Paper Feed

Paper Feed: March 2026

Highlighting research I find interesting and think may deserve more attention (as of 03/16/26), with a heavier-than-usual emphasis on AI economics, oversight, and training dynamics.

AI Economics and Forecasting

The Economics of Transformative AI
Jan Kulveit, David Duvenaud, Raymond Douglas (2026)
Thoughts by a non-economist on AI and economics
Boaz Barak (2025)
Forecasting AI Time Horizon Under Compute Slowdowns
Parker Whitfill, Ben Snodin, Joel Becker (2025)
Anthropic Economic Index report: Economic primitives
Ruth Appel, Maxim Massenkoff, Peter McCrory, [...], Alex Tamkin (2026)
Will compute bottlenecks prevent a software intelligence explosion?
Parker Whitfill, Cheryl Wu (2025)
Will Compute Bottlenecks Prevent an Intelligence Explosion?
Parker Whitfill, Cheryl Wu (2025)
On the Origin of Algorithmic Progress in AI
Hans Gundlach, Alex Fogelson, Jayson Lynch, [...], Neil Thompson (2025)
Most Algorithmic Progress is Data Progress
Beren Millidge (2025)
2025 letter
Zhengdong Wang (2025)
LLM verification
Tom Cunningham (2025)

Security and Control

The AI Infrastructure Security Shortlist
Abbey Chaver (2026)
Defending Against Intelligent Attackers at Large Scales
Andrew J. Lohn (2025)
Frontier AI safety regulations: A reference for lab staff
Miles Kodama, Michael Chen (2026)
Five lessons from having helped run an AI-Biology RCT
Luca Righetti (2026)
Distillation Robustifies Unlearning
Bruce W. Lee, [...], Alex Cloud, Alexander Matt Turner (2025)
Salient Directions in AI Control
Bruce W. Lee (2026)
Optimizing AI Agent Attacks With Synthetic Data
Chloe Loughridge, Paul Colognese, Avery Griffin, [...], Joe Benton (2025)

Alignment and Monitorability

Alignment Faking Mitigations
Johannes Gasteiger, Vlad Mikulik, Hoagy Cunningham, [...], Evan Hubinger (2025)
A Pragmatic Way to Measure Chain-of-Thought Monitorability
Scott Emmons, Roland S. Zimmermann, David K. Elson, Rohin Shah (2025)
RL-Obfuscation: Can Language Models Learn to Evade Latent-Space Monitors?
Rohan Gupta, Erik Jenner (2025)
Training Qwen-1.5B with a CoT legibility penalty
Fabien Roger (2025)
Current language models struggle to reason in ciphered text
Fabien Roger, Shiyuan Guo (2025)
The behavioral selection model for predicting AI motivations
Alex Mallen, Buck Shlegeris (2025)
Refusals that could become catastrophic
Fabien Roger (2026)
Building AIs that do human-like philosophy
Joe Carlsmith (2026)

Training

Relative Scaling Laws for LLMs
William Held, David Hall, Percy Liang, Diyi Yang (2025)
Deriving Neural Scaling Laws from the statistics of natural language
Francesco Cagnetta, Allan Raventos, Surya Ganguli, Matthieu Wyart (2026)
What really matters in matrix-whitening optimizers?
Kevin Frans (2025)
Muon in Modded NanoGPT
Varun Neal (2025)
Better Training Data Attribution via Better Inverse Hessian-Vector Products
Andrew Wang, Elisa Nguyen, Runshi Yang, [...], Roger Grosse (2025)
A Technical Tour of the DeepSeek Models from V3 to V3.2
Sebastian Raschka (2025)

Theory

Steering Language Models with Weight Arithmetic
Constanza Fierro, Fabien Roger (2025)
Training the Untrainable: Introducing Inductive Bias via Representational Alignment
Vighnesh Subramaniam, David Mayo, Colin Conwell, [...], Andrei Barbu (2024)
Excess Description Length of Learning Generalizable Predictors
Elizabeth Donoway, Hailey Joren, Fabien Roger, Jan Leike (2026)
Secret mixtures of experts inside your LLM
Enric Boix-Adsera (2025)
From Entropy to Epiplexity: Rethinking Information for Computationally Bounded Intelligence
Marc Finzi, Shikai Qiu, Yiding Jiang, [...], Andrew Gordon Wilson (2026)