pith. sign in

arxiv: 2605.22154 · v1 · pith:EIYLHDMGnew · submitted 2026-05-21 · 💻 cs.AI

IdleSpec: Exploiting Idle Time via Speculative Planning for LLM Agents

Daewon Choi , Kyunghyun Park , Woomin Song , Saket Dingliwal , Sai Muralidhar Jayanthi , Jinwoo Shin , Aram Galstyan This is my paper

Pith reviewed 2026-05-22 06:17 UTC · model grok-4.3

classification 💻 cs.AI
keywords LLM agentsidle timespeculative planningplan aggregationobservation uncertaintyagentic workflowsGAIA benchmark
0
0 comments X

The pith

IdleSpec turns waiting periods in LLM agents into speculative plan generation that raises task accuracy.

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

The paper shows that LLM agents spend significant time idle while awaiting tool results or environment feedback, and that this time can be repurposed for useful work. IdleSpec generates multiple plan candidates during those intervals by sampling between progressive and recovery drafting strategies drawn from a distribution refined by past outcomes. When the real observation arrives the candidates are aggregated to shape the immediate next reasoning step. This yields measurable gains on multi-step benchmarks while adding no extra end-to-end latency, which matters for any deployment where tool calls or code executions create natural pauses.

Core claim

IdleSpec is a generic inference approach that exploits idle time by iteratively producing plan candidates under observation uncertainty and aggregating them once observations become available. It draws samples from a learned distribution over two complementary drafting strategies—progressive, which extends current information, and recovery, which prepares fallback paths—and updates the distribution via posterior feedback from completed episodes. Experiments confirm that this procedure improves agent performance across varied scenarios without increasing latency.

What carries the argument

Idle-time speculative plan generation followed by observation-triggered aggregation, with sampling between progressive and recovery drafting strategies drawn from a posterior-updated distribution.

If this is right

  • Agent accuracy rises on benchmarks that interleave reasoning with tool calls or code execution.
  • Long-horizon tasks with large execution delays benefit without extra wall-clock time.
  • The method requires no change to the underlying language model and works across different models.
  • Latency overhead stays near zero because all added work occurs inside existing idle windows.

Where Pith is reading between the lines

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

  • Similar idle-time speculation could be inserted into other sequential AI systems that wait on external services.
  • Online adaptation of the drafting distribution might further reduce reliance on completed-task feedback.
  • The approach may encourage agents to maintain multiple contingency plans rather than committing early to a single path.

Load-bearing premise

Plans drafted without the actual observation can still be combined to produce a better next step than would have been chosen without them.

What would settle it

An experiment on GAIA or FRAMES in which the IdleSpec agent shows no accuracy improvement over a matched baseline that performs no idle-time computation.

Figures

Figures reproduced from arXiv: 2605.22154 by Aram Galstyan, Daewon Choi, Jinwoo Shin, Kyunghyun Park, Sai Muralidhar Jayanthi, Saket Dingliwal, Woomin Song.

Figure 1
Figure 1. Figure 1: Overview of IdleSpec. (a) Idle-Time Drafting: during tool execution, the agent iteratively drafts plan candidates by sampling between Progressive and Recovery strategies, and terminates drafting once the observation arrives. (b) Draft Aggregation: the agent aggregates the candidates with the observation into a refined action and forecasts whether the trajectory is on track or requires recovery. (c) Posteri… view at source ↗
Figure 2
Figure 2. Figure 2: How Can We Leverage Idle Time in LLM Agents? (a) Reasoning time vs. tool execution time (i.e., idle time) across benchmarks. (b) Histogram of per-call tool execution times. (c) Accuracy of three idle-time strategies (Summarization, Reflection, Planning) vs. vanilla. progressive and recovery drafts, and exploits idle time to improve task performance rather than to amortize already-required computation. Rece… view at source ↗
Figure 3
Figure 3. Figure 3: Latency–Accuracy Trade-off. All measurements were performed on vLLM using an NVIDIA A6000 GPU [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Idle-Time Utilization (ITU) vs. Accuracy Gain (∆). GAIA with Qwen3.5-4B, one point per difficulty level [PITH_FULL_IMAGE:figures/full_fig_p015_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Progressive drafting prompt used by IdleSpec to speculatively generate the next action [PITH_FULL_IMAGE:figures/full_fig_p022_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Recovery drafting prompt used by IdleSpec to draft an alternative plan that diverges from [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Forecast prompt used by IdleSpec after the observation arrives to choose between the [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Aggregation prompt that consumes the candidate plans together with the just-arrived [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Sequential Revision prompt. The model reflects on the executed action and its observation [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Sleep-Time Compute prompt. The model is asked to pre-compute inferences and [PITH_FULL_IMAGE:figures/full_fig_p023_10.png] view at source ↗
read the original abstract

