Paper Feed: April 2026

Monitoring and Misalignment

• How we monitor internal coding agents for misalignment
  OpenAI (2026)
• Metagaming matters for training, evaluation, and oversight
  Bronson Schoen, Jenny Nitishinskaya (2026)
• Eval awareness in Claude Opus 4.6's BrowseComp performance
  Anthropic (2026)
• Stress-testing asynchronous monitoring of AI coding agents
  AISI (2026); companion paper: Async Control: Stress-testing Asynchronous Control Measures for LLM Agents
• COT control: The Word Disappears, but the Thought Does Not
  Pranjal Garg (2026)
• AIs exhibit a (misaligned) drive to stop early
  Ryan Greenblatt (2026)

Generalization

• How far does alignment midtraining generalize?
  Tomek Korbak, Cameron Raymond, Micah Carroll, [...], Ian Kivlichan (2026)
• Scaling Reward Modeling without Human Supervision
  Jingxuan Fan, Yueying Li, Zhenting Qi, [...], Hanlin Zhang (2026)
• AIs Should Have Proactive Prosocial Drives
  Tom Davidson, William MacAskill (2026)
• Are AIs more likely to pursue on-episode or beyond-episode reward?
  Anders Cairns Woodruff, Alex Mallen (2026)
• Running list of conjectures about neural networks
  Charles Foster (2023 - current)

Security

• Private Post-Training and Inference for Frontier Models
  Rudolf Laine, Tanya Verma, Daniel McCann-Sayles, Jules Drean (2026)
• My computer got self-hacked because of OpenClaw
  Aaron Zhao, Ilia Shumailov, Cheng Zhang, [...], Zehui Li (2026)
• Boundary Point Jailbreaking of Black-Box LLMs
  Xander Davies, Giorgi Giglemiani, Edmund Lau, [...], Yarin Gal (2026)
• Quantifying Frontier LLM Capabilities for Container Sandbox Escape
  Rahul Marchand, Art O Cathain, Jerome Wynne, [...], Harry Coppock (2026)

AI Economics and Forecasting

• When Does Automating Research Produce Explosive Growth?
  Tom Davidson, Basil Halperin, Thomas Houlden, Anton Korinek (2026)
• Measuring AI R&D Automation
  Alan Chan, Ranay Padarath, Joe Kwon, Hilary Greaves, Markus Anderljung (2026)
• How are AI agents used? Evidence from 177,000 MCP tools
  Merlin Stein (2026)
• We spent 2 hours working in the future
  Thomas Kwa, Tom Cunningham (2026)
• Anthropic Economic Index report: Learning curves
  Maxim Massenkoff, Eva Lyubich, Peter McCrory, [...], Ryan Heller (2026)
• AI's capability improvements haven't come from it getting less affordable
  Anders Cairns Woodruff (2026)
• Final training runs account for a minority of R&D compute spending
  Jean-Stanislas Denain (2026)
  
  Why this is notable
  Catch-up effects exist, but they are less pronounced than I naively expected.
• Computational Arbitrage in AI Model Markets
  Ricardo Olmedo, Bernhard Schölkopf, Moritz Hardt (2026)
• Understanding and tracking developments in robotics
  Janvi (2026)
• Concrete projects to prepare for superintelligence
  Will MacAskill (2026)
• Discussion with Daniel Kokotajlo and Eli Lifland on AI worldviews and cruxes
  Yafah Edelman (2026)

Miscellaneous

• Model-Preserving Adaptive Rounding
  Albert Tseng, Zhaofeng Sun, Christopher De Sa (2025)