---
title: "Identifying the Risks of LM Agents with an LM-Emulated Sandbox"
description: "The authors present ToolEmu, a framework where a strong LM (e.g., GPT-4) imitates tool behaviour… 50 references, cited by 89 papers here."
authors: ["Yangjun Ruan", "Honghua Dong", "Andrew Wang", "Silviu Pitis", "Yongchao Zhou", "Jimmy Ba", "Yann Dubois", "Chris J. Maddison", "Tatsunori Hashimoto"]
year: 2023
venue: "arXiv:2309.15817"
source: "https://arxiv.org/abs/2309.15817"
doi: "10.48550/arxiv.2309.15817"
arxiv: "2309.15817"
licence: "CC BY 4.0 (arXiv)"
licence_url: "http://creativecommons.org/licenses/by/4.0/"
summary_licence: "Elamur Cosmic License"
canonical: "https://research.elamur.ai/free/en/papers/ruan-2023-identifying-risks-lm-agents-lm/"
---
[Elamur Research](https://md.research.elamur.ai/free/en/) · [this page in HTML](https://research.elamur.ai/free/en/papers/ruan-2023-identifying-risks-lm-agents-lm/)

[home](https://md.research.elamur.ai/free/en/) / [papers](https://md.research.elamur.ai/free/en/papers/) / [2023](https://md.research.elamur.ai/free/en/papers/2023/page/8/) / Identifying the Risks of LM Agents with an LM-Emulated Sandbox

# Identifying the Risks of LM Agents with an LM-Emulated Sandbox

**Authors:** [Yangjun Ruan](https://md.research.elamur.ai/free/en/authors/yangjun-ruan/), [Honghua Dong](https://md.research.elamur.ai/free/en/authors/honghua-dong/), [Andrew Wang](https://md.research.elamur.ai/free/en/authors/andrew-wang/), [Silviu Pitis](https://md.research.elamur.ai/free/en/authors/silviu-pitis/), [Yongchao Zhou](https://md.research.elamur.ai/free/en/authors/yongchao-zhou/), [Jimmy Ba](https://md.research.elamur.ai/free/en/authors/jimmy-lei-ba/), [Yann Dubois](https://md.research.elamur.ai/free/en/authors/yann-dubois/), [Chris J. Maddison](https://md.research.elamur.ai/free/en/authors/chris-j-maddison/) and [Tatsunori Hashimoto](https://md.research.elamur.ai/free/en/authors/tatsunori-hashimoto/) · **Year:** 2023 · **Venue:** arXiv:2309.15817 · **Licence:** [CC BY 4.0 (arXiv)](http://creativecommons.org/licenses/by/4.0/)

**Links:** [Publisher's page](https://arxiv.org/abs/2309.15817) · [PDF](https://arxiv.org/pdf/2309.15817) · [DOI 10.48550/arxiv.2309.15817](https://doi.org/10.48550/arxiv.2309.15817) · [arXiv 2309.15817](https://arxiv.org/abs/2309.15817) · [OpenAlex](https://openalex.org/W4387164935)

**Citations:** 7 citations (OpenAlex, 2026-10-14) · [89](https://md.research.elamur.ai/free/en/papers/ruan-2023-identifying-risks-lm-agents-lm/cited-by/) in our library, of the 149,439 papers whose bibliographies we read

## Summary

LM agents that call external tools, such as banking APIs, terminals or smart-home devices, can cause serious harm when they misread a request. Finding these rare but costly failures normally means building real tools and test environments by hand, which is slow and limits how many scenarios can be checked. The authors present ToolEmu, a framework where a strong LM (e.g., GPT-4) imitates tool behaviour using only each tool's specification and inputs. An adversarial variant deliberately constructs sandbox states that make mistakes more likely, which supports red-teaming. LM-based evaluators then score each agent trajectory for safety and helpfulness. Human checks show that most emulated runs are realistic and that the evaluators agree with annotators about as well as annotators agree with each other. On a curated benchmark built around underspecified or ambiguous user instructions, even the best-performing agents still produce risky actions. This makes the framework a cheap, scalable way to stress-test agents before deployment.

