Pith. sign in

REVIEW 5 major objections 5 minor 48 references

Speculate While You Reason: Teaching Agents to Predict Their Next Tool Call via Joint Agent-Speculator RL

T0 review · 5 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read A single model can serve as its own tool-call speculator, lifting prediction accuracy from roughly 44% to 61% without hurting task success.

desk verdict The self-speculation method is plausible and worth a referee, but the abstract's 'single model' claim is false as written: Table 2 averages two separately trained checkpoints, and there is no latency measurement anywhere. read the letter →

arxiv 2607.25816 v1 pith:7PFPLZO2 submitted 2026-07-28 cs.AI

Jiabao Ji , Yujian Liu , Li An , Rohit Jain , Gungor Polatkan , Siyu Zhu , Shiyu Chang This is my paper · ORCID
classification cs.AI
keywords tool-callspeculationself-speculatingagentreinforcementlearningLLMagentslatencyhidingnext-callpredictionHit@1on-policytraining
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper argues that the best predictor of an agent's next tool call is the agent itself, and that a single model can be trained to both solve tasks and speculate on its next call. It introduces a training scheme that alternates between agent-mode and speculator-mode reinforcement learning, using the agent's own rollouts as speculation targets. On two agentic task families, this raises next-call prediction accuracy (Hit@1) from around 44–49% to 61–66% while leaving task success roughly unchanged. If the result holds, tool-call latency could be hidden by pre-executing predicted calls, with no separate draft model or trace cache.

What carries the argument

The self-speculating agent: a single policy that behaves as the agent for full trajectories and, when a fixed speculation suffix is appended to a partial trajectory prefix, generates a candidate next tool call. Training alternates between agent-mode updates (task reward, with grouped relative advantages) and speculator-mode updates (a name-conditioned argument-F1 reward) on on-policy rollouts, with optimizer resets at each mode switch and an SFT warmup. The reuse of the prefix KV cache makes self-speculation cheap at inference time.

What would settle it

Measure the average wall-clock time per task for a deployed system running the trained self-speculator with prefetch and compare it to the base agent without speculation, on the same benchmarks. If Hit@1 goes up but latency does not improve (or worsens), the core claim fails. A second falsifier: a larger model scale where joint RL degrades task success, contradicting the 'preserving task performance' claim.

Watch

Extended reading notes

Core claim

The paper's central claim is that a single language model can serve as both the task-solving agent and its own speculator—predicting the next structured tool call from a partial trajectory after a fixed suffix. This works because the speculator-agent gap (the mismatch between an external draft model's predictions and the deployed agent's actual calls) is largest when the speculator is a different model; the agent itself is the best off-the-shelf predictor. The training method, joint agent-speculator RL, samples rollouts from the current policy, uses the agent's own next calls as speculation targets, and alternates between agent and speculator updates with optimizer resets. The result is that

Load-bearing premise

The load-bearing premise is that a higher Hit@1 against the agent's own calls directly yields lower wall-clock latency in a real serving system; the paper never measures end-to-end time.

Editorial extensions

If this is right

  • A single model can replace separate draft models or trace caches for tool-call speculation, cutting serving memory and complexity.
  • Next-call Hit@1 can be improved by roughly 17 percentage points with reinforcement learning without sacrificing task success, across two model families.
  • Call-level speculation ability transfers across domains, but end-task success still requires matched-domain training.
  • The alternating, on-policy training recipe is a general template for learning auxiliary self-predictions without degrading the main task.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The paper's reliance on Hit@1 rather than measured wall-clock time leaves open whether the 17-point Hit@1 gain translates to actual latency hiding; an end-to-end latency benchmark with a prefetch scheduler would settle this.
  • The method assumes read-only tools; extending self-speculation to state-changing actions would require a mechanism to filter or roll back speculative executions.
  • The finding that the agent itself beats smaller same-family drafts suggests any external speculator faces a fundamental distribution-mismatch ceiling; this motivates on-policy, self-referential speculation as a general principle.
  • The alternating-schedule result (4:8 outperforms 1:1) implies that speculation and agent objectives need separated, batched optimization; this could generalize to other multi-objective agent training.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes the self-speculating agent: a single LLM that is trained to act both as a task-solving agent and as a speculator that predicts the agent's own next tool call from a partial trajectory using a fixed suffix and shared prefix KV cache. The method, joint agent-speculator RL, alternates between DAPO-style agent updates and speculator updates, with targets derived from the current policy's own rollouts, plus SFT warmup and optimizer resets for stability. On two 4B backbones, the paper reports large average next-tool-call Hit@1 gains on five benchmarks (44.1→61.2 for Qwen3-4B; 48.9→66.3 for Qwen3.5-4B) while claiming task success is preserved. The paper also includes an off-the-shelf speculator comparison, cross-domain generalization experiments, and ablations of the training recipe.

