LLM-as-Code: Agentic Programming for Agent Harness

Hanyu Yan; Jingtong Gao; Junjia Qi; Wenlin Zhang; Xiangyu Zhao; Xian Wu; Zichuan Fu

arxiv: 2606.15874 · v2 · pith:ASS7EBGJnew · submitted 2026-06-14 · 💻 cs.AI · cs.SE

LLM-as-Code: Agentic Programming for Agent Harness

Junjia Qi , Zichuan Fu , Jingtong Gao , Wenlin Zhang , Hanyu Yan , Xian Wu , Xiangyu Zhao This is my paper

Pith reviewed 2026-06-27 04:13 UTC · model grok-4.3

classification 💻 cs.AI cs.SE

keywords agentic programmingLLM agentscontrol flowcomputer-use agentscall treeDAG contexttoken explosionLLM-as-Code

0 comments

The pith

Giving LLMs control over loops and branches produces token explosion and hallucinations because probabilistic models cannot reliably perform deterministic sequencing.

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

The paper argues that token explosion, control-flow hallucination, and unreliable completion in LLM agents are architectural results of letting the model decide execution paths rather than bugs fixable by better prompts. It introduces Agentic Programming, in which a conventional program owns all control flow and invokes the LLM only as an adaptive component called LLM-as-Code for tasks that need reasoning or generation. Within each invocation the model retains flexibility, yet it cannot change the program's execution path. Context for each call is drawn from the execution history's call tree, which forms a directed acyclic graph, so length is set by call depth instead of accumulating across steps. A computer-use agent case study shows the approach yields more stable performance on long visual operation sequences.

Core claim

Agentic Programming places all control flow under program governance and treats the LLM as LLM-as-Code, an adaptive component invoked solely where reasoning or generation is required. The execution history is recorded as a call tree that becomes a directed acyclic graph for context construction, ensuring each call's context length depends on its depth in the tree rather than on the total number of steps taken.

What carries the argument

LLM-as-Code: the LLM treated as an adaptive component inside a program that fully owns control flow and prevents the model from altering execution paths.

If this is right

Context length for each LLM call is bounded by call depth instead of growing with total steps.
The LLM can still use full flexibility inside each invocation without risking changes to the overall execution path.
Stability of long operation sequences improves because deterministic sequencing is removed from the probabilistic component.
The call-tree DAG structure replaces linear accumulation of history.

Where Pith is reading between the lines

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

The same separation could be tested in non-visual agent domains such as code editing or web navigation to check whether context management benefits generalize.
Program-level control might allow cheaper models to be substituted for expensive ones on reasoning steps without loss of overall reliability.
Explicit call-tree logging could enable new debugging tools that trace exactly which LLM invocations contributed to a final outcome.

Load-bearing premise

Token explosion, control-flow hallucination, and unreliable completion result from assigning deterministic looping, branching, and sequencing to a probabilistic LLM.

What would settle it

Run identical long visual operation sequences in a standard LLM-orchestrated agent versus an Agentic Programming version and compare the rates of hallucinated branches, premature stops, and token usage.

Figures

Figures reproduced from arXiv: 2606.15874 by Hanyu Yan, Jingtong Gao, Junjia Qi, Wenlin Zhang, Xiangyu Zhao, Xian Wu, Zichuan Fu.

**Figure 1.** Figure 1: A comparison of the two agent paradigms on a simple multi-step task. In the left panel (LLM-as-Orchestrator) the [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

**Figure 2.** Figure 2: Context tracks the call graph rather than a flat conversation log. A running call keeps its full ancestor chain, while a [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

read the original abstract

Every major LLM agent framework gives the LLM the role of orchestrator; the model decides what to do next, when to call tools, and when to stop. We argue that token explosion, control-flow hallucination, and unreliable completion are not implementation bugs but architectural consequences of assigning the deterministic work of looping, branching, and sequencing to a probabilistic system. A better prompt or a stronger model cannot guarantee the reliability of the LLM agent. We therefore propose Agentic Programming, in which the program governs all control flow, and the LLM is itself part of it, an adaptive component we call LLM-as-Code and invoke only where a task calls for reasoning or generation. Within each call the model keeps full flexibility, but it cannot alter the program's execution path. With control in the program, the LLM's context is built from the execution history's call tree and forms a directed acyclic graph (DAG). Each call's context length is then determined by its call depth rather than by accumulation over steps. A case study of computer-use agents shows that the design is practical, not just a theoretical stance, substantially improving the stability of long visual operation sequences.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note private letter to a colleague

