ARPA-H RFI 75N992-25-RFI-106 - Revised.pdf

Request for Information – 75N992-25-RFI-106 Advanced Research Projects Agency for Health

Agentic Artificial Intelligence (AI) for Healthcare

The Advanced Research Projects Agency for Health (ARPA-H) is seeking information on possible directions in Agentic Artificial Intelligence for Healthcare.

THIS IS A REQUEST FOR INFORMATION (RFI) ONLY. In accordance with Federal Acquisition Regulation (FAR) 15.201(e), responses to this RFI are not offers and cannot be accepted by the Government to form a binding contract. This RFI is issued solely for information and planning purposes and does not constitute a Request for Proposal (RFP). The Government will not provide reimbursement for any costs incurred in responding to this RFI.

1.0 Background:

The Advanced Research Projects Agency for Health (ARPA-H) is a government agency which seeks to accelerate better health outcomes for everyone by supporting the development of high-impact solutions to society's most challenging health problems.

This RFI seeks to enhance ARPA-H’s understanding of the Agentic AI landscape to aid ARPA-H in setting strategic direction for its research and development initiatives leveraging Agentic AI to improve health outcomes.

The assumptions shaping this request for information are that:

1) Agency is the natural progression of Artificial Intelligence.

2) Currently, autonomous AI agents are generally deployed for low consequence situations and tasks, such as personalized content recommendations or home appliances (e.g. automated vacuums). However, AI agents deployed in high consequence situations almost always have a human in the loop such as self-driving cars or surgical robots.

3) Current and expected advances in Agentic AI will facilitate the deployment of autonomous AI agents in healthcare settings and high consequence situations where the ability for AI agents to collaborate with each other and with humans will have the potential to revolutionize modern medicine, accelerate science, and relieve the burden of overtaxed healthcare professionals and shift many tasks to AI agents.

2.0 Request for Information:

ARPA-H is requesting information on several topics concerning Agentic AI. To accelerate the agency’s mission, ARPA-H is looking for examples related to healthcare when possible. If the examples provided are not currently applicable to healthcare, please explain how they could be applied in the future. For this RFI, you may respond to some or all of the topics below. Please tailor your responses with the assumption that they will be reviewed by persons with working knowledge of modern machine learners as well as Foundation Models, though not necessarily Frontier Models. Please label your responses according to the numbered questions listed in the request (for further response formatting details, see Section 3 below).

1. Agentic AI – What are the most recent implementations for agent instantiation allowing for a.) autonomy, b.) adaptive/continual learning, and c.) cognitive architecture. Are there any mechanisms to instantiate agents that require no or little effort from end users? Can agents be powered by Foundation Models beyond LLMs (should they or should we simply have LLM agents use other foundation models)? Once created, how can third parties summon these agents beyond downloading the source code?

2. Multi-Agentic AI – Beyond instantiating and then deploying multiple agents, what are the potential advances for interacting agents? Provide examples of an LLM agent interacting with other Foundation Models such as Scientific ML (AlphaFold, Climate AI’s, etc.) and describe the overall success/failures? What are the resources and protocols needed to bring together independent agents created by differing groups or entities to solve scientific or applied problems (using different frameworks such as PyTorch|TensorFlow|JAX potentially with different programming languages)?

3. Interoperability of Agents – If independent agents do interact, what is currently in place/needed for communicating between agents created in different platforms/frameworks (e.g. in very large multi-agentic settings created with different frameworks and different languages such as Python and C++)?

Specifically, what methods and protocols are available to control the flow of dialogue and interactions between multiple agents? One example: should the output of one agent be fed into the input of the rest of the agents or only a specific one or a subset thereof for open-ended scientific collaborations to be mindful of context limits and hallucinations? Are there any additional protocols that need to be developed? What about agents in the environment (robots in laboratories, etc.)? Beyond sensors, can they and do they communicate or are they all independent actors?

4. How do we discern agent duplicity (i.e. discriminate as friend or foe for a specific agent)?

5. What is the current state of the art research and development for agentic assurance and trustworthiness?

6. How can we quantify agentic robustness with respect to:

a. performance (doing what they are supposed to do)

b. security

c. privacy

d. agents being and staying on task

e. an agent’s ability for self-behavior modeling/updating

7. How does agentic AI scale? Can we estimate the number of agents for a given task or number of tasks for a given agent? How does the underlying LLM / foundation model limit scalability for the agents? Can we quantify the computational footprint for agents given a specific task?

8. How is agentic autonomy decided (e.g. the levels for human - agent collaboration)? What are the ethical and practical considerations for each level of autonomy?

9. What is the current research for autonomous agents deployed in high con-sequence scenarios?

10. How is resource management optimized (e.g. energy consumption as part of scalability)?

11. What are the known potential risks for agentic AI? What are speculative risks?

For instance, what are the potential cyber-threats arising from agentic AI? Do agents pose new threats for existing/known cyber-systems? What is the realm of possible novel threats?

12. With the construction of an AI Scientist for ML research, what are the barriers to transfer in silico Agentic Science to in vitro Agentic Science, that is how can we get agents to perform bench science?

13. Are there other considerations that ARPA-H should be aware of that are not covered in the previous questions?

3.0 Instructions

3.1 Dates: Interested persons and organizations are invited to submit responses to this RFI on or before 4:00 PM EST on November 29, 2024.

3.2 Addresses: Interested individuals and organizations should submit responses electronically to eureka@arpa-h.gov and include “RFI Response: Directions in Agentic AI” and the name of their institution in the subject line of the email. Due to time constraints, electronic submissions received after the deadline may not be taken into consideration.

3.3 Instructions: Responses should be limited to 10 pages, in legibly sized font.

Responses should address both the current state of the art, including the gaps in current research and development, and near future directions for any topic chosen.

Where possible, quantitative metrics and specifications (how fast, how accurate, how successful, how much computational power, how expensive) describing current and future system performance should be given and substantiated with references to published literature. Respondents may wish to refer to levels of autonomy for the agents 1-5, as follows:

Level Type of Autonomy 1 No autonomy – human in the loop at every step 2 Task Autonomy 3 Conditional Autonomy 4 High Autonomy 5 Full Autonomy

Respondents may also wish to refer to the level of interactions between agents as follows:

Level Number of Interactions Solo 1 Small 2-5 Medium 5-10 Large 10+

Respondents may address multiple topics in their response. Each respondent is requested to submit only one response.

Response to this RFI is voluntary. Responses to this RFI may be viewed by Government personnel and support contractors. The Government will review mailto:eureka@arpa-h.gov responses for market research purposes only. Any information obtained from this RFI is intended to be used by the Government on a non-attribution basis for planning and strategy development. ARPA-H will not respond to individual submissions. A response to this RFI will not be viewed as a binding commitment to develop or pursue the project or ideas discussed. Comments submitted in response to this notice are subject to the Freedom of Information Act (FOIA). Any proprietary information shall be clearly marked in the document.