Large language model (LLM)-based agents solve complex tasks by leveraging multi-step reasoning with iterative tool calls and environment interactions, which incur idle time while waiting for observations. Despite the prevalence of idle time in most agentic scenarios, existing works treat it as an unavoidable overhead or propose restricted solutions that overlook varying computational budgets across different tool calls and future observation uncertainty, thereby leading to suboptimal utilization of idle time. In this paper, we introduce IdleSpec, a scalable and generic inference approach that leverages idle-time computation to improve agent performance while minimizing latency overhead. Specifically, IdleSpec iteratively generates plan candidates during idle periods and, once observations become available, aggregates them to guide the next reasoning step. For effective plan generation under observation uncertainty, IdleSpec samples between complementary drafting strategies (i.e., progressive and recovery) from a learned distribution that is updated via posterior feedback. Our experiments demonstrate that IdleSpec significantly improves agent performance in various agentic scenarios by effectively utilizing idle time. In particular, on the GAIA and FRAMES, IdleSpec achieves 55.6% average accuracy with Gemini-2.5-Flash, surpassing the vanilla baseline without idle-time usage by 5.1%. Furthermore, for MLE-Bench, which involves substantial delay from code executions, IdleSpec achieves performance gains of up to 9.1% on the Any Medal rate, highlighting its generalizability to long-horizon tasks.

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.

Referee Report

2 major / 2 minor

Summary. The paper introduces IdleSpec, an inference-time method for LLM agents that exploits idle time during tool calls and environment interactions by iteratively generating plan candidates under observation uncertainty. It samples drafting strategies (progressive and recovery) from a learned distribution updated via posterior feedback and aggregates the candidates to guide the next reasoning step. Experiments report concrete gains, including 55.6% average accuracy on GAIA and FRAMES with Gemini-2.5-Flash (5.1% above vanilla baseline) and up to 9.1% improvement on Any Medal rate for MLE-Bench.

Significance. If the gains prove robust and stem from the uncertainty-aware aggregation rather than raw extra compute, the work could meaningfully advance practical idle-time utilization in agentic systems, especially for variable-delay and long-horizon tasks. The generic, scalable framing and multi-benchmark evaluation are strengths that would support broader adoption if the mechanism is shown to be load-bearing.

major comments (2)
  1. [§4 Experiments] §4 Experiments: The central performance claim (55.6% accuracy, +5.1% on GAIA/FRAMES) lacks ablations that isolate the aggregation operator and learned distribution from equivalent additional token budget spent on non-speculative planning; without this control, it is unclear whether reported gains exceed what extra compute alone would produce.
  2. [§3 Method] §3 Method: The update rule for the learned distribution over progressive/recovery strategies via posterior feedback and the precise aggregation procedure for plan candidates under observation uncertainty are not specified in sufficient detail to verify that they reduce uncertainty rather than add noise, which directly bears on the soundness of the 5.1% and 9.1% gains.
minor comments (2)
  1. [Abstract] The abstract would be strengthened by briefly naming the aggregation operator used once observations arrive.
  2. [§3.2] Notation for the drafting strategies and posterior update could be clarified with a short pseudocode snippet or additional equation.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive report. We address each major comment below and outline the revisions we will make to improve the manuscript's clarity and rigor.