The paper cleanly separates control flow into the host program and limits the LLM to reasoning steps, which addresses a real failure mode, but the case study provides no numbers to show the improvement.

read the letter

The core claim is that current agent frameworks fail on long tasks because they hand looping, branching, and termination to the LLM. The authors instead keep those in ordinary code and invoke the model only for the parts that need generation or adaptation. Context for each call is then drawn from the call tree as a DAG rather than a growing linear history, so length tracks depth instead of total steps.

This framing as Agentic Programming and LLM-as-Code is new enough to be useful. It gives a concrete alternative to the usual orchestrator pattern and explains why prompt tweaks or stronger models will not fix the underlying reliability issues. The computer-use case study is offered as evidence that the approach works in practice for visual sequences.

The main weakness is the lack of any reported metrics. The abstract mentions improved stability but supplies no success rates, token counts, error breakdowns, or baseline comparisons. Without those, it is difficult to judge whether the DAG construction actually delivers the claimed gains or whether other factors in the implementation mattered more. The logic holds up on its own terms, but the evidential support is thin.

Engineers who build long-horizon agents will find the pattern worth trying even if they adapt the details. It is a clear design stance rather than a finished system. I would send it to referees so they can check the implementation and any quantitative results in the full manuscript.

Referee Report

2 major / 1 minor

Summary. The manuscript claims that token explosion, control-flow hallucination, and unreliable completion in LLM agent frameworks are architectural issues arising from assigning deterministic control flow tasks to probabilistic LLMs. It proposes 'Agentic Programming' as an alternative where the program handles all control flow and invokes the LLM only as an 'LLM-as-Code' component for reasoning or generation tasks. In this setup, the LLM's context is constructed as a directed acyclic graph (DAG) from the execution history's call tree, with context length determined by call depth. A case study on computer-use agents is cited to show that this design improves stability of long visual operation sequences.

Significance. If the architectural separation holds and the case study generalizes, this approach could provide a more reliable foundation for building LLM agents by leveraging the strengths of deterministic programming for sequencing and probabilistic models for adaptive reasoning. The DAG-based context management directly mitigates context accumulation problems. The paper applies a standard engineering pattern to LLM agents, which is a strength in its conceptual clarity.

major comments (2)

[Abstract] Abstract: the claim that the design is 'substantially improving the stability of long visual operation sequences' is presented without quantitative metrics, baselines, error rates, success percentages, or implementation details. This leaves the central practical claim without verifiable support.
[Introduction / Argument] The core argument (that the listed failure modes are architectural consequences rather than implementation bugs) is logically consistent but load-bearing for the proposal; it would benefit from explicit discussion of why alternative mitigations (e.g., structured output constraints or external planners) cannot address the same issues within existing frameworks.

minor comments (1)

[Terminology] Terminology: the newly introduced terms 'LLM-as-Code' and 'Agentic Programming' would benefit from an explicit comparison table or paragraph relating them to existing patterns such as ReAct, tool-calling agents, or program-guided LLM use.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments and the recommendation for major revision. We address each point below and commit to revisions that strengthen the verifiability and argumentation of the manuscript.

read point-by-point responses

Referee: [Abstract] Abstract: the claim that the design is 'substantially improving the stability of long visual operation sequences' is presented without quantitative metrics, baselines, error rates, success percentages, or implementation details. This leaves the central practical claim without verifiable support.

Authors: We agree that the abstract presents the improvement claim without accompanying quantitative evidence. The case study demonstrates practicality through qualitative observation of long-horizon stability, but to make the central claim verifiable we will revise the abstract to report specific metrics from the computer-use experiments, including success rates, error reductions, and baseline comparisons. revision: yes
Referee: [Introduction / Argument] The core argument (that the listed failure modes are architectural consequences rather than implementation bugs) is logically consistent but load-bearing for the proposal; it would benefit from explicit discussion of why alternative mitigations (e.g., structured output constraints or external planners) cannot address the same issues within existing frameworks.

