Rishi Desai
I'm currently building RL environments for agentic coding at Abundant AI. Previously I graduated from Stanford with a BS and MS in computer science and minor in music.
My recent work focuses on building tools for creating RL environments, evaluating dataset quality, and studying agent traces.
I recently released SWE-Marathon, a benchmark for autonomous long-horizon software engineering work. It was featured on the GLM 5.2 model card.
Posts
SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?
June 4, 2026
Multi-hour SWE benchmark spanning library reproductions, full-stack product clones, and MLE.
Building a platform to automate RL env creation and QA
May 27, 2026
Turning RL environment creation into an automated loop.
Designing eval harnesses that prevent reward hacking
May 13, 2026
Two trust boundaries that prevent coding agents from cheating their graders.
SWE-gen: scaling SWE-bench task generation
January 18, 2026
SWE-gen turns any GitHub PR into a SWE-bench style RL environment.
CharForge: character consistency with one reference image
June 27, 2025
A guide for achieving character consistency from a single reference image.
Unlocking Character LoRAs with Structured Captions
May 8, 2025
A guide for image captioning and prompt optimization.
Fixing Face Consistency in GPT-4o Image Gen
April 21, 2025
A guide to enhancing faces in AI generated images.
Publications
SWE-Marathon: Can Agents Autonomously Complete Ultra-Long-Horizon Software Work?
Rishi Desai et al.
arXiv
Privacy-Constrained Policies via Mutual Information Regularized Policy Gradients
Chris Cundy*, Rishi Desai*, and Stefano Ermon
AISTATS 2024
Detecting Human-Object Relationships in Videos
Jingwei Ji, Rishi Desai, Juan Carlos Niebles
ICCV 2021
BioFabric Visualization of Network Alignments
Rishi Desai, William JR Longabaugh, Wayne Hayes
Recent Advances in Biological Network Analysis, Springer, 2021