scalar re￾wards. Recent LLM-based coordinators (Guo et al. 2025b; Zhang et al. 2024) offer a more expressive alternative by interp… view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework. (A) SyncPlan forms a plan-execute-correct loop: collaborative planning generates per-agent action chains; runtime execution advances the chains and detects deadlocks; and the PSD monitors plan validity. (B–C) SFT warm-starts the coordinator with expert and distilled data, followed by planning-oriented RL with execution feedback. 4.2 Collaborative Planning The objective of collaborative p… view at source ↗
Figure 3
Figure 3. Figure 3: Runtime execution. The module advances per￾agent action chains and enforces synchronization through agent- and entity-dependent wait conditions. 4.3 Runtime Execution Considering the generated high-level plan cannot be directly executed by the low-level agents, especially the synchro￾nization primitives, we introduce the execution module to maintain and synchronize the action chain. This module en￾ables ex… view at source ↗
Figure 4
Figure 4. Figure 4: Architecture of PSD. The chain and state pathways encode remaining-plan structure and entity-state changes, whose interaction produces the replanning probability. deployment latency, and a decision quality ceiling bounded by pre-trained knowledge. Therefore, we adopt an SFT and RL training pipeline to achieve both real-time responsiveness and superior performance on smaller models. SFT Warm-Start SFT initi… view at source ↗
Figure 5
Figure 5. Figure 5: (A) The ROC curve of PSD model. (B) RL training curves. (C) Task performance during RL. (D) Deadlock ratios. [PITH_FULL_IMAGE:figures/full_fig_p006_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Effect of coordinator model size on SyncPlan performance. The critical threshold lies between 1B and 1.5B. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance on main backbone models. plans, fails to account for the bot executor’s actual movement timing and collision dynamics, leading to poor coordination quality. RL enhances coordination by providing closed-loop feedback that helps the model learn execution-aligned strate￾gies absent from offline demonstrations. The HoK ablation ( [PITH_FULL_IMAGE:figures/full_fig_p007_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: PSD-triggered replanning in two scenarios. [PITH_FULL_IMAGE:figures/full_fig_p007_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Overview of the two-stage SFT+RL training pipeline for the SyncPlan Coordinator. Stage 1 performs supervised fine-tuning; Stage 2 applies GRPO reinforcement learning with environment-grounded rewards. "agent_2":{action1, action2, ...} ... } Slots in braces (Background, Target, Num of Agents, Descrip￾tion, Usage, State) are template variables filled at planning time; tokens inside the Assistant blocks are g… view at source ↗
Figure 10
Figure 10. Figure 10: Constructing the remaining-plan chain-entity in [PITH_FULL_IMAGE:figures/full_fig_p012_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Event-driven SFT data augmentation. Each dashed [PITH_FULL_IMAGE:figures/full_fig_p013_11.png] view at source ↗
Figure 13
Figure 13. Figure 13: The HoK game map. space: enemy support can arrive from any direction with any composition; held items can drop at any frame; an ally can disconnect or be teleported mid-chain. Enumerating this space exhaustively is impossible. Instead, for evaluation we define a small set of dynamic event types that provides broad coverage of the main failure modes. Each type is triggered by a seeded script in Overcooked … view at source ↗
Figure 12
Figure 12. Figure 12: Our experiments are conducted on three different [PITH_FULL_IMAGE:figures/full_fig_p015_12.png] view at source ↗
Figure 14
Figure 14. Figure 14: In-game performance variation over the course of RL training. [PITH_FULL_IMAGE:figures/full_fig_p016_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Training curves (reward and loss) during RL optimization. [PITH_FULL_IMAGE:figures/full_fig_p016_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Overcooked case study on the Coordination layout. When the active chains create mutually dependent Wait_agents predicates, the incremental reachability test in Section A.4 detects the resulting wait cycle. The re-invoked Plan Module emits a revised chain with an explicit Wait_agents step that serializes tile access, allowing both agents to complete delivery without collision [PITH_FULL_IMAGE:figures/full… view at source ↗
Figure 17
Figure 17. Figure 17: Honor of Kings case study: the PSD detects an [PITH_FULL_IMAGE:figures/full_fig_p017_17.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

53 extracted references · 31 canonical work pages · 3 internal anchors

  1. [1]

    International Conference on Machine Learning , pages=

    QMIX: Monotonic Value Function Factorisation for Deep Multi-Agent Reinforcement Learning , author=. International Conference on Machine Learning , pages=

  2. [2]

    arXiv preprint arXiv:2511.11182 , year=

    Multi-agent Undercover Gaming: Hallucination Removal via Counterfactual Test for Multimodal Reasoning , author=. arXiv preprint arXiv:2511.11182 , year=

    arXiv

  3. [3]

    Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    Adaptive theory of mind for LLM-based multi-agent coordination , author=. Proceedings of the AAAI Conference on Artificial Intelligence , volume=

    work page

  4. [4]

    arXiv preprint arXiv:2507.06261 , year=

    Gemini 2.5: Pushing the frontier with advanced reasoning, multimodality, long context, and next generation agentic capabilities , author=. arXiv preprint arXiv:2507.06261 , year=