Significance. If the headline claim were supported, this would be a practically relevant contribution: it removes the need for a separate draft model or trace cache, reuses the prefix KV cache, and directly aligns the speculator with the deployed agent. The online evaluation protocol (speculating from the true trajectory prefix while the environment only executes the agent's call) is well designed, and the exact-match Hit@1 metric correctly reflects the reuse condition. The method is clearly described and the ablation study addresses important stabilization questions. However, the significance is currently undercut by a load-bearing reporting problem: the main results table appears to average over two separately trained checkpoints, so the 'single model' claim is not actually demonstrated.

major comments (5)
  1. [Table 2, §4.2, Abstract] The 'RL' rows in Table 2 are composite averages over two separately trained checkpoints, not a single model. For Qwen3-4B-RL, Table 2 lists HotpotQA 55.2/MuSiQue 58.4 and τ-bench Airline 68.0/Retail 72.3; Table 3, which evaluates the SearchQA-trained checkpoint on τ-bench, reports the same HotpotQA/MuSiQue values but Airline/Retail 45.6/50.1. The τ-bench numbers therefore come from a different ToolScale-trained run. The same holds for Qwen3.5-4B-RL: Table 2's HotpotQA/MuSiQue 58.6/61.8 differ from Table 3's ToolScale-trained values 35.4/37.9, while the τ-bench columns match. The abstract's 'single model' claim and the average Hit@1 improvements (44.1→61.2, 48.9→66.3) are not supported by any one checkpoint. This is load-bearing because the central design claim is unification. Please report each checkpoint separately and either train/evaluate one joint model or revise the claims according
  2. [§4.1, Tables 2 and 7] No repeated runs, seeds, or error bars are reported. Evaluation sizes are small: τ-bench Airline uses 50 queries, τ-bench Retail 114, BCP 150, and SearchQA subsets 500 per dataset. The 'preserving task success' claim rests on small absolute changes (e.g., average success 26.6→27.7 for Qwen3-4B and 49.2→50.6 for Qwen3.5-4B), which are within plausible noise for these sample sizes. Without variance estimates or at least multiple seeds, the improvement in Hit@1 and the stability of task success are not reliably established. Please provide seed-level results or confidence intervals, especially for the main Table 2 comparisons.
  3. [§4.4, Table 4] The 4:8 agent:speculator schedule is selected from Table 4, which reports results on HotpotQA, MuSiQue, and BCP — the same benchmarks used in the main results of Table 2. Since the schedule is a free hyperparameter and the differences between schedules are large (1:1 gives 31.8 average Hit@1; 4:8 gives 55.2), selection on the evaluation sets can inflate the reported gains. Please either use a held-out validation split for schedule selection, or explicitly acknowledge and correct for this selection effect in the main numbers.
  4. [§1, §4.1] The motivating payoff of tool-call speculation is reduced wall-clock latency, but the paper never measures end-to-end latency or speedup for the trained self-speculating agent. Hit@1 is a proxy: even with high Hit@1, branch decoding overhead, scheduling constraints, or environment restrictions on pre-execution may prevent the predicted calls from being reused in practice. Section 4.1 states that the chosen metrics 'capture the central trade-off,' but no evidence is provided that improved Hit@1 translates into actual latency reduction. Please report at least a simulation or a small serving-system experiment with pre-execution, or temper the motivation claims accordingly.
  5. [§3.2, §4.1, Eq. (1)] The speculation target is the current policy's own call from the same rollout, and the speculation reward Rsp is a shaped version of the exact-match metric reported as Hit@1. Because the objective is to maximize agreement with the policy's own next call, part of the Hit@1 gain is an expected self-consistency improvement rather than evidence of a generally better speculator. The paper should rule out the possibility that the policy becomes easier to predict (e.g., by collapsing toward a small set of template calls) while superficially preserving task success. Reporting output diversity, repeated-call rates, or Hit@1 against a fixed reference agent's calls would strengthen the external validity. This is particularly important because no latency measurement (previous comment) independently validates the usefulness of the improved Hit@1.
minor comments (5)
  1. [§4.2] Typo: 'self-sepculative' should be 'self-speculative'; also 'instanstite' in §4.1 and 'scenerio' in the Introduction.
  2. [Table 3] The table caption and column headers should explicitly state which training domain each checkpoint comes from (SearchQA-trained vs. ToolScale-trained). Currently the reader must infer this from the text, and the shared 'Qwen3-4B' row labels are ambiguous.
  3. [Table 2 caption] The note about external speculator task success being the target agent's success is helpful, but the table should also state that the RL rows combine two separately trained checkpoints, if that remains the presentation.
  4. [§4.1 / Table 7] The SearchQA evaluation uses '500 examples per dataset sampled from the FlashRAG development splits.' Please clarify whether these subsets are disjoint from the FlashQA training distribution and whether they were fixed before running experiments, to avoid train/eval leakage concerns.
  5. [Appendix A.2] The speculation suffix is fixed across all models and tasks, which is good for consistency. However, the suffix contains the word 'I'll look it up,' which may be less appropriate for τ-bench-style API calls than for search; a brief discussion of suffix sensitivity would be useful.