read point-by-point responses

  1. Referee: [§4 Experiments] §4 Experiments: The central performance claim (55.6% accuracy, +5.1% on GAIA/FRAMES) lacks ablations that isolate the aggregation operator and learned distribution from equivalent additional token budget spent on non-speculative planning; without this control, it is unclear whether reported gains exceed what extra compute alone would produce.

    Authors: We agree this is a valuable control and thank the referee for highlighting it. In the revised manuscript we will add ablations that allocate an equivalent additional token budget to non-speculative planning during idle periods (e.g., repeated standard reasoning steps without progressive/recovery drafting or learned aggregation). These experiments will directly compare against IdleSpec to isolate the contribution of the uncertainty-aware components. We have already initiated these runs on the GAIA/FRAMES suite and will report the full results. revision: yes

  2. Referee: [§3 Method] §3 Method: The update rule for the learned distribution over progressive/recovery strategies via posterior feedback and the precise aggregation procedure for plan candidates under observation uncertainty are not specified in sufficient detail to verify that they reduce uncertainty rather than add noise, which directly bears on the soundness of the 5.1% and 9.1% gains.

    Authors: We acknowledge that §3 would benefit from greater precision. In the revision we will expand the method section with the exact posterior update rule (including the likelihood model and feedback weighting) and the full aggregation procedure (e.g., how candidate plans are scored and combined under partial observations). These additions will make the mechanism verifiable and will explicitly show how the approach is designed to reduce rather than amplify uncertainty. revision: yes

Circularity Check

0 steps flagged

IdleSpec method and gains are empirically validated without reducing to self-referential inputs or fitted parameters by construction

full rationale

The paper introduces IdleSpec as a new inference-time approach that generates speculative plan candidates during idle periods, aggregates them upon observation, and samples drafting strategies from a distribution updated by posterior feedback. These elements are presented as algorithmic innovations whose value is demonstrated through benchmark experiments (GAIA, FRAMES, MLE-Bench) comparing against a vanilla baseline without idle-time usage. No equations or derivations in the provided text reduce the reported accuracy improvements (e.g., +5.1% on GAIA/FRAMES) to the inputs by construction, nor does the central claim depend on self-citations or uniqueness theorems imported from prior author work. The performance claims rest on external empirical measurement rather than tautological re-expression of the method itself.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The approach rests on the assumption that idle periods are long enough for useful plan generation and that aggregation of uncertain plans yields net benefit; no explicit free parameters or invented entities are named in the abstract.

axioms (1)
  • domain assumption Plan candidates generated under observation uncertainty can be aggregated to improve the next reasoning step.
    This premise is required for the aggregation step to produce the claimed performance gains.

pith-pipeline@v0.9.0 · 5813 in / 1248 out tokens · 28472 ms · 2026-05-22T06:17:23.701909+00:00 · methodology

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

33 extracted references · 33 canonical work pages · 8 internal anchors

  1. [1]

    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, 2023

    work page 2023

  2. [2]

    Self-generated in-context examples improve LLM agents for sequential decision-making tasks.arXiv preprint arXiv:2505.00234, 2025

    Vishnu Sarukkai, Zhiqiang Xie, and Kayvon Fatahalian. Self-generated in-context examples improve LLM agents for sequential decision-making tasks.arXiv preprint arXiv:2505.00234, 2025

    work page arXiv 2025

  3. [3]

    GAIA: a benchmark for general AI assistants

    Grégoire Mialon, Clémentine Fourrier, Thomas Wolf, Yann LeCun, and Thomas Scialom. GAIA: a benchmark for general AI assistants. InThe Twelfth International Conference on Learning Representations, 2024

    work page 2024

  4. [4]

