Research
AISI works on AI safety research with organizations throughout the broader AI safety community, alongside research conducted at Georgia Tech. This ranges from technical approaches to policy proposals, published at conferences such as ICML, ICLR, CoLM, and more.
Select Publications by AISI members
AuditBench: Evaluating Alignment Auditing Techniques on Models with Hidden Behaviors
Anthropic Science Blog, 2026
Abhay Sheshadri
An Independent Safety Evaluation of Kimi K2.5
Preprint, 2026
Parv Mahajan
Alignment Faking Revisited: Improved Classifiers and Extensions
Anthropic Science Blog, 2025
Abhay Sheshadri
Patterns and Mechanisms of Contrastive Activation Engineering
ICLR Workshop, 2025
Yixiong Hao, Ayush Panda, Stepan Shabalin
Scaling Sparse Feature Circuits For Studying In-Context Learning
ICLR Workshop, 2025
Stepan Shabalin
Interpreting Large Text-to-Image Diffusion Models with Dictionary Learning
CVPR Workshop, 2025
Stepan Shabalin, Ayush Panda, Yixiong Hao
Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs
arXiv preprint, 2024
Abhay Sheshadri
Robust Unlearning via Mechanistic Localizations
ICLR Workshop Spotlight, 2024
Abhay Sheshadri
A Mechanistic Analysis of a Transformer Trained on a Symbolic Multi-Step Reasoning Task
ACL Main Conference, 2024
Abhay Sheshadri
All Publications
A full publication list can be found on our research tracker.
Refusal Boundaries in Language Models Overindex on Superficial Linguistic Cues
NeurIPS 2026, Workshop on Foundations of Language Model Security (FLMSec)
AISI coauthors: Meru Gopalan, Alec Harris
Comparing Sparsified MLPs and Transcoders: Many Transcoder Features Are Neurons in Disguise
NeurIPS 2026, Workshop on Interpretability as a Science
AISI coauthors: Stepan Shabalin
The AI Sedition Acts: Sabotage-by-Omission in LLM Briefing Pipelines
NeurIPS 2026, Workshop on Evaluation of Interactive Agents (IAEval) and Workshop on Agents in the Wild
AISI coauthors: Glenn Matlin
Evaluation Cards: An Interpretive Layer for AI Evaluation Reporting
NeurIPS 2026, Evaluations & Datasets Track (Spotlight)
AISI coauthors: Yixiong Hao, Drishti Sharma, Ishan Khire, Amit Saha
Editing the Goal: Per-Token SAE Steering and Its Depth Limits in π0.5
CoRL 2026, Science of Physical AI Safety (SPAIS) Workshop
AISI coauthors: Rocio Perales Valdes
CoTs Remain Monitorable When Being Unfaithful Requires Computation: Replicating and Extending the Necessity Effect Across Eleven Models
arXiv preprint 2026
AISI coauthors: Yixiong Hao
When Goals Leak Across Relationships: How Cached Goal Representations Shape Cross-Client Decisions
arXiv preprint 2026
AISI coauthors: Glenn Matlin
Prioritization of Risks from Artificial Intelligence: A Delphi Study of 272 International Experts
arXiv preprint 2026, under review at Nature Machine Intelligence
AISI coauthors: Yixiong Hao, Parv Mahajan
Capability Provenance in Language Models: A Case Study in Social Reasoning
COLM 2026, Main Conference; Science of Foundation Models Workshop; Workshop on Social Simulation with LLMs
AISI coauthors: Glenn Matlin
Evaluation-Conditioned Training: Teaching Models to Generalize to Stronger Oversight Regimes
COLM 2026, WAB Workshop
AISI coauthors: Alec Harris, Yixiong Hao
Shall We Play a Game? Language Models for Open-ended Wargames
COLM 2026, Workshop on Social Simulation with LLMs; EMNLP 2025, Wordplay Workshop
AISI coauthors: Glenn Matlin, Yixiong Hao, Parv Mahajan
Role Steering of Language Models for Social Simulations
COLM 2026, Workshop on Social Simulation with LLMs
AISI coauthors: Glenn Matlin
No One Wins in Nuclear War: Social Simulations of High-stakes Military Decision-making
COLM 2026, Workshop on Social Simulation with LLMs
AISI coauthors: Glenn Matlin
Position: AI Is Not Ready for Strategic Conflicts
COLM 2026, Workshop on Social Simulation with LLMs
AISI coauthors: Glenn Matlin
Towards understanding LLM behavior through fine-tuning
ICML 2026, Workshop on Foundations of Deep Generative Models: Understanding Memorization, Generalization, and Reasoning
AISI coauthors: Harshul Basava, Keshav Shenoy
Auditing LLMs for Hidden Behaviors using Model Diffing
ICML 2026, Trustworthy AI for Good (AI4GOOD) Workshop
AISI coauthors: Keshav Shenoy
Introspection Adapters: Training LLMs to Report Their Learned Behaviors
Anthropic Alignment Science Blog 2026
AISI coauthors: Keshav Shenoy, Abhay Sheshadri
Open Source Replication of the Auditing Game Model Organism
Anthropic Alignment Science Blog 2025
AISI coauthors: Abhay Sheshadri
Why Do Some Language Models Fake Alignment While Others Don't?
NeurIPS 2025, Main Conference (Spotlight)
AISI coauthors: Abhay Sheshadri
Transcoders Beat Sparse Autoencoders for Interpretability
arXiv preprint 2025
AISI coauthors: Stepan Shabalin
Obfuscated Activations Bypass LLM Latent-Space Defenses
arXiv preprint 2024
AISI coauthors: Abhay Sheshadri
Eliciting Language Model Behaviors using Reverse Language Models
NeurIPS 2023, Workshop on Socially Responsible Language Modeling Research (Spotlight)
AISI coauthors: Abhay Sheshadri, Ayush Panda