Circularity Check

1 steps flagged · score 4.0 of 10

Speculation target and headline metric are both the agent's own call, making part of the Hit@1 gain a self-consistency fit; external task-success benchmarks keep the paper from full circularity.

  1. self definitional [Sec. 3.1 (Eq. 1) and Sec. 4.1 (Evaluation protocol and metrics)]
    "Rsp = Rname · Rargs ... The target is the agent’s own next call at from the sampled trajectory. Since at comes from the current policy, the targets change as the agent changes, keeping speculation training on-policy with the deployed agent. ... We report Hit@1 exact match, where a hit requires both the tool name and the full argument dictionary to match the agent’s eventual call."

    The RL reward optimized during training (Eq. 1, a macro-F1 smoothed agreement with the agent's own sampled call) is a soft version of the exact-match Hit@1 metric reported as the headline improvement. Both the training target and the evaluation label are defined as the current agent's own next tool call from its own rollouts. Thus the reported Hit@1 gain (44.1→61.2, 48.9→66.3) measures how well the model was fitted to agree with itself, i.e., a self-consistency improvement, rather than an externally grounded predictive claim. The task-success columns are grounded in external benchmarks and are not circular, so the circularity is partial.

full rationale

The derivation chain is largely empirical rather than formal, and most of the paper's machinery is not circular: the joint agent–speculator RL alternation, the optimizer resets, and the DAPO updates are standard optimization choices, and the task-success preservation claim is validated against external benchmarks (HotpotQA, MuSiQue, BCP, τ-bench). The one genuinely self-referential element is the speculation target: Sec. 3.1 defines the target as the current agent's own sampled next call, and Sec. 4.1 defines the headline Hit@1 as exact agreement with that same call. Since the shaped reward in Eq. (1) is a smoothed version of the exact-match metric, the reported Hit@1 gains are partly the expected result of optimizing that objective on the agent's own rollouts; this is a partial self-definitional circularity, not a fully independent prediction. The task-success numbers and the preservation claim are external, which prevents a higher score. No load-bearing self-citation, imported uniqueness theorem, or ansatz-smuggled-via-citation was found; the authors' self-citations are contextual and non-essential. Separately, the cross-referenced Table 2/Table 3 discrepancy (RL rows appearing to average separately trained SearchQA and ToolScale checkpoints) is a correctness and reporting risk, not a circularity, and is therefore excluded from the circularity score under the rule that non-circularity concerns belong in correctness risk.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper's central numbers come from a self-referential training/evaluation loop (agent predicts its own calls and is rewarded for matching them), plus standard RL/optimization assumptions. No new physical or architectural entities are introduced; the fixed speculation suffix is a prompt-level design choice rather than an invented entity.