    Pith/arXiv arXiv

  5. [5]

    HiComm: Hierarchical Communication for Multi-agent Reinforcement Learning

    HiComm: Hierarchical Communication for Multi-agent Reinforcement Learning , author=. arXiv preprint arXiv:2606.29126 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  6. [6]

    ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=

    Doctoragent-rl: A multi-agent collaborative reinforcement learning system for multi-turn clinical dialogue , author=. ICASSP 2026-2026 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP) , pages=. 2026 , organization=

    work page 2026

  7. [7]

    International Conference on Learning Representations , volume=

    Overcookedv2: Rethinking overcooked for zero-shot coordination , author=. International Conference on Learning Representations , volume=

  8. [8]

    International Conference on Machine Learning , pages=

    Cooperative open-ended learning framework for zero-shot coordination , author=. International Conference on Machine Learning , pages=. 2023 , organization=

    work page 2023

  9. [9]

    Advances in neural information processing systems , volume=

    On the utility of learning about humans for human-ai coordination , author=. Advances in neural information processing systems , volume=

  10. [10]

    Proceedings of the Eighth Fact Extraction and VERification Workshop (FEVER) , pages=

    The law of knowledge overshadowing: Towards understanding, predicting, and preventing llm hallucination , author=. Proceedings of the Eighth Fact Extraction and VERification Workshop (FEVER) , pages=

    work page

  11. [11]

    arXiv preprint arXiv:1711.09846 , year=

    Population Based Training of Neural Networks , author=. arXiv preprint arXiv:1711.09846 , year=

    Pith/arXiv arXiv

  12. [12]

    Collab-Overcooked: Benchmarking and Evaluating Large Language Models as Collaborative Agents

    Sun, Haochen and Zhang, Shuwen and Niu, Lujie and Ren, Lei and Xu, Hao and Fu, Hao and Zhao, Fangkun and Yuan, Caixia and Wang, Xiaojie. Collab-Overcooked: Benchmarking and Evaluating Large Language Models as Collaborative Agents. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.249

    doi:10.18653/v1/2025.emnlp-main.249 2025

  13. [13]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    Llmarena: Assessing capabilities of large language models in dynamic multi-agent environments , author=. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    work page

  14. [14]

    Advances in Neural Information Processing Systems , volume=

    Cooperation, competition, and maliciousness: Llm-stakeholders interactive negotiation , author=. Advances in Neural Information Processing Systems , volume=

  15. [15]

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

    Hallulens: Llm hallucination benchmark , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    work page

  16. [16]

    International Conference on Machine Learning , pages=

    QTRAN: Learning to Factorize with Transformation for Cooperative Multi-Agent Reinforcement Learning , author=. International Conference on Machine Learning , pages=

    work page

  17. [17]

    Advances in Neural Information Processing Systems , year=

    The Surprising Effectiveness of PPO in Cooperative Multi-Agent Games , author=. Advances in Neural Information Processing Systems , year=

    work page

  18. [18]

    International Conference on Learning Representations , year=

    Trust Region Policy Optimisation in Multi-Agent Reinforcement Learning , author=. International Conference on Learning Representations , year=

  19. [19]

    arXiv preprint arXiv:2508.21365 , year=

    Think in games: Learning to reason in games via reinforcement learning with large language models , author=. arXiv preprint arXiv:2508.21365 , year=

    Pith/arXiv arXiv

  20. [20]

    Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    Distilling reasoning capabilities into smaller language models , author=. Findings of the Association for Computational Linguistics: ACL 2023 , pages=

    2023

  21. [21]

    arXiv preprint arXiv:2512.04797 , year=

    Sima 2: A generalist embodied agent for virtual worlds , author=. arXiv preprint arXiv:2512.04797 , year=

    arXiv

  22. [22]

    Advances in neural information processing systems , volume=

    Collaborating with humans without human data , author=. Advances in neural information processing systems , volume=

  23. [23]

    AAAI Conference on Artificial Intelligence , year=

    ProAgent: Building Proactive Cooperative Agents with Large Language Models , author=. AAAI Conference on Artificial Intelligence , year=

    work page

  24. [24]

    Guo, Jie and Li, Yingjun and Wang, Yijie and Feng, Zhonghan and Zha, Daochen , booktitle=

    work page

  25. [25]

    Proceedings of NAACL , year=

    Are Large Language Models Good at Coordination? An Extensive Study of LLM Performance in Coordination Games , author=. Proceedings of NAACL , year=

    work page

  26. [26]

    Liu, Siqi and Wang, Hao and Chen, Yifan and Zhang, Jiarui and Yang, Yaodong , booktitle=

    work page

  27. [27]

    Zhang, Hongxin and Du, Zeyuan and Qi, Tielin and Zhao, Dong and others , booktitle=

    work page

  28. [28]

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

    Leveraging dual process theory in language agent framework for real-time simultaneous human-AI collaboration , author=. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , pages=

    work page

  29. [29]

