Portrait de Karam Ghanem n'est pas disponible

Karam Ghanem

Alumni

Publications

TapeAgents: a Holistic Framework for Agent Development and Optimization
Dzmitry Bahdanau
Nicolas Gontier
Gabriel Huang
Ehsan Kamalloo
Rafael Pardinas
Alexandre Piché
Torsten Scholak
Oleh Shliazhko
Jordan Prince Tremblay
Karam Ghanem
Soham Parikh
Mitul Tiwari
Quaizar Vohra
We present TapeAgents, an agent framework built around a granular, structured log tape of the agent session that also plays the role of the … (voir plus)session's resumable state. In TapeAgents we leverage tapes to facilitate all stages of the LLM Agent development lifecycle. The agent reasons by processing the tape and the LLM output to produce new thought and action steps and append them to the tape. The environment then reacts to the agent's actions by likewise appending observation steps to the tape. By virtue of this tape-centred design, TapeAgents can provide AI practitioners with holistic end-to-end support. At the development stage, tapes facilitate session persistence, agent auditing, and step-by-step debugging. Post-deployment, one can reuse tapes for evaluation, fine-tuning, and prompt-tuning; crucially, one can adapt tapes from other agents or use revised historical tapes. In this report, we explain the TapeAgents design in detail. We demonstrate possible applications of TapeAgents with several concrete examples of building monolithic agents and multi-agent teams, of optimizing agent prompts and finetuning the agent's LLM. We present tooling prototypes and report a case study where we use TapeAgents to finetune a Llama-3.1-8B form-filling assistant to perform as well as GPT-4o while being orders of magnitude cheaper. Lastly, our comparative analysis shows that TapeAgents's advantages over prior frameworks stem from our novel design of the LLM agent as a resumable, modular state machine with a structured configuration, that generates granular, structured logs and that can transform these logs into training text -- a unique combination of features absent in previous work.
2024-12-11
ArXiv (prépublication)
doi.org
arxiv.org
TapeAgents: a Holistic Framework for Agent Development and Optimization
Dzmitry Bahdanau
Nicolas Gontier
Gabriel Huang
Ehsan Kamalloo
Rafael Pardinas
Alex Pich'e
Torsten Scholak
Oleh Shliazhko
Jordan Prince Tremblay
Karam Ghanem
Soham Parikh
Mitul Tiwari
Quaizar Vohra
We present TapeAgents, an agent framework built around a granular, structured log tape of the agent session that also plays the role of the … (voir plus)session's resumable state. In TapeAgents we leverage tapes to facilitate all stages of the LLM Agent development lifecycle. The agent reasons by processing the tape and the LLM output to produce new thought and action steps and append them to the tape. The environment then reacts to the agent's actions by likewise appending observation steps to the tape. By virtue of this tape-centred design, TapeAgents can provide AI practitioners with holistic end-to-end support. At the development stage, tapes facilitate session persistence, agent auditing, and step-by-step debugging. Post-deployment, one can reuse tapes for evaluation, fine-tuning, and prompt-tuning; crucially, one can adapt tapes from other agents or use revised historical tapes. In this report, we explain the TapeAgents design in detail. We demonstrate possible applications of TapeAgents with several concrete examples of building monolithic agents and multi-agent teams, of optimizing agent prompts and finetuning the agent's LLM. We present tooling prototypes and report a case study where we use TapeAgents to finetune a Llama-3.1-8B form-filling assistant to perform as well as GPT-4o while being orders of magnitude cheaper. Lastly, our comparative analysis shows that TapeAgents's advantages over prior frameworks stem from our novel design of the LLM agent as a resumable, modular state machine with a structured configuration, that generates granular, structured logs and that can transform these logs into training text -- a unique combination of features absent in previous work.
2024-12-11
ArXiv (prépublication)
arxiv.org
arxiv.org
TapeAgents: a Holistic Framework for Agent Development and Optimization
Dzmitry Bahdanau
Nicolas Gontier
Gabriel Huang
Ehsan Kamalloo
Rafael Pardinas
Alex Pich'e
Torsten Scholak
Oleh Shliazhko
Jordan Prince Tremblay
Karam Ghanem
Soham Parikh
Mitul Tiwari
Quaizar Vohra
We present TapeAgents, an agent framework built around a granular, structured log tape of the agent session that also plays the role of the … (voir plus)session's resumable state. In TapeAgents we leverage tapes to facilitate all stages of the LLM Agent development lifecycle. The agent reasons by processing the tape and the LLM output to produce new thought and action steps and append them to the tape. The environment then reacts to the agent's actions by likewise appending observation steps to the tape. By virtue of this tape-centred design, TapeAgents can provide AI practitioners with holistic end-to-end support. At the development stage, tapes facilitate session persistence, agent auditing, and step-by-step debugging. Post-deployment, one can reuse tapes for evaluation, fine-tuning, and prompt-tuning; crucially, one can adapt tapes from other agents or use revised historical tapes. In this report, we explain the TapeAgents design in detail. We demonstrate possible applications of TapeAgents with several concrete examples of building monolithic agents and multi-agent teams, of optimizing agent prompts and finetuning the agent's LLM. We present tooling prototypes and report a case study where we use TapeAgents to finetune a Llama-3.1-8B form-filling assistant to perform as well as GPT-4o while being orders of magnitude cheaper. Lastly, our comparative analysis shows that TapeAgents's advantages over prior frameworks stem from our novel design of the LLM agent as a resumable, modular state machine with a structured configuration, that generates granular, structured logs and that can transform these logs into training text -- a unique combination of features absent in previous work.
2024-12-11
ArXiv (prépublication)
doi.org
arxiv.org
Multivariate analytical approaches for investigating brain-behavior relationships
E. Leighton Durham
Karam Ghanem
Andrew J. Stier
Carlos Cardenas-Iniguez
Gabrielle E. Reimann
Hee Jung Jeong
Randolph M. Dupont
Xiaoyu Dong
Tyler M. Moore
Marc G. Berman
Benjamin B. Lahey
Danilo Bzdok
Antonia N. Kaczkurkin
2023-07-31
Frontiers in Neuroscience (publié)
doi.org