free parameters (5)
  • Agent:speculator update schedule = 4:8 (four agent updates, eight speculator updates)
    Chosen via ablation on SearchQA (Table 4) and applied to all reported settings; no held-out validation of the schedule choice.
  • Rollout group sizes Gag, Gsp = 8, 8
    Default DAPO-style grouped optimization sizes; not swept in this paper.
  • Learning rate = 3e-6
    Fixed default; could affect dual-mode stability but is not analyzed.
  • RL budget = 200 steps per setting
    Fixed step count with no early stopping tied to evaluation; stopping point is arbitrary.
  • Speculation suffix = Fixed '<think>...' prompt
    Ad hoc prompt design; paper claims it does not leak benchmark/tool/argument names, but no sensitivity analysis is provided.
assumptions (5)
  • domain assumption The agent's actual next call from its own rollout is the correct speculation target (predict-the-agent, not predict-the-ideal tool).
    Defines Rsp in Eq. (1) and the on-policy targets in Algorithm 1; external optimality of tool calls is never used.
  • domain assumption Speculative calls are read-only and can be discarded without environment side effects.
    Required for safe pre-execution; acknowledged in Limitations as not valid for state-mutating tools such as placing orders or sending messages.
  • domain assumption Hit@1 exact structured-call match is a sufficient proxy for the latency benefit of tool-call speculation.
    Evaluation Section 4.1 uses Hit@1 as the primary metric; no wall-clock latency is measured post-training.
  • standard math DAPO grouped relative advantage estimation with group size 8 provides valid policy-gradient estimates in both modes.
    Borrowed from DAPO (Yu et al., 2025); not re-derived and no bias analysis for the dual-mode setting is given.
  • ad hoc to paper Optimizer reset at mode switches and SFT warmup stabilize dual-mode training.
    Empirical components of the recipe (Section 3.3, Figure 3); no theoretical justification, only ablations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Speculate While You Reason: Teaching Agents to Predict Their Next Tool Call via Joint Agent-Speculator RL." pith.science (2026). https://pith.science/paper/7PFPLZO2

@misc{pith2026260725816,
  author       = {Pith},
  title        = {Pith review of: Speculate While You Reason: Teaching Agents to Predict Their Next Tool Call via Joint Agent-Speculator RL},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7PFPLZO2}},
  note         = {Machine review of arXiv:2607.25816}
}
read the original abstract

Large language model agents often spend substantial wall-clock time waiting for tool call results. Tool-call speculation can hide this latency by predicting and pre-executing an agent's next tool call if the prediction matches the agent's eventual tool call, but existing speculators are typically separate draft models or cached traces that are poorly aligned with the deployed agent's own behavior. We identify this speculator-agent gap and show that the target agent itself is a strong next-call speculator. This points to a simpler design: unifying the agent and speculator within the same model. In this paper, we introduce the self-speculating agent, a single model that both solves tasks in agent mode and predicts its next tool call from partial trajectories in speculator mode, fully reusing prefix KV cache. To enable this dual-mode agent without degrading performance, we propose a joint agent-speculator reinforcement learning method, which derives speculation targets from the agent's own rollouts and alternates agent and speculator updates. Across agentic search QA and conversational tool-use agentic tasks, our method improves average next tool-call Hit@1 from 44.1 to 61.2 for Qwen3-4B and from 48.9 to 66.3 for Qwen3.5-4B, while preserving agent task success.

Figures

Figures reproduced from arXiv: 2607.25816 by the authors.

Figure 1
Figure 1. External versus self tool-call speculation. The [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Online tool-call speculation example. From [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Reward trend over the first 100 steps for each [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 5 linked inside Pith

  1. [1]

    Shao, Zhihong and Wang, Peiyi and Zhu, Qihao and Xu, Runxin and Song, Junxiao and Bi, Xiao and Zhang, Haowei and Zhang, Mingchuan and Li, Y. K. and Wu, Yu and Guo, Daya , journal =. 2024 , url =

    2024

  2. [2]

    2025 , url =

    Jin, Bowen and Zeng, Hansi and Yue, Zhenrui and Yoon, Jinsung and Arik, Sercan and Wang, Dong and Zamani, Hamed and Han, Jiawei , journal =. 2025 , url =

    2025

  3. [3]

    2025 , url =

    Song, Huatong and Jiang, Jinhao and Min, Yingqian and Chen, Jie and Chen, Zhipeng and Zhao, Wayne Xin and Fang, Lei and Wen, Ji-Rong , journal =. 2025 , url =

