Central evaluation examples include:
AI Researchers' Views on Automating AI R&D and Intelligence Explosions
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Severin Field
Date:
March 14, 2026
Citations:
0
Frontier Models Can Take Actions at Low Probabilities
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Alex Serrano, Wen Xing
Date:
March 11, 2026
Citations:
0
Can Aha Moments Be Fake? Identifying True and Decorative Thinking Steps in Chain-of-Thought
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jiachen Zhao
Date:
March 10, 2026
Citations:
2
SL5 Standard for AI Security
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Yoav Tzfati
Date:
March 11, 2026
Citations:
0
Emergent Misalignment is Easy, Narrow Misalignment is Hard
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Anna Soligo, Ed Turner
Date:
March 10, 2026
Citations:
0
Haiku to Opus in Just 10 bits: LLMs Unlock Massive Compression Gains
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Roy Rinberg
Date:
March 11, 2026
Citations:
0
Censored LLMs as a Natural Testbed for Secret Knowledge Elicitation
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Helena Casademunt , Bartosz Cywiński, Khoi Tran, Arya Jakkli
Date:
March 10, 2026
Citations:
0
Reasoning Models Struggle to Control their Chains of Thought
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
John Chen, Robert McCarthy, Bruce Lee
Date:
March 11, 2026
Citations:
0
Training Agents to Self-Report Misbehavior
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Bruce Lee, John Chen
Date:
March 11, 2026
Citations:
0
Automatically Finding Reward Model Biases
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Atticus Wang
Date:
March 10, 2026
Citations:
0
Large-scale online deanonymization with LLMs
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Simon Lermen
Date:
March 9, 2026
Citations:
0
The Assistant Axis: Situating and Stabilizing the Default Persona of Language Models
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Christina Lu
Date:
March 10, 2026
Citations:
10
The 2025 AI Agent Index: Documenting Technical and Safety Features of Deployed Agentic AI Systems
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Leon Staufer, Mick Yang
Date:
March 10, 2026
Citations:
0
Eliciting Harmful Capabilities by Fine-Tuning on Safeguarded Outputs
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jackson Kaunismaa
Date:
March 10, 2026
Citations:
4
SynthSAEBench: Evaluating Sparse Autoencoders on Scalable Realistic Synthetic Data
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
David Chanin
Date:
March 10, 2026
Citations:
0
How does information access affect LLM monitors' ability to detect sabotage?
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Raja Moreno, Rohan Subramani
Date:
February 16, 2026
Citations:
0
Prompt Optimization Makes Misalignment Legible
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Caleb Biddulph
Date:
March 11, 2026
Citations:
0
Simple LLM Baselines are Competitive for Model Diffing
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Elias Kempf, Bartosz Cywiński
Date:
March 10, 2026
Citations:
0
Biases in the Blind Spot: Detecting What LLMs Fail to Mention
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Iván Arcuschin Moreno
Date:
March 10, 2026
Citations:
1
What Happens When Superhuman AIs Compete for Control?
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Steven Veld
Date:
March 11, 2026
Citations:
0
Bias-Augmented Consistency Training Reduces Biased Reasoning in Chain-of-Thought
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
James Chua, Edward Rees, Hunar Batra
Date:
March 10, 2026
Citations:
25
AI Futures Model: Timelines & Takeoff
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Brendan Halstead, Alex Kastner
Date:
March 11, 2026
Citations:
0
Neural Chameleons: Language Models Can Learn to Hide Their Thoughts from Unseen Activation Monitors
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Max McGuinness
Date:
March 10, 2026
Citations:
2
Recontextualization Mitigates Specification Gaming without Modifying the Specification
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Ariana Azarbal, Victor Gillioz
Date:
March 10, 2026
Citations:
6
Bloom: an open source tool for automated behavioral evaluations
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Isha Gupta
Date:
March 11, 2026
Citations:
0
Activation Oracles: Training and Evaluating LLMs as General-Purpose Activation Explainers
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Adam Karvonen
Date:
March 10, 2026
Citations:
5
Measuring Chain-of-Thought Monitorability Through Faithfulness and Verbosity
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Austin James Meek, Iván Arcuschin Moreno
Date:
March 10, 2026
Citations:
1
Petri: An open-source auditing tool to accelerate AI safety research
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Isha Gupta
Date:
March 11, 2026
Citations:
0
RE-Bench: Evaluating frontier AI R&D capabilities of language model agents against human experts
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jai Dhyani
Date:
March 10, 2026
Citations:
71
Incentivizing honest performative predictions with proper scoring rules
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Johannes Treutlein, Jeremy Rubinoff (Rubi J. Hudson)
Date:
March 10, 2026
Citations:
11
Probing by Analogy: Decomposing Probes into Activations for Better Interpretability and Inter-Model Generalization
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Patrick Leask
Date:
March 11, 2026
Citations:
0
Real-Time Detection of Hallucinated Entities in Long-Form Generation
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Oscar Balcells Obeso, Andy Arditi, Javier Ferrando Monsonis
Date:
March 10, 2026
Citations:
10
Reasoning Under Pressure: How do Training Incentives Influence Chain-of-Thought Monitorability?
