Research Reading
RL methods for language models: what the RLxx names refer to and which results hold up
Do a survey of the state of the art in various RL methods being experimented to improve language models. RLHF and RLVR are known, but Jev uses RLCD and in the process I started seeing a variety of RLxx algorithms. They seem to describe the shape of the data that the reward is being derived off. What are people up to and what has been working?
2026-09-20 · 106 papers · 578 links
How Jev answers many questions in one pass
How does Jev, TypeSafe AI's "System One" model, answer many questions in parallel, and what is the most likely architecture behind it?
2026-09-19 · 127 papers · 579 links
Hardware for single-user (batch size 1) LLM inference on a handheld
Model inference but for a single user, on a local machine. You lose batching at scale and the property that all MoE experts activate at large batch size. For privacy. Find the exact nature of computation for a single thread of inference and spec out what the ideal dedicated hardware would look like. Goal: run 100B-500B models on a handheld.
2026-09-04 · 68 papers · 357 links
Looped transformers and latent chain of thought
OpenAI claims to be running Astra as a Looped Transformer with chain of thought replaced from decode to english to embedding representation. I'd like to learn and play around with this in small models. Find all modern reputable research that take a stab at this and prepare my report.
2026-09-02 · 58 papers · 295 links
Model size vs. reasoning tokens: is there a conversion rate?
You can increase the amount an LLM thinks in two axes: by increasing the size of the model, or by increasing the number of tokens it outputs. Can we quantify the mathematical equivalence between the two? Given an architecture (depth, embed dim, param count), what's the equation that establishes the relationship between model size and reasoning effort (test-time compute)?
2026-08-23 · 55 papers · 263 links
Structure of the residual-mixing matrices in mHC
What kind of structures emerge in the linear transformations added by Manifold Constrained Hyper Connections (mHC) used in modern language models to solve the information bottleneck problem in the residual stream? Studied from geometric, eigenvalue, and other perspectives.
2026-06-24 · 33 papers · 149 links
Why did language models need BPE tokenization?
Byte-pair encoding (BPE) tokenization is widely credited with making language models trainable. Why? Tokenization is a form of compression, so why couldn't character-level models reach the same quality? Are character-level models provably able to learn the same things given more training (just less efficiently), or does subword tokenization give models a capability they would otherwise lack?
2026-06-23 · 44 papers · 182 links