IC-064The TD learning mechanism identified in Llama 3 70B generalizes to Gemma-2-27B and Qwen-2.5-72B across all three tasks

Can Demircan, Tankred Saanum, Akshay Kumar Jagadish, Marcel Binz, Eric Schulz

SourceSparse Autoencoders Reveal Temporal Difference Learning in Large Language Models

The authors replicate the two-step, grid world, and graph learning analyses on Gemma-2-27B and Qwen-2.5-72B. For both models, SAE features show strong correlations with TD-driven value and error signals that outperform alternative models (myopic, transition matrix, surprise). Behavioral patterns are qualitatively similar, though neither model performs the two-step task as well as Llama 3 70B, so RL model fits are less accurate. In the graph task, both models show strong SR correlations and community structures emerging over transformer blocks.

Evidence
correlational
Caveat
Neither model can do the two-step task as well as Llama, so RL models predict behavior less accurately; no interventional (lesioning) experiments were performed on these models.
Model
Gemma 2, Qwen2.5
Concepts
Linear representation, Depth-dependent structure
Methods
Sparse autoencoder / Sparse autoencoders / K-sparse autoencoder / Topk sparse autoencoder / Scaling and Evaluating Sparse Autoencoders / Cunningham et al. 2023 (sparse autoencoders) [primary], Q-learning [compared-to]
Related findings
IC-062, IC-063
Extraction
automatic-extraction