Emily McMilin

Researcher at Meta FAIR working on code agents, world models, reinforcement learning, and causal approaches to language-model behavior.

Google Scholar LinkedIn Twitter

Research Interests

Code agents and software engineering benchmarks
World models for code generation
Self-play and self-improvement
Causal inference for language-model behavior

News

May 2026 ProgramBench, challenging agents to reproduce programs from only the compiled binary, is released.
Apr 2026 Self-Play SWE-RL, RL task generation and auto-curricula via self-play, accepted to ICML 2026.
Apr 2026 Invited speaker at VerifAI: AI Verification in the Wild workshop at ICLR 2026 (slides) (video).

Selected Work

ProgramBench: Can Language Models Rebuild Programs From Scratch?

John Yang*, Kilian Lieret*, Jeffrey Ma, Parth Thakkar, Dmitrii Pedchenko, Sten Sootla, Emily McMilin, Pengcheng Yin, Rui Hou, Gabriel Synnaeve, Diyi Yang, Ofir Press · 2026 benchmark

Toward Training Superintelligent Software Agents through Self-Play SWE-RL

Yuxiang Wei*, Zhiqing Sun, Emily McMilin, Jonas Gehring, David Zhang, Gabriel Synnaeve, Daniel Fried, Lingming Zhang, Sida Wang · ICML 2026 self-improvement

CWM: An Open-Weights LLM for Research on Code Generation with World Models

FAIR CodeGen Team · 2025 world models for code

Underspecification in Language Modeling Tasks

Emily McMilin · AAAI 2024 causal inference

Misc

Previously worked in electromagnetics; earlier work includes Full Duplex Radios (SIGCOMM 2013).
Contributed lessons to Udacity's Generative AI Nanodegree and Generative AI Fluency programs.
Served as the technical reviewer for Manning's Causal AI book (Ness, 2025).