    Guo, Daya and Yang, Dejian and Zhang, He and Song, Junxiao and Zhang, Runxin and Xu, Ruoyu and Zhu, Qihao and Ma, Shirong and Wang, Peiyi and Bi, Xiao and others , journal=

    work page

  30. [30]

    Conference on Robot Learning , year=

    Do As I Can, Not As I Say: Grounding Language in Robotic Affordances , author=. Conference on Robot Learning , year=

  31. [31]

    Advances in Neural Information Processing Systems , year=

    Voyager: An Open-Ended Embodied Agent with Large Language Models , author=. Advances in Neural Information Processing Systems , year=

    work page

  32. [32]

    Conference on Robot Learning , year=

    Inner Monologue: Embodied Reasoning through Planning with Language Models , author=. Conference on Robot Learning , year=

    work page

  33. [33]

    Proceedings of the AAAI Conference on Artificial Intelligence , year=

    MindAgent: Emergent Gaming Interaction , author=. Proceedings of the AAAI Conference on Artificial Intelligence , year=

    work page

  34. [34]

    arXiv preprint arXiv:2309.17277 , year=

    Suspicion-Agent: Playing Imperfect Information Games with Theory of Mind Aware GPT-4 , author=. arXiv preprint arXiv:2309.17277 , year=

    Pith/arXiv arXiv

  35. [35]

    Advances in Neural Information Processing Systems , year=

    SPRING: Studying the Paper and Reasoning to Play Games , author=. Advances in Neural Information Processing Systems , year=

    work page

  36. [36]

    AvalonBench: Evaluating

    Light, Jonathan and Cai, Min and Gupta, Sheng and Hu, Ziniu and others , booktitle=. AvalonBench: Evaluating

    work page

  37. [37]

    ER-FSL: Experience Replay with Feature Subspace Learning for Online Continual Learning

    Werewolf Among Us: A Multimodal Dataset for Modeling Persuasion Behaviors in Social Deduction Games , author=. arXiv preprint arXiv:2407.12279 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  38. [38]

    International Conference on Machine Learning , pages=

    Machine Theory of Mind , author=. International Conference on Machine Learning , pages=

    work page

  39. [39]

    International Conference on Autonomous Agents and Multi-Agent Systems , pages=

    Learning with Opponent-Learning Awareness , author=. International Conference on Autonomous Agents and Multi-Agent Systems , pages=

    work page

  40. [40]

    Geng, Minghong and Pateria, Shubham and others , booktitle=

    work page

  41. [41]

    Xu, Zhiwei and Bai, Yunpeng and Zhang, Bin and Li, Dapeng and Fan, Guoliang , booktitle=

    work page

  42. [42]

    ReCAPA: Hierarchical Predictive Correction to Mitigate Cascading Failures

    Recapa: Hierarchical predictive correction to mitigate cascading failures , author=. arXiv preprint arXiv:2604.21232 , year=

    work page internal anchor Pith review Pith/arXiv arXiv

  43. [43]

    Applied Sciences , volume=

    Target-oriented multi-agent coordination with hierarchical reinforcement learning , author=. Applied Sciences , volume=. 2024 , publisher=

    work page 2024

  44. [44]

    IEEE International Conference on Computer Vision , pages=

    Focal Loss for Dense Object Detection , author=. IEEE International Conference on Computer Vision , pages=

    work page

  45. [45]

    McDermott, Drew and Ghallab, Malik and Howe, Adele and Knoblock, Craig and Ram, Ashwin and Veloso, Manuela and Weld, Daniel and Wilkins, David , institution=

    work page

  46. [46]

    Behavior Trees in Robotics and

    Colledanchise, Michele and. Behavior Trees in Robotics and. CRC Press , year=

    work page

  47. [47]

    5-coder technical report , author=

    Qwen2. 5-coder technical report , author=. arXiv preprint arXiv:2409.12186 , year=

    Pith/arXiv arXiv

  48. [48]

    arXiv preprint arXiv:2407.21783 , year=

    The llama 3 herd of models , author=. arXiv preprint arXiv:2407.21783 , year=

    Pith/arXiv arXiv

  49. [49]

    2026 , eprint=

    GLM-5: from Vibe Coding to Agentic Engineering , author=. 2026 , eprint=

    2026

  50. [50]

    arXiv preprint arXiv:2303.08774 , year=

    Gpt-4 technical report , author=. arXiv preprint arXiv:2303.08774 , year=

    Pith/arXiv arXiv

  51. [51]

    arXiv preprint arXiv:2312.11805 , year=

    Gemini: a family of highly capable multimodal models , author=. arXiv preprint arXiv:2312.11805 , year=

    Pith/arXiv arXiv

  52. [52]

    arXiv preprint arXiv:2504.07491 , year=

    Kimi-vl technical report , author=. arXiv preprint arXiv:2504.07491 , year=

    Pith/arXiv arXiv

  53. [53]

    arXiv preprint arXiv:2605.26494 , year=

    The minimax-m2 series: Mini activations unleashing max real-world intelligence , author=. arXiv preprint arXiv:2605.26494 , year=

    Pith/arXiv arXiv

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.