COLLABORATIVE RESEARCH: III: TOWARDS RELIABLE, EFFICIENT, AND ADAPTIVE MULTI-AGENT COLLABORATION: A GRAPH-C…

Assistance award details

Assistance award identifier

ASST_NON_2616632_049

FAIN

2616632

Recipient

ARIZONA STATE UNIVERSITY

Awarding agency

National Science Foundation

Funding agency

National Science Foundation

Assistance listing (CFDA)

47.070

Total obligated

$250,000

Period

Aug 1, 2026 – Jul 31, 2029

Fiscal year

2026

Place of performance state

AZ

Source file

FY2026_All_Assistance_Full_20261006.zip

Description

COLLABORATIVE RESEARCH: III: TOWARDS RELIABLE, EFFICIENT, AND ADAPTIVE MULTI-AGENT COLLABORATION: A GRAPH-CENTRIC DESIGN -LARGE LANGUAGE MODELS ARE INCREASINGLY BEING USED TO BUILD MULTI-AGENT ARTIFICIAL INTELLIGENCE (AI) SYSTEMS, IN WHICH MULTIPLE AI AGENTS COMMUNICATE AND COLLABORATE TO SOLVE COMPLEX PROBLEMS. THESE SYSTEMS HAVE THE POTENTIAL TO SUPPORT APPLICATIONS SUCH AS SCIENTIFIC DISCOVERY, SOFTWARE ENGINEERING, DECISION SUPPORT, EDUCATION, AND AI-ASSISTED ANALYSIS. HOWEVER, CURRENT MULTI-AGENT SYSTEMS OFTEN RELY ON STATIC AND MANUALLY DESIGNED COMMUNICATION PATTERNS. AS A RESULT, THEY CAN BE VULNERABLE TO UNRELIABLE OR ADVERSARIAL AGENTS, EXCESSIVE AND REDUNDANT COMMUNICATION, AND UNEXPECTED CHANGES IN AGENT AVAILABILITY, INTERACTION STRUCTURE, OR TASK OBJECTIVES. THESE LIMITATIONS MAKE IT DIFFICULT TO DEPLOY MULTI-AGENT SYSTEMS RELIABLY AND EFFICIENTLY IN REAL-WORLD ENVIRONMENTS. THIS PROJECT DEVELOPS GRAPH-CENTRIC MULTI-AGENT SYSTEMS (GRAPH-MAS), A NEW FRAMEWORK THAT REPRESENTS AGENTS AS NODES AND THEIR INTERACTIONS AS EDGES IN A DYNAMIC COMMUNICATION GRAPH. BY MAKING THE COLLABORATION STRUCTURE EXPLICIT AND OPTIMIZABLE, THE PROJECT ADVANCES THE NATIONAL INTEREST THROUGH CONTRIBUTIONS TO RELIABLE, EFFICIENT, AND ADAPTIVE ARTIFICIAL INTELLIGENCE, WHILE ALSO TRAINING STUDENTS IN EMERGING AREAS OF MACHINE LEARNING, LARGE LANGUAGE MODELS, AND AGENTIC AI. THIS PROJECT DEVELOPS NEW GRAPH-BASED METHODS FOR DESIGNING, EXECUTING, AND ADAPTING MULTI-AGENT SYSTEMS POWERED BY LARGE LANGUAGE MODELS. THE RESEARCH INVESTIGATES THREE COMPLEMENTARY DIRECTIONS. FIRST, THE PROJECT DEVELOPS RELIABILITY-AWARE METHODS THAT DETECT UNRELIABLE OR ADVERSARIAL AGENTS, ESTIMATE UNCERTAINTY IN AGENT COMMUNICATION, AND CONTAIN THE SPREAD OF INCORRECT INFORMATION THROUGH THE COLLABORATION GRAPH. SECOND, THE PROJECT DEVELOPS EFFICIENCY-ORIENTED METHODS THAT REDUCE UNNECESSARY COMMUNICATION AND COMPUTATION BY IDENTIFYING REDUNDANT AGENTS, EDGES, AND MESSAGES WHILE PRESERVING SYSTEM-LEVEL