browse by category
browse by tag
research notes
In-depth research writeups, organized by topic. Browse all notes →
all posts
-
Tournament Style RL: A Stable Reward Signal for Problems With No “Right Answer”
TSRL replaces noisy scalar/pairwise rewards with win-rate against a fixed anchor ladder, cutting reward variance by 1/k and beating strong baselines on four non-verifiable RL tasks.
-
A Brief Summary on Positional Encoding Methods
A summary of positional encoding methods in transformers
-
Representation Engineering: A Top-Down Approach to AI Transparency
Exploring Representation Engineering (RepE) as a parallel research paradigm for interpretability and controllability in LLMs.