    Humanity's Last Exam

    Long Phan, Alice Gatti, Ziwen Han, Nathaniel Li, Josephina Hu, Hugh Zhang, Chen Bo Calvin Zhang, Mohamed Shaaban, John Ling, Sean Shi, et al. Humanity’s last exam.arXiv preprint arXiv:2501.14249, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    WebArena: A Realistic Web Environment for Building Autonomous Agents

    Shuyan Zhou, Frank F Xu, Hao Zhu, Xuhui Zhou, Robert Lo, Abishek Sridhar, Xianyi Cheng, Tianyue Ou, Yonatan Bisk, Daniel Fried, et al. WebArena: A realistic web environment for building autonomous agents.arXiv preprint arXiv:2307.13854, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  6. [6]

    Mind2web: Towards a generalist agent for the web.Advances in Neural Information Processing Systems, 36:28091–28114, 2023

    Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Sam Stevens, Boshi Wang, Huan Sun, and Yu Su. Mind2web: Towards a generalist agent for the web.Advances in Neural Information Processing Systems, 36:28091–28114, 2023

    work page 2023

  7. [7]

    SWE-bench: Can Language Models Resolve Real-World GitHub Issues?

    Carlos E Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. SWE-bench: Can language models resolve real-world GitHub issues?arXiv preprint arXiv:2310.06770, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  8. [8]

    MLE-bench: Evaluating Machine Learning Agents on Machine Learning Engineering

    Jun Shern Chan, Neil Chowdhury, Oliver Jaffe, James Aung, Dane Sherburn, Evan Mays, Giulio Starace, Kevin Liu, Leon Maksin, Tejal Patwardhan, et al. MLE-bench: Evaluating machine learning agents on machine learning engineering.arXiv preprint arXiv:2410.07095, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    The AI Scientist: Towards Fully Automated Open-Ended Scientific Discovery

    Chris Lu, Cong Lu, Robert Tjarko Lange, Jakob Foerster, Jeff Clune, and David Ha. The AI scientist: Towards fully automated open-ended scientific discovery.arXiv preprint arXiv:2408.06292, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  10. [10]

    CodeAgent: Enhancing code generation with tool-integrated agent systems for real-world repo-level coding challenges.arXiv preprint arXiv:2401.07339, 2024

    Kechi Zhang, Jia Li, Ge Li, Xianjie Shi, and Zhi Jin. CodeAgent: Enhancing code generation with tool-integrated agent systems for real-world repo-level coding challenges.arXiv preprint arXiv:2401.07339, 2024

    work page arXiv 2024

  11. [11]

    Agent-as-Tool: A study on the hierarchical decision making with reinforcement learning.arXiv preprint arXiv:2507.01489, 2025

    Yanfei Zhang. Agent-as-Tool: A study on the hierarchical decision making with reinforcement learning.arXiv preprint arXiv:2507.01489, 2025

    work page arXiv 2025

  12. [12]

    Adaptation of agentic AI: A survey of post-training, memory, and skills.arXiv preprint arXiv:2512.16301, 2025

    Pengcheng Jiang, Jiacheng Lin, Zhiyi Shi, Zifeng Wang, Luxi He, Yichen Wu, Ming Zhong, Peiyang Song, Qizheng Zhang, Heng Wang, et al. Adaptation of agentic AI: A survey of post-training, memory, and skills.arXiv preprint arXiv:2512.16301, 2025

    work page arXiv 2025

  13. [13]

    Asynchronous LLM function calling.arXiv preprint arXiv:2412.07017, 2024

    In Gim, Seung-seob Lee, and Lin Zhong. Asynchronous LLM function calling.arXiv preprint arXiv:2412.07017, 2024

    work page arXiv 2024

  14. [14]