Authors: We accept that an explicit discussion of alternatives would strengthen the load-bearing argument. We will add a subsection to the introduction that directly compares Agentic Programming against structured-output constraints and external-planner approaches, explaining why those mitigations leave token explosion, control-flow hallucination, and unreliable completion unresolved when the LLM retains the orchestrator role. revision: yes

Circularity Check

0 steps flagged

No significant circularity

full rationale

The paper is a conceptual architectural argument with no equations, fitted parameters, derivations, or quantitative predictions. The central claim—that token explosion, control-flow hallucination, and unreliable completion are architectural consequences of assigning deterministic control flow to a probabilistic model—follows directly from the proposed separation of concerns (program handles sequencing; LLM invoked only for reasoning). The DAG context construction is a straightforward consequence of fixing control flow outside the model rather than a redefinition or fit. No self-citations, uniqueness theorems, or ansatzes appear in a load-bearing role. The proposal is a standard engineering pattern and remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 2 invented entities

The central claim rests on the domain assumption that control-flow tasks are inherently mismatched to LLMs and introduces two new conceptual entities without external falsifiable evidence.

axioms (1)

domain assumption LLMs are probabilistic systems unsuitable for deterministic control flow tasks such as looping, branching, and sequencing
Explicitly stated in the abstract as the root cause of observed failures.

invented entities (2)

LLM-as-Code no independent evidence
purpose: Adaptive component invoked by the program only for reasoning or generation tasks
New term and role introduced to describe selective LLM use within program-controlled flow.
Agentic Programming no independent evidence
purpose: Programming paradigm in which the program governs all control flow and the LLM is an invoked component
Core new architectural concept proposed as the solution.

pith-pipeline@v0.9.1-grok · 5743 in / 1432 out tokens · 63717 ms · 2026-06-27T04:13:10.570106+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

25 extracted references · 2 canonical work pages

[1]

Islem Bouzenia and Michael Pradel. 2025. Understanding Software Engineering Agents: A Study of Thought-Action-Result Trajectories. InProceedings of the 40th IEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE/ACM, New York, NY, USA, 12 pages. https://arxiv.org/abs/2506.18824

arXiv 2025
[2]

Harrison Chase and Nuno Campos. 2024. LangGraph: Building Stateful, Multi- Actor Applications with LLMs. https://github.com/langchain-ai/langgraph. Ac- cessed: 2026-05-30

2024
[3]

Tse-Hsun Chen. 2026. Towards Structured, State-Aware, and Execution-Grounded Reasoning for Software Engineering Agents. InProceedings of the 7th International Workshop on Bots and Agents in Software Engineering (BoatSE ’26). ACM, New York, NY, USA, 6 pages. doi:10.1145/3786161.3788456

work page doi:10.1145/3786161.3788456 2026
[4]

Gonzalo Gonzalez-Pumariega, Vincent Tu, Chih-Lun Lee, Jiachen Yang, Ang Li, and Xin Eric Wang. 2025. The Unreasonable Effectiveness of Scaling Agents for Computer Use. arXiv:2510.02250 [cs.AI] https://arxiv.org/abs/2510.02250

arXiv 2025
[5]

Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jinlin Wang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. 2024. MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework. In International Conference on Learning Representations (ICLR). O...

Pith/arXiv arXiv 2024
[6]

Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024. Large Language Models Cannot Self- Correct Reasoning Yet. InInternational Conference on Learning Representations (ICLR). OpenReview.net, Vienna, Austria, 14 pages. https://arxiv.org/abs/2310. 01798

2024
[7]

Yuxin Jiang, Yufei Wang, Xingshan Zeng, Wanjun Zhong, Liangyou Li, Fei Mi, Lifeng Shang, Xin Jiang, Qun Liu, and Wei Wang. 2024. FollowBench: A Multi-level Fine-grained Constraints Following Benchmark for Large Lan- guage Models. InAnnual Meeting of the Association for Computational Linguistics (ACL). Association for Computational Linguistics, Bangkok, Th...

arXiv 2024
[8]

Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan

Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. 2024. SWE-bench: Can Language Models Resolve Real- World GitHub Issues?. InInternational Conference on Learning Representations (ICLR). OpenReview.net, Vienna, Austria, 23 pages

2024
[9]

Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts

Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2024. DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines. InInternational Conference on Learning Representations (IC...