## Key points

- **Benchmark size:** 144 test cases spanning 36 toolkits in 18 categories and 9 risk types. Most of these tools lack existing sandboxed evaluations.
- **Validation of the emulator and evaluator:**
  - Over 80% of 200 emulated trajectories were rated realistic by humans.
  - 68.8% of the flagged failures were confirmed as genuine, realistic risks.
- **Agreement with humans:** the automatic safety evaluator detects 73.1% of failures found by a 3-annotator majority vote. A single held-out human annotator detects 78.8% on average.
- **Real-world transfer:** of 7 severe ChatGPT-3.5 failures on the emulated terminal, 6 were reproduced on a real bash shell. Reproducing them took about 8 hours manually, versus under 15 minutes in ToolEmu.
- **Agent results:** GPT-4 and Claude-2 scored best on both safety and helpfulness, and prompt tuning improved results further. Still, the safest agent failed in 23.9% of test cases.

*Our summary of the paper, in our own words — not its text. The work belongs to its authors; read it at the links above.*

## Institutions

- [Stanford University](https://md.research.elamur.ai/free/en/universities/stanford-university/) — University, Stanford, United States · 506 works · 330,053 citations
- [University of Toronto](https://md.research.elamur.ai/free/en/universities/university-of-toronto/) — University, Toronto, Canada · 110 works · 179,062 citations
- [Vector Institute](https://md.research.elamur.ai/free/en/labs/vector-institute/) — Research institute, Toronto, Canada · 37 works · 5,703 citations

📚 Shelves (5): [🖥️ Computer Science](https://md.research.elamur.ai/free/en/shelves/computer-science/) · 18,614 works · [🤖 Artificial Intelligence](https://md.research.elamur.ai/free/en/shelves/computer-science/artificial-intelligence/) · 17,965 works · [🤝 Social Sciences](https://md.research.elamur.ai/free/en/shelves/social-sciences/) · 1,954 works · [🤝 Safety Research](https://md.research.elamur.ai/free/en/shelves/social-sciences/safety-research/) · 626 works · [💻 Software](https://md.research.elamur.ai/free/en/shelves/computer-science/software/) · 119 works

## Cite

Cite as: Yangjun Ruan, Honghua Dong, Andrew Wang, Silviu Pitis, Yongchao Zhou, Jimmy Ba, Yann Dubois, Chris J. Maddison and Tatsunori Hashimoto. 2023. Identifying the Risks of LM Agents with an LM-Emulated Sandbox. arXiv:2309.15817.

The paper itself, not this page. BibTeX from the library's record of the work.

```bibtex
@misc{ruan2023identifying,
  title = {{Identifying the Risks of LM Agents with an LM-Emulated Sandbox}},
  author = {Yangjun Ruan, Honghua Dong, Andrew Wang, Silviu Pitis, Yongchao Zhou, Jimmy Ba, Yann Dubois, Chris J. Maddison and Tatsunori Hashimoto},
  year = {2023},
  doi = {10.48550/arxiv.2309.15817},
  eprint = {2309.15817},
  archivePrefix = {arXiv},
  url = {https://doi.org/10.48550/arxiv.2309.15817}
}
```

- [← previous Ghost in the Minecraft: Generally Capable Agents for Open-World Environments via Large Language Models with Text-based Knowledge and Memory](https://md.research.elamur.ai/free/en/papers/zhu-2023-ghost-minecraft-generally-capable-agents/)
- [next → INSTRUCTEVAL: Towards Holistic Evaluation of Instruction-Tuned Large Language Models](https://md.research.elamur.ai/free/en/papers/chia-2023-instructeval-towards-holistic-evaluation-instruction/)
- [↑ back to papers of 2023](https://md.research.elamur.ai/free/en/papers/2023/page/8/)

## References

50 references, 22 of them with a card of ours. ↗ leads to the work, «find» searches for one we have no id for yet.