    Gonzalez

    Kevin Lin, Charlie Snell, Yu Wang, Charles Packer, Sarah Wooders, Ion Stoica, and Joseph E. Gonzalez. Sleep-time compute: Beyond inference scaling at test-time, 2025. URL https: //arxiv.org/abs/2504.13171

    work page arXiv 2025

  15. [15]

    Fact, fetch, and reason: A unified evaluation of retrieval-augmented generation, 2024

    Satyapriya Krishna, Kalpesh Krishna, Anhad Mohananey, Steven Schwarcz, Adam Stambler, Shyam Upadhyay, and Manaal Faruqui. Fact, fetch, and reason: A unified evaluation of retrieval-augmented generation, 2024. URLhttps://arxiv.org/abs/2409.12941

    work page arXiv 2024

  16. [16]

    Chain-of-thought prompting elicits reasoning in large language models

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022. 10

    work page 2022

  17. [17]

    Pre-Act: Multi-step planning and reasoning improves acting in LLM agents.arXiv preprint arXiv:2505.09970, 2025

    Mrinal Rawat, Ambuje Gupta, Rushil Goomer, Alessandro Di Bari, Neha Gupta, and Roberto Pieraccini. Pre-Act: Multi-step planning and reasoning improves acting in LLM agents.arXiv preprint arXiv:2505.09970, 2025

    work page arXiv 2025

  18. [18]

    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

  19. [19]

    Tool learning with large language models: A survey.Frontiers of Computer Science, 19(8):198343, 2025

    Changle Qu, Sunhao Dai, Xiaochi Wei, Hengyi Cai, Shuaiqiang Wang, Dawei Yin, Jun Xu, and Ji-Rong Wen. Tool learning with large language models: A survey.Frontiers of Computer Science, 19(8):198343, 2025

    work page 2025

  20. [20]

    Magentic-One: A Generalist Multi-Agent System for Solving Complex Tasks

    Adam Fourney, Gagan Bansal, Hussein Mozannar, Cheng Tan, Eduardo Salinas, Friederike Niedtner, Grace Proebsting, Griffin Bassman, Jack Gerrits, Jacob Alber, et al. Magentic-One: A generalist multi-agent system for solving complex tasks.arXiv preprint arXiv:2411.04468, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  21. [21]

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

    Wentao Zhang, Ce Cui, Yilei Zhao, Yang Liu, and Bo An. AgentOrchestra: A hierarchical multi-agent framework for general-purpose task solving.arXiv preprint arXiv:2506.12508, 2025

    work page arXiv 2025

  22. [22]

    Demystifying long chain-of-thought reasoning in LLMs.arXiv preprint arXiv:2502.20379,2025

    Shalev Lifshitz, Sheila A McIlraith, and Yilun Du. Multi-agent verification: Scaling test-time compute with multiple verifiers.arXiv preprint arXiv:2502.20379, 2025

    work page arXiv 2025

  23. [23]

    Continuum: Efficient and Robust Multi-Turn LLM Agent Scheduling with KV Cache Time-to-Live

    Hanchen Li, Runyuan He, Qiuyang Mang, Qizheng Zhang, Huanzhi Mao, Xiaokun Chen, Hangrui Zhou, Alvin Cheung, Joseph Gonzalez, and Ion Stoica. Continuum: Efficient and robust multi-turn LLM agent scheduling with KV cache time-to-live.arXiv preprint arXiv:2511.02230, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  24. [24]

    Sutradhara: An Intelligent Orchestrator-Engine Co-design for Tool-based Agentic Inference

    Anish Biswas, Kanishk Goel, Jayashree Mohan, Alind Khare, Anjaly Parayil, Ramachandran Ramjee, and Chetan Bansal. Sutradhara: An intelligent orchestrator-engine co-design for tool-based agentic inference.arXiv preprint arXiv:2601.12967, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  25. [25]