    2025

  4. [4]

    2025 , url =

    Tan, Zhiwen and Huang, Jiaming and Wu, Qintong and Zhang, Hongxuan and Zhuang, Chenyi and Gu, Jinjie , journal =. 2025 , url =

    2025

  5. [5]

    2025 , url =

    Yu, Qiying and Zhang, Zheng and Zhu, Ruofei and Yuan, Yufeng and Zuo, Xiaochen and Yue, Yu and Fan, Tiantian and Liu, Gaohong and Liu, Lingjun and Liu, Xin and Lin, Haibin and Lin, Zhiqi and Ma, Bole and Sheng, Guangming and Tong, Yuxuan and Zhang, Chi and Zhang, Mofan and Zhang, Wang and Zhu, Hang and Zhu, Jinhua and Chen, Jiaze and Chen, Jiangjie and Wa...

    2025

  6. [6]

    arXiv preprint arXiv:2302.04761 , year =

    Schick, Timo and Dwivedi-Yu, Jane and Dess. arXiv preprint arXiv:2302.04761 , year =

    arXiv

  7. [7]

    and Zhang, Tianjun and Wang, Xin and Gonzalez, Joseph E

    Patil, Shishir G. and Zhang, Tianjun and Wang, Xin and Gonzalez, Joseph E. , journal =. 2023 , url =

    2023

  8. [8]

    2023 , url =

    Chen, Baian and Shu, Chang and Shareghi, Ehsan and Collier, Nigel and Narasimhan, Karthik and Yao, Shunyu , journal =. 2023 , url =

    2023

Show all 48 references
  1. [9]

    2024 , url =

    Liu, Zuxin and Hoang, Thai and Zhang, Jianguo and Zhu, Ming and Lan, Tian and Kokane, Shirley and Tan, Juntao and Yao, Weiran and Liu, Zhiwei and Feng, Yihao and Murthy, Rithesh and Yang, Liangwei and Savarese, Silvio and Niebles, Juan Carlos and Wang, Huan and Heinecke, Shelb...

    2024

  2. [10]

    arXiv preprint arXiv:2504.13958 , year =

    Qian, Cheng and Acikgoz, Emre Can and He, Qi and Wang, Hongru and Chen, Xiusi and Hakkani-T. arXiv preprint arXiv:2504.13958 , year =

    arXiv

  3. [11]

    2025 , url =

    Zhang, Yabo and Zeng, Yihan and Li, Qingyun and Hu, Zhen and Han, Kavin and Zuo, Wangmeng , journal =. 2025 , url =

    2025

  4. [12]

    Reinforcement Learning for Long-Horizon Interactive

    Chen, Kevin and Cusumano-Towner, Marco and Huval, Brody and Petrenko, Aleksei and Hamburger, Jackson and Koltun, Vladlen and Kr. Reinforcement Learning for Long-Horizon Interactive. arXiv preprint arXiv:2502.01600 , year =

    arXiv

  5. [13]

    Reinforcing Multi-Turn Reasoning in

    Wei, Quan and Zeng, Siliang and Li, Chenliang and Brown, William and Frunza, Oana and Deng, Wei and Schneider, Anderson and Nevmyvaka, Yuriy and Zhao, Yang Katie and Garcia, Alfredo and Hong, Mingyi , journal =. Reinforcing Multi-Turn Reasoning in. 2025 , url =

    2025

  6. [14]

    and Yang, Yuqing , journal =

    Luo, Xufang and Zhang, Yuge and He, Zhiyuan and Wang, Zilong and Zhao, Siyun and Li, Dongsheng and Qiu, Luna K. and Yang, Yuqing , journal =. Agent Lightning: Train. 2025 , url =

    2025

  7. [15]

    Proceedings of the 40th International Conference on Machine Learning , pages =

    Fast Inference from Transformers via Speculative Decoding , author =. Proceedings of the 40th International Conference on Machine Learning , pages =. 2023 , url =

    2023

  8. [16]

    Proceedings of the 33rd International Conference on Machine Learning , pages =

    Asynchronous Methods for Deep Reinforcement Learning , author =. Proceedings of the 33rd International Conference on Machine Learning , pages =. 2016 , url =