1. Language model agents Building intelligent agents is a long-standing goal in artificial intelligence (Wooldridge and Jennings, 1995; Maes, 1995; Russell, 2010; Bostrom, 2014). Recent advances in LMs (Brown et al., 2020; Raffel et al., 2020; Ouyang et al., 2022; OpenAI, 2023a) have paved the way for… [find](https://scholar.google.com/scholar?q=Language+model+agents+Building+intelligent+agents+is+a+long%2Dstanding+goal+in+artificial+intelligence+%28Wooldridge+and+Jennings%2C+1995%3B+Maes%2C+1995%3B+Russell%2C+2010%3B+Bostrom%2C+2014%29%2E+Recent+advances+in+LMs+%28)
2. Evaluation of LM agents Existing benchmarks for evaluating LM agents primarily focus on assessing their capabilities. Specialized benchmarks have been established to evaluate domain-specific LM agents in code execution (Yang et al., 2023) and web (Yao et al., 2022; Deng et al., 2023; Zhou et al., 20… [find](https://scholar.google.com/scholar?q=Evaluation+of+LM+agents+Existing+benchmarks+for+evaluating+LM+agents+primarily+focus+on+assessing+their+capabilities%2E+Specialized+benchmarks+have+been+established+to+evaluate+domain%2Dspecific+LM+agents)
3. Notably, Kinniment et al. (2023) sought to assess the long-term risks of LM agents but from a capability perspective. Our evaluation framework distinguishes itself for several reasons: (i) To the best of our knowledge, our work is the first initiative in directly assessing the risks associated with… [find](https://scholar.google.com/scholar?q=Notably%2C+Kinniment+et+al%2E+%282023%29+sought+to+assess+the+long%2Dterm+risks+of+LM+agents+but+from+a+capability+perspective%2E+Our+evaluation+framework+distinguishes+itself+for+several+reasons%3A+%28i%29+To+the+best)
4. Language model as an emulator Prior work has demonstrated the strong emulation capabilities of current LMs. These LMs have been applied to emulate human behaviours (Park et al., 2022; 2023; Liu et al., 2023b; Gao et al., 2023a) or feedback (Bai et al., 2022b; Dubois et al., 2023; Zheng et al., 2023)… [find](https://scholar.google.com/scholar?q=Language+model+as+an+emulator+Prior+work+has+demonstrated+the+strong+emulation+capabilities+of+current+LMs%2E+These+LMs+have+been+applied+to+emulate+human+behaviours+%28Park+et+al%2E%2C+2022%3B+2023%3B+Liu+et+al%2E)
5. Simulation environments Due to the complexity of real-world environments and potential safety concerns, many experiments for developing and evaluating agents, especially in domains of autonomous driving and robotics, are performed in simulation environments. Notable examples include OpenAI Gym (Broc… [find](https://scholar.google.com/scholar?q=Simulation+environments+Due+to+the+complexity+of+real%2Dworld+environments+and+potential+safety+concerns%2C+many+experiments+for+developing+and+evaluating+agents%2C+especially+in+domains+of+autonomous+drivi)
6. L IMITATIONS & F UTURE D IRECTIONS Quality of the emulators & evaluators Both our emulators and evaluators are based on prompt-engineered LMs. We carefully designed their prompts to inject the requirements that they should satisfy to ensure the emulation quality (Appx. A.1) and evaluation accuracy (… [find](https://scholar.google.com/scholar?q=L+IMITATIONS+%26+F+UTURE+D+IRECTIONS+Quality+of+the+emulators+%26+evaluators+Both+our+emulators+and+evaluators+are+based+on+prompt%2Dengineered+LMs%2E+We+carefully+designed+their+prompts+to+inject+the+require)
