7 papers
Training LLMs for Honesty via Confessions
Manas Joglekar, Jeremy Chen, Gabriel Wu +4
Large language models (LLMs) can be dishonest when reporting on their actions and beliefs -- for example, they may overstate their confidence in factual claims or cover up evidence…
Stress Testing Deliberative Alignment for Anti-Scheming Training
Bronson Schoen, Evgenia Nitishinskaya, Mikita Balesni +16
Highly capable AI systems could secretly pursue misaligned goals -- what we call "scheming". Because a scheming AI would deliberately try to hide its misaligned goals and actions,…
BrowseComp: A Simple Yet Challenging Benchmark for Browsing Agents
Jason Wei, Zhiqing Sun, Spencer Papay +7
We present BrowseComp, a simple yet challenging benchmark for measuring the ability for agents to browse the web. BrowseComp comprises 1,266 questions that require persistently nav…
PaperBench: Evaluating AI's Ability to Replicate AI Research
Giulio Starace, Oliver Jaffe, Dane Sherburn +10
We introduce PaperBench, a benchmark evaluating the ability of AI agents to replicate state-of-the-art AI research. Agents must replicate 20 ICML 2024 Spotlight and Oral papers fro…
Trading Inference-Time Compute for Adversarial Robustness
Wojciech Zaremba, Evgenia Nitishinskaya, Boaz Barak +8
We conduct experiments on the impact of increasing inference-time compute in reasoning models (specifically OpenAI o1-preview and o1-mini) on their robustness to adversarial attack…
Deliberative Alignment: Reasoning Enables Safer Language Models
Melody Y. Guan, Manas Joglekar, Eric Wallace +12
As large-scale language models increasingly impact safety-critical domains, ensuring their reliable adherence to well-defined principles remains a fundamental challenge. We introdu…