2406.19501
arxiv.org · 7,426 words · saved by 1 readers
N/A
Preprint. M ONITORING L ATENT W ORLD S TATES IN L ANGUAGE M ODELS WITH P ROPOSITIONAL P ROBES Jiahai Feng∗, Stuart Russell & Jacob Steinhardt UC Berkeley A BSTRACT Language models (LMs) are susceptible to bias, sycophancy, backdoors, and other arXiv:2406.19501v2 [cs.CL] 6 Dec 2024…
related reading
- [2406.19501] Monitoring Latent World States in Language Models with Propositional Probesarxiv.org
- Verbalizable Representations Form a Global Workspace in Language Modelstransformer-circuits.pub
- Surfacing Pathological Behaviors in Language Models | Transluce AItransluce.org
- Chapter 1: Transformer Interpretability - ARENAlearn.arena.education
- Structure and Interpretation of Deep Networkssidn.baulab.info
- 2212.03827arxiv.org
- Language Models, World Models, and Human Model-Buildinglingo.csail.mit.edu
- Large Language Model: world models or surface statistics?thegradient.pub
- How well do truth probes generalise? — LessWronglesswrong.com
- [2210.13382] Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Taskarxiv.org
- Features as Rewards: Using Interpretability to Reduce Hallucinationsgoodfire.ai
- Actually, Othello-GPT Has A Linear Emergent World Representation - Neel Nandaneelnanda.io