IARPA teases event for large language model vulnerability mitigation tools

wildpixel/Getty Images

The intelligence agency offered details on a new potential contract and October event to showcase tools that can detect and remediate vulnerabilities in large language models.

The Intelligence Advanced Research Projects Agency is crafting a new contract to examine what tools are available to address threats and vulnerabilities within the large language models that support artificial intelligence and machine learning tools. 

The agency issued a pre-solicitation notice for a broad agency announcement on Aug. 25, detailing its plan for a research and development procurement dubbed Bias Effects and Notable Generative AI Limitations — or BENGAL — and a proposer’s day event scheduled for Oct. 24 in Washington, D.C. 

“The goal of BENGAL is to understand LLM threat modes, quantify them and to find novel

methods to address threats and vulnerabilities or to work resiliently with imperfect models,” the announcement said. “IARPA seeks to develop and incorporate novel technologies to efficiently probe large language models to detect and characterize LLM threat modes and vulnerabilities.”

The notice comes on the heels of a request for information IARPA issued earlier this month to gain more insight into what threats and vulnerabilities could be inherent in LLM. 

According to the notice, stakeholders participating in the proposer’s day will focus on multiple topic domains, identifying threat modes within each domain with a crafted taxonomy and developing tools “to efficiently probe LLM models to detect, characterize and mitigate biases, threats or vulnerabilities.”

Topic areas and additional requirements are expected to be introduced at the proposer’s day event, but potential participants can detail their expertise to potentially be paired with teaming partners working on the project.

The proposer’s day will feature a maximum of 20 lightning talks lasting five minutes to highlight capabilities that align with the project. 

For participants looking to attend the event, in-person access will be available for a single representative — with the possible exception of professors from the same university — and virtual access will accommodate multiple attendees remotely. 

The deadline to submit lightning talk presentations is by 12 p.m. EST on Oct. 11, while attendance registration must be complete by 12 p.m. EST on Oct. 18.