- llm
- nlp
- reasoning
- rl
- rlhf
- ssm
- theory
- information-theory
- representation-learning
- optimization
- rag
- alignment
- paper-review
- survey
- notes
•
•
•
•
•
•
•
•
•
•
•
•
•
•
-
SOAR — Teaching Models to Teach Themselves: Reasoning at the Edge of Learnability
Paper review of SOAR (Sundaram et al., 2026) — a bilevel meta-RL recipe where a teacher LLM is rewarded for the student LLM's measured progress on hard problems, unlocking learning on math benchmarks where the base model scores 0/128.
-
Learning Fast and Slow: A New Recipe for Adapting LLMs Without Forgetting
Trend note on the GEPA team's "Learning, Fast and Slow" — Fast-Slow Training (FST) interleaves slow RL weight updates with fast in-context prompt evolution via GEPA, hitting RL's ceiling with up to 3x fewer samples and staying close enough to the base model to keep learning new tasks afterward.
-
Parcae: Scaling Laws for Stable Looped Language Models
Beginner-friendly review of Parcae (Prairie et al., 2026) — a stable recipe for *looped* transformers that lets a 770M model match a 1.3B fixed-depth Transformer by running the same layers multiple times, derived from a residual-stream stability analysis.
-
Meta-Harness: End-to-End Optimization of Model Harnesses
Paper review of Meta-Harness (Lee et al., 2026) — an outer-loop search over the *code around the LLM*, driven by a coding-agent proposer that reads prior candidates and their raw execution traces through the filesystem.
-
Mamba: Linear-Time Sequence Modeling with Selective State Spaces
Selective state space model that matches Transformer quality with linear-time inference scaling.