7. R EFERENCES Michael Ahn, Anthony Brohan, Noah Brown, Yevgen Chebotar, Omar Cortes, Byron David, Chelsea Finn, Chuyuan Fu, Keerthana Gopalakrishnan, Karol Hausman, et al. [Do as i can, not as i say: Grounding language in robotic affordances](https://md.research.elamur.ai/free/en/papers/ahn-2022-i-i-say-grounding-language/). arXiv preprint arXiv:2204.01691, 2022.
8. Anthropic. Claude 2, July 2023. URL [https://www.anthropic.com/index/claude-2](https://www.anthropic.com/index/claude-2). Accessed: 2023-08-31. [↗](https://www.anthropic.com/index/claude-2)
9. Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback. arXiv preprint arXiv:2204.05862, 2022a. [find](https://scholar.google.com/scholar?q=Yuntao+Bai%2C+Andy+Jones%2C+Kamal+Ndousse%2C+Amanda+Askell%2C+Anna+Chen%2C+Nova+DasSarma%2C+Dawn+Drain%2C+Stanislav+Fort%2C+Deep+Ganguli%2C+Tom+Henighan%2C+et+al%2E+Training+a+helpful+and+harmless+assistant+with+reinforcem)
10. Yuntao Bai, Saurav Kadavath, Sandipan Kundu, Amanda Askell, Jackson Kernion, Andy Jones, Anna Chen, Anna Goldie, Azalia Mirhoseini, Cameron McKinnon, et al. [Constitutional ai: Harmlessness from ai feedback](https://md.research.elamur.ai/free/en/papers/bai-2022-constitutional-ai-harmlessness-ai-feedback/). arXiv preprint arXiv:2212.08073, 2022b.
11. Charles Beattie, Joel Z Leibo, Denis Teplyashin, Tom Ward, Marcus Wainwright, Heinrich Küttler, Andrew Lefrancq, Simon Green, Víctor Valdés, Amir Sadik, et al. [Deepmind lab](https://md.research.elamur.ai/free/en/papers/beattie-2016-deepmind-lab/). arXiv preprint arXiv:1612.03801, 2016.
12. Raja Ben Abdessalem, Shiva Nejati, Lionel C Briand, and Thomas Stifter. Testing advanced driver assistance systems using multi-objective search and neural networks. In Proceedings of the 31st IEEE/ACM international conference on automated software engineering, pages 63–74, 2016. [find](https://scholar.google.com/scholar?q=Raja+Ben+Abdessalem%2C+Shiva+Nejati%2C+Lionel+C+Briand%2C+and+Thomas+Stifter%2E+Testing+advanced+driver+assistance+systems+using+multi%2Dobjective+search+and+neural+networks%2E+In+Proceedings+of+the+31st+IEEE%2FACM)
13. Nick Bostrom. Superintelligence: Paths, Dangers, Strategies. Oxford University Press, 2014. [find](https://scholar.google.com/scholar?q=Nick+Bostrom%2E+Superintelligence%3A+Paths%2C+Dangers%2C+Strategies%2E+Oxford+University+Press%2C+2014%2E)
14. Greg Brockman, Vicki Cheung, Ludwig Pettersson, Jonas Schneider, John Schulman, Jie Tang, and Wojciech Zaremba. Openai gym. arXiv preprint arXiv:1606.01540, 2016. [find](https://scholar.google.com/scholar?q=Greg+Brockman%2C+Vicki+Cheung%2C+Ludwig+Pettersson%2C+Jonas+Schneider%2C+John+Schulman%2C+Jie+Tang%2C+and+Wojciech+Zaremba%2E+Openai+gym%2E+arXiv+preprint+arXiv%3A1606%2E01540%2C+2016%2E)
15. Anthony Brohan, Noah Brown, Justice Carbajal, Yevgen Chebotar, Xi Chen, Krzysztof Choromanski, Tianli Ding, Danny Driess, Avinava Dubey, Chelsea Finn, et al. Rt-2: Vision-language-action models transfer web knowledge to robotic control. arXiv preprint arXiv:2307.15818, 2023. [find](https://scholar.google.com/scholar?q=Anthony+Brohan%2C+Noah+Brown%2C+Justice+Carbajal%2C+Yevgen+Chebotar%2C+Xi+Chen%2C+Krzysztof+Choromanski%2C+Tianli+Ding%2C+Danny+Driess%2C+Avinava+Dubey%2C+Chelsea+Finn%2C+et+al%2E+Rt%2D2%3A+Vision%2Dlanguage%2Daction+models+transf)
16. Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. [Language models are few-shot learners](https://md.research.elamur.ai/free/en/papers/brown-2020-language-models-few-shot-learners/). Advances in neural information processing systems, 33:1877–1901, 2020.
17. Yevgen Chebotar, Ankur Handa, Viktor Makoviychuk, Miles Macklin, Jan Issac, Nathan Ratliff, and Dieter Fox. Closing the sim-to-real loop: Adapting simulation randomization with real world experience. In 2019 International Conference on Robotics and Automation (ICRA), pages 8973–8979. [find](https://scholar.google.com/scholar?q=Yevgen+Chebotar%2C+Ankur+Handa%2C+Viktor+Makoviychuk%2C+Miles+Macklin%2C+Jan+Issac%2C+Nathan+Ratliff%2C+and+Dieter+Fox%2E+Closing+the+sim%2Dto%2Dreal+loop%3A+Adapting+simulation+randomization+with+real+world+experience%2E+)
18. Mark Chen, Jerry Tworek, Heewoo Jun, Qiming Yuan, Henrique Ponde de Oliveira Pinto, Jared Kaplan, Harri Edwards, Yuri Burda, Nicholas Joseph, Greg Brockman, et al. [Evaluating large language models trained on code](https://md.research.elamur.ai/free/en/papers/chen-2021-evaluating-large-language-models-trained/). arXiv preprint arXiv:2107.03374, 2021.
19. Wei-Lin Chiang, Zhuohan Li, Zi Lin, Ying Sheng, Zhanghao Wu, Hao Zhang, Lianmin Zheng, Siyuan Zhuang, Yonghao Zhuang, Joseph E. Gonzalez, Ion Stoica, and Eric P. Xing. Vicuna: An open-source chatbot impressing gpt-4 with 90%* chatgpt quality, March 2023. URL [https://lmsys.org/blog/2023-03-30-vicuna…](https://lmsys.org/blog/2023-03-30-vicuna…) [↗](https://lmsys.org/blog/2023-03-30-vicuna…)
20. Ishita Dasgupta, Christine Kaeser-Chen, Kenneth Marino, Arun Ahuja, Sheila Babayan, Felix Hill, and Rob Fergus. Collaborating with language models for embodied reasoning. arXiv preprint arXiv:2302.00763, 2023. [find](https://scholar.google.com/scholar?q=Ishita+Dasgupta%2C+Christine+Kaeser%2DChen%2C+Kenneth+Marino%2C+Arun+Ahuja%2C+Sheila+Babayan%2C+Felix+Hill%2C+and+Rob+Fergus%2E+Collaborating+with+language+models+for+embodied+reasoning%2E+arXiv+preprint+arXiv%3A2302%2E007)
21. Jonas Degrave. Building a virtual machine inside chatgpt, 12 2022. URL [https://www.engraved.blog/building-a-virtual-machine-inside/](https://www.engraved.blog/building-a-virtual-machine-inside/). Accessed: 2023-08-28. [↗](https://www.engraved.blog/building-a-virtual-machine-inside/)
22. Xiang Deng, Yu Gu, Boyuan Zheng, Shijie Chen, Samuel Stevens, Boshi Wang, Huan Sun, and Yu Su. [Mind2web: Towards a generalist agent for the web](https://md.research.elamur.ai/free/en/papers/deng-2023-mind2web-towards-generalist-agent-web/). arXiv preprint arXiv:2306.06070, 2023.
23. Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1–16. PMLR, 2017. [find](https://scholar.google.com/scholar?q=Alexey+Dosovitskiy%2C+German+Ros%2C+Felipe+Codevilla%2C+Antonio+Lopez%2C+and+Vladlen+Koltun%2E+Carla%3A+An+open+urban+driving+simulator%2E+In+Conference+on+robot+learning%2C+pages+1%E2%80%9316%2E+PMLR%2C+2017%2E)
24. Yann Dubois, Xuechen Li, Rohan Taori, Tianyi Zhang, Ishaan Gulrajani, Jimmy Ba, Carlos Guestrin, Percy Liang, and Tatsunori B Hashimoto. [Alpacafarm: A simulation framework for methods that learn from human feedback](https://md.research.elamur.ai/free/en/papers/dubois-2023-alpacafarm-simulation-framework-methods-that/). arXiv preprint arXiv:2305.14387, 2023.
