Correspondence - Small Language Models are the Future of Agentic AI

Correspondence and Feedback

This page contains correspondence, feedback, and discussions related to our paper "Small Language Models are the Future of Agentic AI." We believe in fostering open dialogue within the research community and welcome constructive critiques, suggestions, and contributions that can help refine our position and advance the discussion on the effective use of AI resources.

All correspondence published here has been shared with permission from the corresponding authors. If you would like to contribute to this discussion, please reach out to us via email at agents@nvidia.com.

Discussion on Reasoning LLMs and Agentic Workflows

July 8, 2025

From: Alexei Novikov, Opplane
Subject: About Small Language Models are the Future of Agentic AI

Dear authors,

Thank you very much for the excellent article. I thoroughly enjoyed reading it. I may be wrong, but I think you missed one point.

I'm seeing more and more often that visioneers and even sometimes practitioners are making statements that reasoning LLMs are the genuine alternative to agentic workflows. So it may be worth extending your section 2.2

The counterarguments to these are:

The reasoning LLMs are just standard LLMs with an additional Chain-of-Thought post-training. Therefore, if your particular problem was not in the training set, they are not necessarily effective.
Additional "reasoning" post-training requires more resources, making the LLMs more expensive.

I think it will also be worth noting that using SLMs in the agentic workflow will help facilitate the broader adoption of agentic workflows, as for many financial institutions and companies, using third-party SaaS LLMs is impossible due to the handling of PII data and/or bank and business secrecy.

Sincerely yours,
Alexei Novikov

From: Peter Belcak, NVIDIA Research
Subject: Re: About Small Language Models are the Future of Agentic AI

Hi Alexei,

Thanks for the feedback on the paper. Your contribution is much appreciated, you make good points.

From our perspective when writing the paper, reasoning has been considered as model self-agency that may still be used for more dynamism in specialized tasks even by SLMs — but you are right that reasoning has been put on a pedestal to the point where some argue it could take place of carefully orchestrated agentic workflows.

Do you wish our correspondence to be published on our website?

Best,
Peter

Contribute to the Discussion

We welcome additional correspondence, critiques, and contributions to this discussion. If you have feedback on our position, questions about implementation, or would like to share your own research or experiences related to small language models and agentic AI, please reach out to us.

Contact: agents@nvidia.com
Subject: Please include "SLM Agents Correspondence" in your subject line

All correspondence will be reviewed and, with permission from the corresponding author, published on this page to foster continued discussion within the research community.

BibTeX

@misc{belcak2025small,
  title        = {Small Language Models are the Future of Agentic AI},
  author       = {Belcak, Peter and Heinrich, Greg and Diao, Shizhe and Fu, Yonggan and Dong, Xin and Muralidharan, Saurav and Lin, Yingyan Celine and Molchanov, Pavlo},
  year         = {2025},
  eprint       = {2506.02153},
  archivePrefix= {arXiv},
  primaryClass = {cs.AI},
  url          = {https://arxiv.org/abs/2506.02153},
  doi          = {10.48550/arXiv.2506.02153}
}