// RESEARCH

PAPER
OF THE WEEK

Each week our team discusses a research paper in depth. These summaries are generated by AI from the recorded discussion.

"Ruiyao Xu", "Yanfei Chen", "Zhongying CuiZhu", "Bhavana Dalvi Mishra", "Yifei Ming", "Han Yu", "Rujun Han", "Chen-Yu Lee", "Tomas Pfister" — Harness2: Recursive Agent Harnessing for an Open World

Harness2 — A Fresh Harness for Every Task, and a Library of Lessons on How to Build It

We read Google Cloud AI Research's Harness2, which rebuilds an agent's harness for each task without a score to guide it, and asked whether a stronger model editing a weaker one's harness really counts as self-improvement

2026-11-05

"Jacky Kwok", "Hangoo Kang", "Tarun Suresh", "Jon Saad-Falcon", "Marco Pavone", "Christopher Ré", "Azalia Mirhoseini" — Contrastive Language Models

Contrastive Language Models — A Fast Decision Model Built from Two Frozen Encoders and a Dot Product"

We read Stanford and NVIDIA's CLM, a System One model that picks actions by embedding similarity instead of generation, and looked hard at what its speedups, fine-tuned results and probabilities really show.

2026-10-16

Anthony Hu", "Václav Volhejn", "Adrien Ramanana Rahary", "Chris Mulder", "Aditya Makkar", "Alyx Liao", "Amélie Royer", "Manu Orsini", "Adam Jelley", "Eloi Alonso", "Florian Laurent", "Fredrik Norén", "James Swingos", "Jan Hünermann", "Kent Rollins", "Lucas Hosseini", "Matthieu Le Cauchois", "Maxim Peter", "Pim de Witte", "Tim Brown", "Vincent Micheli", "Moritz Böhle", "Gabriel de Marmiesse", "Viktoriia Sharmanska", "Lucia Specia", "Michael Black", "Patrick Pérez — MIRA: Multiplayer Interactive World Models with Representation Autoencoders

MIRA — Making Multiplayer World Models Stable by Removing Hidden Information

We discuss MIRA, the real-time multiplayer Rocket League world model from General Intuition, Kyutai and Epic Games: why conditioning on all four players' views keeps the simulation coherent, how flow matching with per-frame noise makes hour-long rollouts possible, and what still separates this from a simulator you could train robots in.

2026-08-04

Amir Zandieh, Majid Daliri, Majid Hadian, Vahab Mirrokni — TurboQuant: Online Vector Quantization with Near-optimal Distortion Rate

TurboQuant — Making Quantization Less Lossy by Randomly Rotating the Problem

2026-04-27

Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, Y.K. Li, Y. Wu, Daya Guo — DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

Unpacking GRPO and the Evolution of RL in LLMs