25. Kevin Ellis, Maxwell Nye, Yewen Pu, Felix Sosa, Josh Tenenbaum, and Armando Solar-Lezama. Write, execute, assess: Program synthesis with a repl. Advances in Neural Information Processing Systems, 32, 2019. [find](https://scholar.google.com/scholar?q=Kevin+Ellis%2C+Maxwell+Nye%2C+Yewen+Pu%2C+Felix+Sosa%2C+Josh+Tenenbaum%2C+and+Armando+Solar%2DLezama%2E+Write%2C+execute%2C+assess%3A+Program+synthesis+with+a+repl%2E+Advances+in+Neural+Information+Processing+Systems%2C+32%2C+)
26. Gordon Fraser and Andrea Arcuri. Evosuite: automatic test suite generation for object-oriented software. In Proceedings of the 19th ACM SIGSOFT symposium and the 13th European conference on Foundations of software engineering, pages 416–419, 2011. [find](https://scholar.google.com/scholar?q=Gordon+Fraser+and+Andrea+Arcuri%2E+Evosuite%3A+automatic+test+suite+generation+for+object%2Doriented+software%2E+In+Proceedings+of+the+19th+ACM+SIGSOFT+symposium+and+the+13th+European+conference+on+Foundation)
27. Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer, Kamal Ndousse, et al. [Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned](https://md.research.elamur.ai/free/en/papers/ganguli-2022-red-teaming-language-models-reduce/). arXiv preprint arXiv:2209.07858, 2022a.
28. Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer, Kamal Ndousse, et al. [Red teaming language models to reduce harms: Methods, scaling behaviors, and lessons learned](https://md.research.elamur.ai/free/en/papers/ganguli-2022-red-teaming-language-models-reduce/). arXiv preprint arXiv:2209.07858, 2022b.
29. Chen Gao, Xiaochong Lan, Zhihong Lu, Jinzhu Mao, Jinghua Piao, Huandong Wang, Depeng Jin, and Yong Li. S 3 : Social-network simulation system with large language model-empowered agents. arXiv preprint arXiv:2307.14984, 2023a. — [arXiv:2307.14984](https://arxiv.org/abs/2307.14984) [find](https://scholar.google.com/scholar?q=Chen+Gao%2C+Xiaochong+Lan%2C+Zhihong+Lu%2C+Jinzhu+Mao%2C+Jinghua+Piao%2C+Huandong+Wang%2C+Depeng+Jin%2C+and+Yong+Li%2E+S+3+%3A+Social%2Dnetwork+simulation+system+with+large+language+model%2Dempowered+agents%2E+arXiv+preprint)
30. Luyu Gao, Aman Madaan, Shuyan Zhou, Uri Alon, Pengfei Liu, Yiming Yang, Jamie Callan, and Graham Neubig. Pal: Program-aided language models. In International Conference on Machine Learning, pages 10764–10799. PMLR, 2023b. [find](https://scholar.google.com/scholar?q=Luyu+Gao%2C+Aman+Madaan%2C+Shuyan+Zhou%2C+Uri+Alon%2C+Pengfei+Liu%2C+Yiming+Yang%2C+Jamie+Callan%2C+and+Graham+Neubig%2E+Pal%3A+Program%2Daided+language+models%2E+In+International+Conference+on+Machine+Learning%2C+pages+1076)
31. Patrice Godefroid, Nils Klarlund, and Koushik Sen. Dart: Directed automated random testing. In Proceedings of the 2005 ACM SIGPLAN conference on Programming language design and implementation, pages 213–223, 2005. [find](https://scholar.google.com/scholar?q=Patrice+Godefroid%2C+Nils+Klarlund%2C+and+Koushik+Sen%2E+Dart%3A+Directed+automated+random+testing%2E+In+Proceedings+of+the+2005+ACM+SIGPLAN+conference+on+Programming+language+design+and+implementation%2C+pages+2)
32. Izzeddin Gur, Hiroki Furuta, Austin Huang, Mustafa Safdari, Yutaka Matsuo, Douglas Eck, and Aleksandra Faust. [A real-world webagent with planning, long context understanding, and program synthesis](https://md.research.elamur.ai/free/en/papers/gur-2023-real-world-webagent-planning-long/). arXiv preprint arXiv:2307.12856, 2023.
