Alexander Miller

Timeline

1. 2016

   Key-Value Memory Networks

   He is first author of Key-Value Memory Networks for Directly Reading Documents, published at EMNLP 2016 in Austin with Adam Fisch, Jesse Dodge, Amir-Hossein Karimi, Antoine Bordes and Jason Weston. The paper introduces the WikiMovies question-answering dataset, and its method reaches state-of-the-art results on the WikiQA benchmark.

2. 2017

   ParlAI for dialog research

   In May 2017, he, Jason Weston and Will Feng announce ParlAI, an open-source platform from Facebook AI Research for training and testing dialog models across many tasks at once. The EMNLP 2017 system-demonstration paper, with him as first author, reports over 20 tasks in the first release.

3. 2019

   Language models as knowledge bases

   In 2019, he coauthors Language Models as Knowledge Bases?, which finds that BERT holds relational knowledge competitive with traditional NLP methods without fine-tuning. He also coauthors the ConvAI2 report on the NeurIPS competition for open-domain chatbots.

4. 2020

   The NetHack Learning Environment

   He is third author of The NetHack Learning Environment in 2020, written with Heinrich Küttler, Nantas Nardelli, Roberta Raileanu, Marco Selvatici, Edward Grefenstette and Tim Rocktäschel. NLE is a scalable, procedurally generated environment for reinforcement learning research based on the game NetHack.

5. 2022–2023

   Diplomacy research and an ICLR honor

   In October 2022, he coauthors Mastering the Game of No-Press Diplomacy via Human-Regularized Reinforcement Learning and Planning, which trains the Diplodocus agents; two of them place first and third by an Elo ratings model in a 200-game tournament with 62 human participants. The paper receives an Honorable Mention at ICLR 2023. At Meta AI, he supports a project to get AI to play complex multiplayer board games such as Hanabi and Diplomacy.

6. By 2023

   Research engineering manager at Meta

   By 2023, he is a Research Engineering Manager at Meta AI. He starts as a Research Engineer in New York, spends two years in London helping grow Meta AI's presence in Europe, and returns to New York. He holds a Bachelor of Arts in Computer Science from Cornell and is enrolled in NYU's Master of Science program in Computer Science.

7. 2025

   Benchmarks for AI research agents

   In 2025, he coauthors the Automated LLM Speedrunning Benchmark with colleagues at Meta; it uses the NanoGPT speedrun, a competition to train a GPT-2 model in the shortest time, to test whether AI agents can reproduce known improvements. The July 2025 paper AI Research Agents for Machine Learning, which he also coauthors, raises the Kaggle-medal success rate on MLE-bench lite from 39.6% to 47.7%.

8. As of October 2026

   Core contributor to Muse Glimmer

   In 2026, he coauthors Asking the Right Questions: Improving Reasoning with Generated Stepping Stones with colleagues at FAIR at Meta. In October 2026, he is among the core contributors to Muse Glimmer 30B, an open-source 30-billion-parameter on-device agentic model from Meta Superintelligence Labs.