5 papers
Assign and Add: A Mechanistic Study of Compositional Arithmetic
Brady Exoo, Alberto Bietti, John Sous
Large language models are able to compose skills in order to perform complex tasks, many of which might not have been seen during training. The details of how exactly this composit…
Memory Inception: Latent-Space KV Cache Manipulation for Steering LLMs
Andy Zeyi Liu, Michael Zhang, Ilana Greenberg +3
Steering large language models (LLMs) is usually done by either instruction prompting or activation steering. Prompting often gives strong control, but caches guidance tokens at ev…
Spectral Lens: Activation and Gradient Spectra as Diagnostics of LLM Optimization
Andy Zeyi Liu, Elliot Paquette, John Sous
Training loss and throughput can hide distinct internal representation in language-model training. To examine these hidden mechanics, we use spectral measurements as practical and…
(Im)possibility of Automated Hallucination Detection in Large Language Models
Amin Karbasi, Omar Montasser, John Sous +1
Is automated hallucination detection possible? In this work, we introduce a theoretical framework to analyze the feasibility of automatically detecting hallucinations produced by l…
PHYSICS: Benchmarking Foundation Models on University-Level Physics Problem Solving
Kaiyue Feng, Yilun Zhao, Yixin Liu +4
We introduce PHYSICS, a comprehensive benchmark for university-level physics problem solving. It contains 1297 expert-annotated problems covering six core areas: classical mechanic…