Sriram Gopalakrishnan, Ph.D.

Sriram Gopalakrishnan, Ph.D.

AI Research Lead at J.P. Morgan Chase | Sequential-Decision Making Problems, Human-AI Interaction, & Generative AI

Synthetic Data Generation for Multi-Hop Logical Reasoning

Published:

Pillar: Hybrid-Agent Systems
Synthetic Data Multi-Hop Reasoning Multi-Agent Systems Boolean SAT

Overview

Evaluating whether Large Language Models genuinely perform multi-hop deductive reasoning rather than memorizing surface statistical correlations is a major bottleneck in AI evaluation.

At J.P. Morgan AI Research, we developed:

  • Multi-Agent Document Synthesis: Multi-agent architectures that construct synthetic document corpora with intricate, inter-document cross-references and multi-step logical dependencies.
  • Formal SAT Grounding: Translating Boolean satisfiability (SAT) mathematical problem instances into coherent natural language narratives to benchmark LLM logical consistency.
  • Token-Efficient In-Browser Coding Agents: Built secure coding agent tooling running client-side in browsers via Pyodide to allow non-coders to formulate data queries efficiently.