Pith/arXiv arXiv 2024
[10]

Mosh Levy, Alon Jacoby, and Yoav Goldberg. 2024. Same Task, More Tokens: the Impact of Input Length on the Reasoning Performance of Large Language Models. InAnnual Meeting of the Association for Computational Linguistics (ACL). Association for Computational Linguistics, Bangkok, Thailand, 15 pages. https: //arxiv.org/abs/2402.14848

arXiv 2024
[11]

Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the Middle: How Language Models Use Long Contexts.Transactions of the Association for Computational Linguistics (TACL)12 (2024), 157–173. https://aclanthology.org/2024.tacl-1.9/

2024
[12]

Norman Mu, Sarah Chen, Zifan Wang, Sizhe Chen, David Karamardian, Lulwa Aljeraisy, Basel Alomair, Dan Hendrycks, and David Wagner. 2023. Can LLMs Follow Simple Rules? arXiv:2311.04235 [cs.AI] https://arxiv.org/abs/2311.04235

arXiv 2023
[13]

Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language Models Can Teach Themselves to Use Tools. InAdvances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., Red Hook, NY, USA, 14 pages

2023
[14]

Chi, Nathanael Schärli, and Denny Zhou

Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H. Chi, Nathanael Schärli, and Denny Zhou. 2023. Large Language Models Can Be Easily Distracted by Irrelevant Context. InInternational Conference on Machine Learning (ICML). PMLR, Honolulu, HI, USA, 14 pages. https://arxiv.org/abs/2302.00093

arXiv 2023
[15]

Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language Agents with Verbal Reinforcement Learning. InAdvances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., Red Hook, NY, USA, 18 pages. https://arxiv. org/abs/2303.11366

Pith/arXiv arXiv 2023
[16]

Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, et al
[17]

doi: 10.1007/s11704-024-40231-1

A Survey on Large Language Model based Autonomous Agents.Frontiers of Computer Science18, 6 (2024), 186345. doi:10.1007/s11704-024-40231-1

work page doi:10.1007/s11704-024-40231-1 2024
[18]

Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. 2024. Executable Code Actions Elicit Better LLM Agents. InProceedings of the 41st International Conference on Machine Learning (ICML). PMLR, Vienna, Austria, 15 pages. https://arxiv.org/abs/2402.01030

arXiv 2024
[19]

Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H

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. 2025. OpenHands: An Open Platform for A...

Pith/arXiv arXiv 2025
[20]

White, Doug Burger, and Chi Wang

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W. White, Doug Burger, and Chi Wang. 2024. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. InConference on Language Modeling (COLM). OpenReview.net, Philadelphia, PA, USA, 41 pages

2024
[21]

Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, et al
[22]

https://arxiv.org/abs/2309.07864

The Rise and Potential of Large Language Model Based Agents: A Survey. https://arxiv.org/abs/2309.07864

Pith/arXiv arXiv
[23]

Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu
[24]

InAdvances in Neural Information Processing Systems (NeurIPS)

OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments. InAdvances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., Red Hook, NY, USA, 64 pages
[25]

next step

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. InInternational Conference on Learning Representations (ICLR). OpenRe- view.net, Kigali, Rwanda, 33 pages. A Why Standard Patches Miss The category error is not for lack of effort, since the communit...

2023

[1] [1]

Islem Bouzenia and Michael Pradel. 2025. Understanding Software Engineering Agents: A Study of Thought-Action-Result Trajectories. InProceedings of the 40th IEEE/ACM International Conference on Automated Software Engineering (ASE). IEEE/ACM, New York, NY, USA, 12 pages. https://arxiv.org/abs/2506.18824

arXiv 2025

[2] [2]

Harrison Chase and Nuno Campos. 2024. LangGraph: Building Stateful, Multi- Actor Applications with LLMs. https://github.com/langchain-ai/langgraph. Ac- cessed: 2026-05-30

2024

[3] [3]

Tse-Hsun Chen. 2026. Towards Structured, State-Aware, and Execution-Grounded Reasoning for Software Engineering Agents. InProceedings of the 7th International Workshop on Bots and Agents in Software Engineering (BoatSE ’26). ACM, New York, NY, USA, 6 pages. doi:10.1145/3786161.3788456

work page doi:10.1145/3786161.3788456 2026

[4] [4]