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Qiyao Wei
Date:
March 10, 2026
Citations:
1
Uncertainty-Aware Policy-Preserving Abstractions with Abstention for One-Shot Decisions
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Sandy Tanwisuth
Date:
March 11, 2026
Citations:
0
Inoculation Prompting: Eliciting traits from LLMs during training can suppress them at test-time
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Daniel Tan
Date:
March 10, 2026
Citations:
8
Towards a unified and verified understanding of group-operation networks
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Wilson Wu, Louis Jaburi, Jacob Drori
Date:
March 10, 2026
Citations:
3
Unelicitable Backdoors in Language Models via Cryptographic Transformer Circuits
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Sumeet Motwani, Andis Draguns, Andrew Gritsevskiy
Date:
March 10, 2026
Citations:
5
Simple Mechanistic Explanations for Out-Of-Context Reasoning
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Atticus Wang, Joshua Engels
Date:
March 10, 2026
Citations:
4
Compact Proofs of Model Performance via Mechanistic Interpretability
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Thomas Kwa
Date:
March 10, 2026
Citations:
13
Are Sparse Autoencoders Useful? A Case Study in Sparse Probing
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Joshua Engels
Date:
March 10, 2026
Citations:
65
Tell, don't show: Declarative facts influence how LLMs generalize
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Alexander Meinke
Date:
March 10, 2026
Citations:
10
Latent Adversarial Training Improves Robustness to Persistent Harmful Behaviors in LLMs
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Aidan Ewart, Aengus Lynch, Phillip Guo, Cindy Wu, Vivek Hebbar
Date:
March 10, 2026
Citations:
114
A Causal Model of Theory-of-Mind in AI Agents
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jack Foxabbott
Date:
December 19, 2025
Citations:
0
Towards Understanding Sycophancy in Language Models
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Meg Tong
Date:
March 10, 2026
Citations:
582
Explorations of Self-Repair in Language Models
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Cody Rushing
Date:
March 10, 2026
Citations:
21
Open Character Training: Shaping the Persona of AI Assistants through Constitutional AI
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Sharan Maiya
Date:
March 10, 2026
Citations:
4
Sparse but Wrong: Incorrect L0 Leads to Incorrect Features in Sparse Autoencoders
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
David Chanin
Date:
March 10, 2026
Citations:
5
Linear Representations of Sentiment in Large Language Models
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Oskar John Hollinsworth, Curt Tigges
Date:
March 10, 2026
Citations:
135
Analyzing Probabilistic Methods for Evaluating Agent Capabilities
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Axel Højmark, Govind Pimpale, Arjun Panickssery
Date:
March 10, 2026
Citations:
4
Overcoming Sparsity Artifacts in Crosscoders to Interpret Chat-Tuning
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Caden Juang, Julian Minder, Clément Dumas, Bilal Chughtai
Date:
March 10, 2026
Citations:
10
Adversarial Circuit Evaluation
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Niels uit de Bos
Date:
March 10, 2026
Citations:
2
SolidGoldMagikarp (plus, prompt generation)
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jessica Cooper (Rumbelow), Matthew Watkins
Date:
March 10, 2026
Citations:
18
The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Annah Dombrowski
Date:
March 10, 2026
Citations:
346
Base Models Know How to Reason, Thinking Models Learn When
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Constantin Venhoff, Iván Arcuschin Moreno
Date:
March 10, 2026
Citations:
9
Catastrophic Goodhart: regularizing RLHF with KL divergence does not mitigate heavy-tailed reward misspecification
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Thomas Kwa, Drake Thomas
Date:
March 10, 2026
Citations:
6
Estimating the Empowerment of Language Model Agents
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jinyeop Song
Date:
March 10, 2026
