PuSH - Publication Server of Helmholtz Zentrum München: Sparse autoencoders reveal temporal difference learning in large language models.

Information

Clues to quality of journals

Open Access Policy of Helmholtz Association 2016

CC Licencences

Publication Metrics

Navigation

Home

Deutsch

Research

Advanced Search

Browse by ...

... HMGU-Authors/Consortia

... Organizational Structure

... Journal

... Publication Type

... Research Data

... Arbeitsgruppen

... Publication Year

Publication overview

Statistics

Statistics (last 5 years)

OA Publications

Publish

Submit Publication

Import publication from...

...EVA

Report missing publication

Highlights

Support & Contact

Contact persons

Help

Data protection

Helmholtz Open Science

JANE Journal Estimator

SHERPA/RoMEO

DOAJ

Export:

Text

Endnote (RIS) BIB

BibTeX

Demircan, C. ; Saanum, T.* ; Jagadish, A.K. ; Binz, M. ; Schulz, E.

Sparse autoencoders reveal temporal difference learning in large language models.

In: (13th International Conference on Learning Representations Iclr 2025, 24 - 28 April 2025, Singapur). 2025. 4972-4997 (13th International Conference on Learning Representations Iclr 2025)

Publ. Version/Full Text

Abstract
Metrics
Extra information

In-context learning, the ability to adapt based on a few examples in the input prompt, is a ubiquitous feature of large language models (LLMs). However, as LLMs' in-context learning abilities continue to improve, understanding this phenomenon mechanistically becomes increasingly important. In particular, it is not well-understood how LLMs learn to solve specific classes of problems, such as reinforcement learning (RL) problems, in-context. Through three different tasks, we first show that Llama 3 70B can solve simple RL problems in-context. We then analyze the residual stream of Llama using Sparse Autoencoders (SAEs) and find representations that closely match temporal difference (TD) errors. Notably, these representations emerge despite the model only being trained to predict the next token. We verify that these representations are indeed causally involved in the computation of TD errors and Q-values by performing carefully designed interventions on them. Taken together, our work establishes a methodology for studying and manipulating in-context learning with SAEs, paving the way for a more mechanistic understanding.