    2016

  9. [17]

    and Quan, John and Kirkpatrick, James and Hadsell, Raia and Heess, Nicolas and Pascanu, Razvan , booktitle =

    Teh, Yee and Bapst, Victor and Czarnecki, Wojciech M. and Quan, John and Kirkpatrick, James and Hadsell, Raia and Heess, Nicolas and Pascanu, Razvan , booktitle =. 2017 , url =

    2017

  10. [18]

    Proceedings of the 35th International Conference on Machine Learning , pages =

    Espeholt, Lasse and Soyer, Hubert and Munos, R. Proceedings of the 35th International Conference on Machine Learning , pages =. 2018 , url =

    2018

  11. [19]

    Proceedings of the 30th International Conference on Machine Learning , pages =

    On the Importance of Initialization and Momentum in Deep Learning , author =. Proceedings of the 30th International Conference on Machine Learning , pages =. 2013 , url =

    2013

  12. [20]

    International Conference on Learning Representations , year =

    Adam: A Method for Stochastic Optimization , author =. International Conference on Learning Representations , year =

  13. [21]

    International Conference on Learning Representations , year =

    Speculative Actions: A Lossless Framework for Faster Agentic Systems , author =. International Conference on Learning Representations , year =

  14. [22]

    arXiv preprint arXiv:2512.15834 , year =

    Optimizing Agentic Language Model Inference via Speculative Tool Calls , author =. arXiv preprint arXiv:2512.15834 , year =

  15. [23]

    2026 , url =

    Zhong, Shuzhang and Lu, Baotong and Chen, Qi and Liu, Chuanjie and Yang, Fan and Li, Meng , journal =. 2026 , url =

    2026

  16. [24]

    Act While Thinking: Accelerating

    Sui, Yifan and Zhao, Han and Ma, Rui and He, Zhiyuan and Wang, Hao and Li, Jianxun and Yang, Yuqing , journal =. Act While Thinking: Accelerating. 2026 , url =

    2026

  17. [25]

    2025 , url =

    Qi, Zehan and Liu, Xiao and Iong, Iat Long and Lai, Hanyu and Sun, Xueqiao and Zhao, Wenyi and Yang, Yu and Yang, Xinyue and Sun, Jiadai and Yao, Shuntian and Zhang, Tianjie and Xu, Wei and Tang, Jie and Dong, Yuxiao , journal =. 2025 , url =

    2025

  18. [26]

    2025 , url =

    Zheng, Yuxiang and Fu, Dayuan and Hu, Xiangkun and Cai, Xiaojie and Ye, Lyumanshan and Lu, Pengrui and Liu, Pengfei , journal =. 2025 , url =

    2025

  19. [27]

    2025 , url =

    Li, Xuefeng and Zou, Haoyang and Liu, Pengfei , journal =. 2025 , url =

    2025

  20. [28]

    2025 , url =

    Feng, Jiazhan and Huang, Shijue and Qu, Xingwei and Zhang, Ge and Qin, Yujia and Zhong, Baoquan and Jiang, Chengquan and Chi, Jinxin and Zhong, Wanjun , journal =. 2025 , url =

    2025

  21. [29]

    Agentic Reasoning and Tool Integration for

    Singh, Joykirat and Magazine, Raghav and Pandya, Yash and Nambi, Akshay , journal =. Agentic Reasoning and Tool Integration for. 2025 , url =

    2025

  22. [30]

    Reducing Latency of

    Huang, Zixiao and Zeng, Wen and Fu, Tianyu and Liu, Tengxuan and Sun, Yizhou and Hong, Ke and Yang, Xinhao and Liu, Chengchun and Li, Yan and Zhang, Quanlu and Dai, Guohao and Zhu, Zhenhua and Wang, Yu , journal =. Reducing Latency of. 2025 , url =

    2025

  23. [31]

    and Shao, Yakun Sophia and Gholami, Amir and Keutzer, Kurt , journal =

    Hooper, Coleman and Kang, Minwoo and Moon, Suhong and Lee, Nicholas and Wen, Eric and Wawrzynek, John and Mahoney, Michael W. and Shao, Yakun Sophia and Gholami, Amir and Keutzer, Kurt , journal =. Speculative Interaction Agents: Building Real-Time Agents with Asynchronous. 20...