33. Sirui Hong, Xiawu Zheng, Jonathan Chen, Yuheng Cheng, Ceyao Zhang, Zili Wang, Steven Ka Shing Yau, Zijuan Lin, Liyang Zhou, Chenyu Ran, et al. Metagpt: Meta programming for multi-agent collaborative framework. arXiv preprint arXiv:2308.[00352, 2023](https://md.research.elamur.ai/free/en/papers/hong-2023-metagpt-meta-programming-multi/).
34. Wenlong Huang, Pieter Abbeel, Deepak Pathak, and Igor Mordatch. [Language models as zero-shot planners: Extracting actionable knowledge for embodied agents](https://md.research.elamur.ai/free/en/papers/huang-2022-language-models-zero-shot-planners/). In International Conference on Machine Learning, pages 9118–9147. PMLR, 2022.
35. Jared Kaplan, Sam McCandlish, Tom Henighan, Tom B Brown, Benjamin Chess, Rewon Child, Scott Gray, Alec Radford, Jeffrey Wu, and Dario Amodei. [Scaling laws for neural language models](https://md.research.elamur.ai/free/en/papers/kaplan-2020-scaling-laws-neural-language-models/). arXiv preprint arXiv:2001.08361, 2020.
36. Megan Kinniment, Lucas Jun, Koba Sato, Haoxing Du, Brian Goodrich, Max Hasin, Lawrence Chan, Luke Harold, Miles R. Lin Tao, Hjalmar Wijk, Joel Burget, Aaron Ho, Elizabeth Barnes, and Paul Christiano. Evaluating language-model agents on realistic autonomous tasks. 2023. URL [https://evals.alignment.or…](https://evals.alignment.or…) [↗](https://evals.alignment.or…)
37. Takeshi Kojima, Shixiang Shane Gu, Machel Reid, Yutaka Matsuo, and Yusuke Iwasawa. [Large language models are zero-shot reasoners](https://md.research.elamur.ai/free/en/papers/kojima-2022-large-language-models-zero-shot/). Advances in neural information processing systems, 35: 22199–22213, 2022.
38. Hung Le, Yue Wang, Akhilesh Deepak Gotmare, Silvio Savarese, and Steven Chu Hong Hoi. [Coderl: Mastering code generation through pretrained models and deep reinforcement learning](https://md.research.elamur.ai/free/en/papers/le-2022-coderl-mastering-code-generation-through/). Advances in Neural Information Processing Systems, 35:21314–21328, 2022.
39. Guohao Li, Hasan Abed Al Kader Hammoud, Hani Itani, Dmitrii Khizbullin, and Bernard Ghanem. Camel: Communicative agents for" mind" exploration of large scale language model society. arXiv preprint arXiv:2303.[17760, 2023a](https://md.research.elamur.ai/free/en/papers/li-2023-camel-communicative-agents-mind/).
40. Minghao Li, Feifan Song, Bowen Yu, Haiyang Yu, Zhoujun Li, Fei Huang, and Yongbin Li. Api-bank: A benchmark for tool-augmented llms. arXiv preprint arXiv:2304.08244, 2023b. — [arXiv:2304.08244](https://arxiv.org/abs/2304.08244) [find](https://scholar.google.com/scholar?q=Minghao+Li%2C+Feifan+Song%2C+Bowen+Yu%2C+Haiyang+Yu%2C+Zhoujun+Li%2C+Fei+Huang%2C+and+Yongbin+Li%2E+Api%2Dbank%3A+A+benchmark+for+tool%2Daugmented+llms%2E+arXiv+preprint+arXiv%3A2304%2E08244%2C+2023b%2E+%E2%80%94+%5BarXiv%3A2304%2E08244%5D%28https%3A)
41. Yujia Li, David Choi, Junyoung Chung, Nate Kushman, Julian Schrittwieser, Rémi Leblond, Tom Eccles, James Keeling, Felix Gimeno, Agustin Dal Lago, et al. [Competition-level code generation with alphacode](https://md.research.elamur.ai/free/en/papers/li-2022-competition-level-code-generation-alphacode/). Science, 378(6624):1092–1097, 2022.
42. Jacky Liang, Wenlong Huang, Fei Xia, Peng Xu, Karol Hausman, Brian Ichter, Pete Florence, and Andy Zeng. Code as policies: Language model programs for embodied control. In 2023 IEEE International Conference on Robotics and Automation (ICRA), pages 9493–9500. [find](https://scholar.google.com/scholar?q=Jacky+Liang%2C+Wenlong+Huang%2C+Fei+Xia%2C+Peng+Xu%2C+Karol+Hausman%2C+Brian+Ichter%2C+Pete+Florence%2C+and+Andy+Zeng%2E+Code+as+policies%3A+Language+model+programs+for+embodied+control%2E+In+2023+IEEE+International+Conf)
