Papers

What we've read, and what we've considered. The graph shows papers we've actually discussed in a session; an edge connects two papers that share at least one tag, and gets heavier the more tags they share.

Read

Papers we've discussed in a session.

• A Geometric Calculator Inside a Neural Network
  Goodfire · 2025
  mech-interpmanifoldsinterpretability

  Discussed in: BASIS 05

• A Global Workspace in Language Models
  Anthropic · 2026
  mech-interpinterpretabilitymonitoringcognition

  Discussed in: BASIS 08 , BASIS 07

• Auditing Language Models for Hidden Objectives
  Anthropic · 2025 · arXiv:2503.10965
  auditingdeceptionalignmentmech-interp

  Discussed in: BASIS 03

• Emergent Misalignment
  Betley et al. · 2025
  alignmentfine-tuningmisalignment

  Discussed in: BASIS 01

• Model Forensics: Investigating Whether Concerning Behavior Reflects Misalignment
  Singh, Kroiz, Rajamanoharan, Nanda · 2026 · arXiv:2606.26071
  misalignmentdeceptionchain-of-thoughtevalsagentsauditing

  Discussed in: BASIS 09

• Steering Along Manifolds to Control Neural Networks
  Wurgaft, Goodman, Rager, Fel, Kowal, Geiger, Shyam, Lubana, Feucht, Bhalla, Haklay, Bigelow, Sarfati, McGrath, Lewis, Merullo · 2025
  mech-interpinterpretabilitymanifoldssteering

  Discussed in: BASIS 06

• The Assistant Axis
  Anthropic · 2025
  assistantpersonaalignmentmech-interp

  Discussed in: BASIS 02

• Towards Guaranteed Safe AI: A Framework for Ensuring Robust and Reliable AI Systems
  2024 · arXiv:2405.06624
  safe-by-designalignment

  Discussed in: BASIS 04

Considered

Papers that the group surfaced, skimmed, and voted on, but ultimately did not dedicate a full session to.

• Agentic Misalignment: How LLMs Could Be Insider Threats
  Anthropic · 2025 · arXiv:2510.05179
• AI 2027
  Kokotajlo et al. · 2025
• AI Control: Improving Safety Despite Intentional Subversion
  Greenblatt et al. · 2024
• Can SAEs Capture Neural Geometry?
  Bhalla, Geiger, Fel, Lubana, Rager, Feucht, Haklay, Wurgaft, Boppana, Kowal, Shyam, Lewis, McGrath, Merullo · 2026
• Claudini: Autoresearch Discovers State-of-the-Art Adversarial Attack Algorithms for LLMs
  2026 · arXiv:2603.24511
• Constitutional Classifiers: Defending against Universal Jailbreaks across Thousands of Hours of Red Teaming
  Anthropic · 2025 · arXiv:2501.18837
• Ctrl-Z: Controlling AI Agents via Resampling
  2025 · arXiv:2504.10374
• Emergent Introspective Awareness in Large Language Models
  Anthropic · 2025
• Evaluating Frontier Models for Dangerous Capabilities
  Phuong et al. · 2024 · arXiv:2403.13793
• Language Models Transmit Behavioural Traits Through Hidden Signals in Data
  2026
• MONA: Myopic Optimization with Non-myopic Approval Can Mitigate Multi-step Reward Hacking
  2025 · arXiv:2501.13011
• Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitors
  2025 · arXiv:2512.11949
• On the Biology of a Large Language Model
  Anthropic · 2025
• Reasoning Models Don't Always Say What They Think
  Anthropic · 2025 · arXiv:2505.05410
• Steering Evaluation-Aware Language Models to Act Like They Are Deployed
  2025 · arXiv:2510.20487
• Values in the Wild
  Anthropic · 2025
• You Are What You Eat (SLT)
  2025 · arXiv:2502.05475