Gonzalo Gonzalez-Pumariega, Vincent Tu, Chih-Lun Lee, Jiachen Yang, Ang Li, and Xin Eric Wang. 2025. The Unreasonable Effectiveness of Scaling Agents for Computer Use. arXiv:2510.02250 [cs.AI] https://arxiv.org/abs/2510.02250

arXiv 2025

[5] [5]

Sirui Hong, Mingchen Zhuge, Jonathan Chen, Xiawu Zheng, Yuheng Cheng, Ceyao Zhang, Jinlin Wang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, Lingfeng Xiao, Chenglin Wu, and Jürgen Schmidhuber. 2024. MetaGPT: Meta Programming for a Multi-Agent Collaborative Framework. In International Conference on Learning Representations (ICLR). O...

Pith/arXiv arXiv 2024

[6] [6]

Jie Huang, Xinyun Chen, Swaroop Mishra, Huaixiu Steven Zheng, Adams Wei Yu, Xinying Song, and Denny Zhou. 2024. Large Language Models Cannot Self- Correct Reasoning Yet. InInternational Conference on Learning Representations (ICLR). OpenReview.net, Vienna, Austria, 14 pages. https://arxiv.org/abs/2310. 01798

2024

[7] [7]

Yuxin Jiang, Yufei Wang, Xingshan Zeng, Wanjun Zhong, Liangyou Li, Fei Mi, Lifeng Shang, Xin Jiang, Qun Liu, and Wei Wang. 2024. FollowBench: A Multi-level Fine-grained Constraints Following Benchmark for Large Lan- guage Models. InAnnual Meeting of the Association for Computational Linguistics (ACL). Association for Computational Linguistics, Bangkok, Th...

arXiv 2024

[8] [8]

Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan

Carlos E. Jimenez, John Yang, Alexander Wettig, Shunyu Yao, Kexin Pei, Ofir Press, and Karthik Narasimhan. 2024. SWE-bench: Can Language Models Resolve Real- World GitHub Issues?. InInternational Conference on Learning Representations (ICLR). OpenReview.net, Vienna, Austria, 23 pages

2024

[9] [9]

Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts

Omar Khattab, Arnav Singhvi, Paridhi Maheshwari, Zhiyuan Zhang, Keshav Santhanam, Sri Vardhamanan, Saiful Haq, Ashutosh Sharma, Thomas T. Joshi, Hanna Moazam, Heather Miller, Matei Zaharia, and Christopher Potts. 2024. DSPy: Compiling Declarative Language Model Calls into Self-Improving Pipelines. InInternational Conference on Learning Representations (IC...

Pith/arXiv arXiv 2024

[10] [10]

Mosh Levy, Alon Jacoby, and Yoav Goldberg. 2024. Same Task, More Tokens: the Impact of Input Length on the Reasoning Performance of Large Language Models. InAnnual Meeting of the Association for Computational Linguistics (ACL). Association for Computational Linguistics, Bangkok, Thailand, 15 pages. https: //arxiv.org/abs/2402.14848

arXiv 2024

[11] [11]

Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang

Nelson F. Liu, Kevin Lin, John Hewitt, Ashwin Paranjape, Michele Bevilacqua, Fabio Petroni, and Percy Liang. 2024. Lost in the Middle: How Language Models Use Long Contexts.Transactions of the Association for Computational Linguistics (TACL)12 (2024), 157–173. https://aclanthology.org/2024.tacl-1.9/

2024

[12] [12]

Norman Mu, Sarah Chen, Zifan Wang, Sizhe Chen, David Karamardian, Lulwa Aljeraisy, Basel Alomair, Dan Hendrycks, and David Wagner. 2023. Can LLMs Follow Simple Rules? arXiv:2311.04235 [cs.AI] https://arxiv.org/abs/2311.04235

arXiv 2023

[13] [13]

Timo Schick, Jane Dwivedi-Yu, Roberto Dessì, Roberta Raileanu, Maria Lomeli, Eric Hambro, Luke Zettlemoyer, Nicola Cancedda, and Thomas Scialom. 2023. Toolformer: Language Models Can Teach Themselves to Use Tools. InAdvances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., Red Hook, NY, USA, 14 pages