43. Bo Liu, Yuqian Jiang, Xiaohan Zhang, Qiang Liu, Shiqi Zhang, Joydeep Biswas, and Peter Stone. Llm+ p: Empowering large language models with optimal planning proficiency. arXiv preprint arXiv:2304.[11477, 2023a](https://md.research.elamur.ai/free/en/papers/liu-2023-llm-p-empowering-large-language/).
44. Ruibo Liu, Jason Wei, Shixiang Shane Gu, Te-Yen Wu, Soroush Vosoughi, Claire Cui, Denny Zhou, and Andrew M Dai. Mind’s eye: Grounded language model reasoning through simulation. arXiv preprint arXiv:2210.05359, 2022. [find](https://scholar.google.com/scholar?q=Ruibo+Liu%2C+Jason+Wei%2C+Shixiang+Shane+Gu%2C+Te%2DYen+Wu%2C+Soroush+Vosoughi%2C+Claire+Cui%2C+Denny+Zhou%2C+and+Andrew+M+Dai%2E+Mind%E2%80%99s+eye%3A+Grounded+language+model+reasoning+through+simulation%2E+arXiv+preprint+arXiv%3A2)
45. Ruibo Liu, Ruixin Yang, Chenyan Jia, Ge Zhang, Denny Zhou, Andrew M Dai, Diyi Yang, and Soroush Vosoughi. Training socially aligned language models in simulated human society. arXiv preprint arXiv:2305.[16960, 2023b](https://md.research.elamur.ai/free/en/papers/liu-2023-training-socially-aligned-language-models/).
46. Xiao Liu, Hao Yu, Hanchen Zhang, Yifan Xu, Xuanyu Lei, Hanyu Lai, Yu Gu, Hangliang Ding, Kaiwen Men, Kejuan Yang, et al. [Agentbench: Evaluating llms as agents](https://md.research.elamur.ai/free/en/papers/liu-2023-agentbench-evaluating-llms-agents/). arXiv preprint arXiv:2308.03688, 2023c.
47. Pablo Alvarez Lopez, Michael Behrisch, Laura Bieker-Walz, Jakob Erdmann, Yun-Pang Flötteröd, Robert Hilbrich, Leonhard Lücken, Johannes Rummel, Peter Wagner, and Evamarie Wiesafe-tyner. Microscopic traffic simulation using sumo. In The 21st IEEE International Conference on Intelligent Transportation… [find](https://scholar.google.com/scholar?q=Pablo+Alvarez+Lopez%2C+Michael+Behrisch%2C+Laura+Bieker%2DWalz%2C+Jakob+Erdmann%2C+Yun%2DPang+Fl%C3%B6tter%C3%B6d%2C+Robert+Hilbrich%2C+Leonhard+L%C3%BCcken%2C+Johannes+Rummel%2C+Peter+Wagner%2C+and+Evamarie+Wiesafe%2Dtyner%2E+Microscopic+tr)
48. Pattie Maes. Agents that reduce work and information overload. In Readings in human–computer interaction, pages 811–821. Elsevier, 1995. [find](https://scholar.google.com/scholar?q=Pattie+Maes%2E+Agents+that+reduce+work+and+information+overload%2E+In+Readings+in+human%E2%80%93computer+interaction%2C+pages+811%E2%80%93821%2E+Elsevier%2C+1995%2E)
49. Reiichiro Nakano, Jacob Hilton, Suchir Balaji, Jeff Wu, Long Ouyang, Christina Kim, Christopher Hesse, Shantanu Jain, Vineet Kosaraju, William Saunders, et al. [Webgpt: Browser-assisted question-answering with human feedback](https://md.research.elamur.ai/free/en/papers/nakano-2021-webgpt-browser-assisted-question-answering/). arXiv preprint arXiv:2112.09332, 2021.
50. Maxwell Nye, Anders Johan Andreassen, Guy Gur-Ari, Henryk Michalewski, Jacob Austin, David Bieber, David Dohan, Aitor Lewkowycz, Maarten Bosma, David Luan, et al. [Show your work: Scratchpads for intermediate computation with language models](https://md.research.elamur.ai/free/en/papers/al-2021-show-your-work-scratchpads-intermediate/). arXiv preprint arXiv:2112.00114, 2021.

*Updated 2 November 2026*