Citations:
1
Taken out of context: On measuring situational awareness in LLMs
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Lukas Berglund, Asa Cooper Stickland , Max Kaufmann, Meg Tong
Date:
March 10, 2026
Citations:
109
Gradient Routing: Masking Gradients to Localize Computation in Neural Networks
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Joseph Miller, Alex Cloud, Jacob Goldman-Wetzler, Evžen Wybitul
Date:
March 10, 2026
Citations:
18
Eliciting Secret Knowledge from Language Models
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Bartosz Cywiński, Emil Ryd, Rowan Wang
Date:
March 10, 2026
Citations:
8
Control Tax: The Price of Keeping AI in Check
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Zhen Ning David Liu, Mikhail Terekhov
Date:
March 10, 2026
Citations:
3
Resisting RL Elicitation of Biosecurity Capabilities: Reasoning Models Exploration Hacking on WMDP
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Joschka Braun, Damon Falck, Yeonwoo Jang
Date:
March 11, 2026
Citations:
0
Large Language Models Often Know When They Are Being Evaluated
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Joseph Needham, Giles Edkins, Govind Pimpale
Date:
March 10, 2026
Citations:
35
Identifying Functionally Important Features with End-to-End Sparse Dictionary Learning
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jordan Taylor
Date:
March 10, 2026
Citations:
58
Planning in a recurrent neural network that plays Sokoban
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Maximilian Li
Date:
March 10, 2026
Citations:
11
Steering Evaluation-Aware Language Models to Act Like They Are Deployed
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Tim Hua, Andrew Qin
Date:
March 10, 2026
Citations:
4
Optimizing AI Agent Attacks With Synthetic Data
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jonathan Kutasov, Chloe Loughridge, Tyler Tracy
Date:
March 10, 2026
Citations:
3
Public Perspectives on AI Governance: A Survey of Working Adults in California, Illinois, and New York
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Claire Short
Date:
January 27, 2026
Citations:
0
Failures to Find Transferable Image Jailbreaks Between Vision-Language Models
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Dan Valentine, James Chua
Date:
March 10, 2026
Citations:
25
AI Sandbagging: Language Models can Strategically Underperform on Evaluations
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Felix Hofstätter, Teun van der Weij
Date:
March 10, 2026
Citations:
72
Using Degeneracy in the Loss Landscape for Mechanistic Interpretability
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Cindy Wu
Date:
March 10, 2026
Citations:
12
Secret Collusion Among Generative AI Agents
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Sumeet Motwani
Date:
March 10, 2026
Citations:
61
Distributed and Decentralised Training: Technical Governance Challenges in a Shifting AI Landscape
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Yashvardhan Sharma, Jakub Kryś
Date:
January 21, 2026
Citations:
0
Feature Hedging: Correlated Features Break Narrow Sparse Autoencoders
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
David Chanin
Date:
March 10, 2026
Citations:
7
Distillation Robustifies Unlearning
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Adeline (Addie) Foote, Alexander Infanger, Eleni Shor, Harish Kamath, Jacob Goldman-Wetzler
Date:
March 10, 2026
Citations:
6
The Attacker Moves Second: Stronger Adaptive Attacks Bypass Defenses Against Llm Jailbreaks and Prompt Injections
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Sander Schulhoff
Date:
March 10, 2026
Citations:
29
Narrow Finetuning Leaves Clearly Readable Traces in Activation Differences
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Clément Dumas, Julian Minder, Helena Casademunt
Date:
March 10, 2026
Citations:
5
BatchTopK Sparse Autoencoders
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Patrick Leask, Bart Bussmann
Date:
March 10, 2026
Citations:
67
Probabilistic Modeling of Latent Agentic Substructures in Deep Neural Networks
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Su Hyeong Lee
Date:
December 14, 2025