    2026

  24. [32]

    Advances in Neural Information Processing Systems , volume =

    Training Language Models to Follow Instructions with Human Feedback , author =. Advances in Neural Information Processing Systems , volume =. 2022 , url =

    2022

  25. [33]

    Constitutional

    Bai, Yuntao and Kadavath, Saurav and Kundu, Sandipan and Askell, Amanda and Kernion, Jackson and others , journal =. Constitutional. 2022 , url =

    2022

  26. [34]

    arXiv preprint arXiv:2501.12948 , year =

    arXiv

  27. [35]

    arXiv preprint arXiv:2501.12599 , year =

    arXiv

  28. [36]

    2025 , url =

    Hou, Bairu and Zhang, Yang and Ji, Jiabao and Liu, Yujian and Qian, Kaizhi and Andreas, Jacob and Chang, Shiyu , journal =. 2025 , url =

    2025

  29. [37]

    2025 , url =

    Liu, Yujian and Ji, Jiabao and Zhang, Yang and Guo, Wenbo and Jaakkola, Tommi and Chang, Shiyu , journal =. 2025 , url =

    2025

  30. [38]

    2026 , url =

    He, Zhongmou and Choi, Yee Man and Zhang, Kexun and Bercovich, Ivan and Ji, Jiabao and Zhou, Junting and Xu, Dejia and Zhang, Aidan and Zeng, Yixiao and Li, Lei , booktitle =. 2026 , url =

    2026

  31. [39]

    How Well Do Agentic Skills Work in the Wild: Benchmarking

    Liu, Yujian and Ji, Jiabao and An, Li and Jaakkola, Tommi and Zhang, Yang and Chang, Shiyu , journal =. How Well Do Agentic Skills Work in the Wild: Benchmarking. 2026 , url =

    2026

  32. [40]

    Collision- and Reachability-Aware Multi-Robot Control with Grounded

    Ji, Jiabao and Chen, Yongchao and Zhang, Yang and Kompella, Ramana Rao and Fan, Chuchu and Liu, Gaowen and Chang, Shiyu , journal =. Collision- and Reachability-Aware Multi-Robot Control with Grounded. 2025 , url =

    2025

  33. [41]

    2024 , url =

    Yao, Shunyu and Shinn, Noah and Razavi, Pedram and Narasimhan, Karthik , journal =. 2024 , url =

    2024

  34. [42]

    2025 , url =

    Barres, Victor and Dong, Honghua and Ray, Soham and Si, Xujie and Narasimhan, Karthik , journal =. 2025 , url =

    2025

  35. [43]

    2025 , url =

    Su, Hongjin and Diao, Shizhe and Lu, Ximing and Liu, Mingjie and Xu, Jiacheng and Dong, Xin and Fu, Yonggan and Belcak, Peter and Ye, Hanrong and Yin, Hongxu and Dong, Yi and Bakhturina, Evelina and Yu, Tao and Choi, Yejin and Kautz, Jan and Molchanov, Pavlo , journal =. 2025 , url =

    2025

  36. [44]

    2025 , url =

    Yang, An and Li, Anfeng and Yang, Baosong and Zhang, Beichen and Hui, Binyuan and Zheng, Bo and Yu, Bowen and Gao, Chang and Huang, Chengen and Lv, Chenxu and others , journal =. 2025 , url =

    2025

  37. [45]

    2026 , howpublished =

    2026

  38. [46]

    2025 , url =

    Chen, Zijian and Ma, Xueguang and Zhuang, Shengyao and Nie, Ping and Zou, Kai and Liu, Andrew and Green, Joshua and Patel, Kshama and Meng, Ruoxi and Su, Mingyi and Sharifymoghaddam, Sahel and Li, Yanxi and Hong, Haoran and Shi, Xinyu and Liu, Xuye and Thakur, Nandan and Zhang...

    2025

  39. [47]

    Scaling Long-Horizon

    Sun, Yunhao and Zhang, Yadong and Shi, Weijia and Hu, Yushi and Li, Xinze and Chen, Mingyang and Xiao, Zilin and Zhang, Chi , journal =. Scaling Long-Horizon. 2025 , url =

    2025

  40. [48]

    2021 , journal =

    MuSiQue: Multihop Questions via Single-hop Question Composition , author =. 2021 , journal =

    2021

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.