    Speculative actions: A lossless framework for faster AI agents

    Naimeng Ye, Arnav Ahuja, Georgios Liargkovas, Yunan Lu, Kostis Kaffes, and Tianyi Peng. Speculative actions: A lossless framework for faster AI agents. InThe Fourteenth International Conference on Learning Representations, 2026. URL https://openreview.net/forum? id=P0GOk5wslg

    work page 2026

  26. [26]

    Optimizing agentic language model inference via speculative tool calls.arXiv preprint arXiv:2512.15834, 2025

    Daniel Nichols, Prajwal Singhania, Charles Jekel, Abhinav Bhatele, and Harshitha Menon. Optimizing agentic language model inference via speculative tool calls.arXiv preprint arXiv:2512.15834, 2025

    work page arXiv 2025

  27. [27]

    Interactive speculative planning: Enhance agent efficiency through co-design of system and user interface

    Wenyue Hua, Mengting Wan, Jagannath Shashank Subramanya Sai Vadrevu, Ryan Nadel, Yongfeng Zhang, and Chi Wang. Interactive speculative planning: Enhance agent efficiency through co-design of system and user interface. InThe Thirteenth International Conference on Learning Representations, 2025. URL https://openreview.net/forum?id=BwR8t91yqh

    work page 2025

  28. [28]

    Analysis of thompson sampling for the multi-armed bandit problem

    Shipra Agrawal and Navin Goyal. Analysis of thompson sampling for the multi-armed bandit problem. In Shie Mannor, Nathan Srebro, and Robert C. Williamson, editors,Proceedings of the 25th Annual Conference on Learning Theory, volume 23 ofProceedings of Machine Learning Research, pages 39.1–39.26, Edinburgh, Scotland, 25–27 Jun 2012. PMLR. URL https://proce...

    work page 2012

  29. [29]

    Scaling test-time compute for LLM agents

    King Zhu, Hanhao Li, Siwei Wu, Tianshun Xing, Dehua Ma, Xiangru Tang, Minghao Liu, Jian Yang, Jiaheng Liu, Yuchen Eleanor Jiang, et al. Scaling test-time compute for LLM agents. arXiv preprint arXiv:2506.12928, 2025

    work page arXiv 2025

  30. [30]

    OAgents: An empirical study of building effective agents.arXiv preprint arXiv:2506.15741, 2025

    He Zhu, Tianrui Qin, King Zhu, Heyuan Huang, Yeyi Guan, Jinxiang Xia, Yi Yao, Hanhao Li, Ningning Wang, Pai Liu, et al. OAgents: An empirical study of building effective agents.arXiv preprint arXiv:2506.15741, 2025. 11

    work page arXiv 2025

  31. [31]

    The original prompt is designed for mathematical problem solving. Please minimally adapt it to better support {task}

    Xingyao Wang, Boxuan Li, Yufan Song, Frank F. Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H. Tran, Fuqiang Li, Ren Ma, Mingzhang Zheng, Bill Qian, Yanjun Shao, Niklas Muennighoff, Yizhe Zhang, Binyuan Hui, Junyang Lin, Robert Brennan, Hao Peng, Heng Ji, and Graham Neubig. OpenHands: An open platform for AI soft...

    work page 2025

  32. [32]

    years only

    Final code applies the template verbatim, prints integer ages 22 and 34, and submits final_answer(12). Why This Works.The injected plan turns the failure modes of the other two methods into pinned constraints in the executor’s context: the pre-verified birthdates close the door on Sleep-Time Compute’s hallucination, and the explicit “years only” instructi...

    work page 1987

  33. [33]

    Li Peng” as the unique match. 5.final_answer(

    Inline cross-reference between the two lists prints “Li Peng” as the unique match. 5.final_answer("Li Peng"). 20 Why This Works.The decisive contribution is the first idle window’s plan, which widens retrieval scope before the second action is chosen — once the executor’s step 2 query is “contributors list of 4.0.0” instead of “commit author of the Mask-R...

    work page