2023

[14] [14]

Chi, Nathanael Schärli, and Denny Zhou

Freda Shi, Xinyun Chen, Kanishka Misra, Nathan Scales, David Dohan, Ed H. Chi, Nathanael Schärli, and Denny Zhou. 2023. Large Language Models Can Be Easily Distracted by Irrelevant Context. InInternational Conference on Machine Learning (ICML). PMLR, Honolulu, HI, USA, 14 pages. https://arxiv.org/abs/2302.00093

arXiv 2023

[15] [15]

Noah Shinn, Federico Cassano, Edward Berman, Ashwin Gopinath, Karthik Narasimhan, and Shunyu Yao. 2023. Reflexion: Language Agents with Verbal Reinforcement Learning. InAdvances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., Red Hook, NY, USA, 18 pages. https://arxiv. org/abs/2303.11366

Pith/arXiv arXiv 2023

[16] [16]

Lei Wang, Chen Ma, Xueyang Feng, Zeyu Zhang, Hao Yang, Jingsen Zhang, et al

[17] [17]

doi: 10.1007/s11704-024-40231-1

A Survey on Large Language Model based Autonomous Agents.Frontiers of Computer Science18, 6 (2024), 186345. doi:10.1007/s11704-024-40231-1

work page doi:10.1007/s11704-024-40231-1 2024

[18] [18]

Xingyao Wang, Yangyi Chen, Lifan Yuan, Yizhe Zhang, Yunzhu Li, Hao Peng, and Heng Ji. 2024. Executable Code Actions Elicit Better LLM Agents. InProceedings of the 41st International Conference on Machine Learning (ICML). PMLR, Vienna, Austria, 15 pages. https://arxiv.org/abs/2402.01030

arXiv 2024

[19] [19]

Xu, Xiangru Tang, Mingchen Zhuge, Jiayi Pan, Yueqi Song, Bowen Li, Jaskirat Singh, Hoang H

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. 2025. OpenHands: An Open Platform for A...

Pith/arXiv arXiv 2025

[20] [20]

White, Doug Burger, and Chi Wang

Qingyun Wu, Gagan Bansal, Jieyu Zhang, Yiran Wu, Beibin Li, Erkang Zhu, Li Jiang, Xiaoyun Zhang, Shaokun Zhang, Jiale Liu, Ahmed Hassan Awadallah, Ryen W. White, Doug Burger, and Chi Wang. 2024. AutoGen: Enabling Next-Gen LLM Applications via Multi-Agent Conversation. InConference on Language Modeling (COLM). OpenReview.net, Philadelphia, PA, USA, 41 pages

2024

[21] [21]

Zhiheng Xi, Wenxiang Chen, Xin Guo, Wei He, Yiwen Ding, Boyang Hong, et al

[22] [22]

https://arxiv.org/abs/2309.07864

The Rise and Potential of Large Language Model Based Agents: A Survey. https://arxiv.org/abs/2309.07864

Pith/arXiv arXiv

[23] [23]

Tianbao Xie, Danyang Zhang, Jixuan Chen, Xiaochuan Li, Siheng Zhao, Ruisheng Cao, Toh Jing Hua, Zhoujun Cheng, Dongchan Shin, Fangyu Lei, Yitao Liu, Yiheng Xu, Shuyan Zhou, Silvio Savarese, Caiming Xiong, Victor Zhong, and Tao Yu

[24] [24]

InAdvances in Neural Information Processing Systems (NeurIPS)

OSWorld: Benchmarking Multimodal Agents for Open-Ended Tasks in Real Computer Environments. InAdvances in Neural Information Processing Systems (NeurIPS). Curran Associates, Inc., Red Hook, NY, USA, 64 pages

[25] [25]

next step

Shunyu Yao, Jeffrey Zhao, Dian Yu, Nan Du, Izhak Shafran, Karthik Narasimhan, and Yuan Cao. 2023. ReAct: Synergizing Reasoning and Acting in Language Models. InInternational Conference on Learning Representations (ICLR). OpenRe- view.net, Kigali, Rwanda, 33 pages. A Why Standard Patches Miss The category error is not for lack of effort, since the communit...

2023