Jatin Ganhotra

I am a Senior Software Engineer at IBM Research, AI for Code, Thomas J. Watson Research Center, where I lead the development of autonomous SWE-Agents for intelligent code generation, issue resolution, and software testing.

I am the project lead and architect of iSWE-Agent, IBM Research’s software engineering agent that secured the top position on the Multi-SWE-Bench and SWE-PolyBench Java leaderboards. I also created SWE-Bench-Arena, a platform for rigorous blind evaluation of AI-generated code — assessing quality dimensions like maintainability, readability, and production readiness beyond just passing tests.

My research interests span AI-driven software engineering, conversational AI, dialog systems, and natural language processing. I have published across top-tier AI and software engineering venues, including ICML, ICSE, EMNLP, ACL, NAACL, TACL, Interspeech, and ICASSP.

I earned my M.S. in Computer Science from the University of Illinois, Urbana-Champaign and my B.Tech in Computer Engineering from the National Institute of Technology (NIT) Kurukshetra.

You can learn more about my current work here: IBM SWE Agents and iSWE-Agent.

Latest posts

News

Sep 21, 2026 On September 17, 2026, I gave an invited talk titled Green Means Stop: Building the Reward Signal Code Optimization Never Had at the Fastcode session at IEEE HPEC 2026. The talk asked why capable coding agents stop after producing correct code with only a modest speedup.
IEEE HPEC 2026 speaker card for Jatin Ganhotra, Senior Software Engineer at IBM, 14–18 September 2026
Event announcement shared by IEEE Boston Section before HPEC 2026.
Aug 20, 2026 My opinion piece The Code Optimization Flywheel Won’t Spin Itself is live on BLOG@CACM. On SWE-bench Verified, the top agent resolves 79.20% issues. On GSO, a benchmark of optimization tasks, the best system reaches 47.1% Opt@1. I argue that the second number will not climb the way the first one did. Every repository already contains a reward for correctness. The test suite sets the target, tells an agent when it gets closer, and says when to stop. However, a request for speed gives only a timer.
BLOG@CACM header: The Code Optimization Flywheel Won't Spin Itself, by Jatin Ganhotra
The Code Optimization Flywheel Won't Spin Itself on BLOG@CACM (20 August 2026)
Jun 18, 2026 iSWE-Agent, IBM Research’s software engineering agent, achieved Rank #1 on the SWE-PolyBench (Verified) Python leaderboard with a 58.41% resolution rate (66/113 resolved), surpassing the previous best from Atlassian Rovo Dev (54.87%). This is iSWE-Agent’s first submission on Python — after topping the Java splits of Multi-SWE-Bench and SWE-PolyBench — and its first to run on OpenAI’s GPT models, demonstrating that the same agent generalizes across both programming languages and model ecosystems.
SWE-PolyBench Verified leaderboard sorted by Python: iSWE-Agent first at 58.41%, Atlassian Rovo Dev second at 54.87%
iSWE-Agent achieves Rank-1 on SWE-PolyBench (Verified) Python leaderboard (18 June 2026)
Related resources:
  1. iSWE-Agent arXiv paper - Resolving Java Code Repository Issues with iSWE Agent
Feb 06, 2026 iSWE-Agent, IBM Research’s software engineering agent, achieved Rank #1 on the SWE-PolyBench (full) Java leaderboard with a 33.33% resolution rate. On the Verified subset, iSWE-Agent scores 46.38% on Java, matching Atlassian Rovo Dev and significantly outperforms PrometheusV1.2 + GPT-5 (33.33%).
SWE-PolyBench: Multi-Language Evaluation of Coding Agents
SWE-PolyBench full leaderboard sorted by Java: iSWE-Agent first at 33.33%, ahead of Amazon Q Developer Agent at 26.67%
iSWE-Agent achieves Rank-1 on SWE-PolyBench (Full) Java leaderboard (6 February 2026)
SWE-PolyBench Verified leaderboard sorted by Java: iSWE-Agent and Atlassian Rovo Dev both at 46.38%
iSWE-Agent on SWE-PolyBench (Verified) Java — matches Atlassian Rovo Dev, outperforms PrometheusV1.2
Dec 01, 2025 iSWE-Agent, IBM Research’s software engineering agent, achieved Rank 1 on the Multi-SWE-Bench Java leaderboard with a 33% resolution rate, substantially outperforming the previous best score of 28.9%.
Multi-SWE-bench Java leaderboard with IBM's iSWE-Agent first at 33.59% and iSWE-OpenModels second at 31.25%
iSWE-Agent achieves Rank-1 on Multi-SWE-Bench Java leaderboard (1 December 2025)
Related resources:
  1. IBM Research Blog - iSWE-Agent