Citations:
0
Identifying Sparsely Active Circuits Through Local Loss Landscape Decomposition
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Brianna Chrisman
Date:
March 10, 2026
Citations:
3
Do the Rewards Justify the Means? Measuring Trade-Offs Between Rewards and Ethical Behavior in the MACHIAVELLI Benchmark
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jonathan Ng, Hanlin Zhang
Date:
March 10, 2026
Citations:
178
Auditing language models for hidden objectives
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Florian Dietz, Kei Nishimura-Gasparian, Jeanne Salle
Date:
March 10, 2026
Citations:
33
Towards eliciting latent knowledge from LLMs with mechanistic interpretability
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Bartosz Cywiński, Emil Ryd
Date:
March 10, 2026
Citations:
6
On Defining Neural Averaging
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Su Hyeong Lee
Date:
December 17, 2025
Citations:
0
Transcoders Find Interpretable LLM Feature Circuits
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jacob Dunefsky, Philippe Chlenski
Date:
March 10, 2026
Citations:
108
Too Late to Recall: The Two-Hop Problem in Multimodal Knowledge Retrieval
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Constantin Venhoff
Date:
December 17, 2025
Citations:
0
Convergent Linear Representations of Emergent Misalignment
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Ed Turner, Anna Soligo
Date:
March 10, 2026
Citations:
26
Understanding Reasoning in Thinking Language Models via Steering Vectors
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Constantin Venhoff
Date:
March 10, 2026
Citations:
50
Quantifying stability of non-power-seeking in artificial agents
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Evan Ryan Gunter, Yevgeny Liokumovich
Date:
March 10, 2026
Citations:
2
Understanding and Controlling a Maze-Solving Policy Network
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Ulisse Mini, Peli Grietzer
Date:
March 10, 2026
Citations:
22
Steering Language Models With Activation Engineering
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Lisa Thiergart, David Udell, Ulisse Mini
Date:
March 10, 2026
Citations:
384
Learning Multi-Level Features with Matryoshka Sparse Autoencoders
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Bart Bussmann, Adam Karvonen
Date:
March 10, 2026
Citations:
68
MISR: Measuring Instrumental Self-Reasoning in Frontier Models
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Kai Fronsdal
Date:
March 10, 2026
Citations:
1
Sparse Autoencoders Find Highly Interpretable Features in Language Models
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Hoagy Cunningham
Date:
March 10, 2026
Citations:
911
Interpretability as Compression: Reconsidering SAE Explanations of Neural Activations with MDL-SAEs
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Kola Ayonrinde, Michael Pearce
Date:
March 10, 2026
Citations:
16
Evaluating Sparse Autoencoders on Targeted Concept Erasure Tasks
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Adam Karvonen, Can Rager
Date:
March 10, 2026
Citations:
9
Studying Cross-cluster Modularity in Neural Networks
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Satvik Golechha
Date:
March 10, 2026
Citations:
1
Audit Cards: Contextualizing AI Evaluations
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Leon Staufer, Mick Yang
Date:
March 10, 2026
Citations:
7
Teaching Models to Verbalize Reward Hacking in Chain-of-Thought Reasoning
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Marvin Li, Andy Arditi
Date:
March 10, 2026
Citations:
8
A Benchmark for Scalable Oversight Protocols
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jackson Kaunismaa, Arjun Panickssery
Date:
March 10, 2026
Citations:
1
Language Models Learn to Mislead Humans via RLHF
Lorem ipsum dolor sit amet, consectetur adipiscing elit. Suspendisse varius enim in eros elementum tristique. Duis cursus, mi quis viverra ornare, eros dolor interdum nulla, ut commodo diam libero vitae erat. Aenean faucibus nibh et justo cursus id rutrum lorem imperdiet. Nunc ut sem vitae risus tristique posuere.
Authors:
Jiaxin Wen
Date:
March 10, 2026
Citations:
91