alphaXiv

@askalphaxiv

High fidelity research

Joined November 2023

Pinned
alphaXiv
@askalphaxiv
May 12
Reinforcing Recursive Language Models Can a 4B model learn to recursively call itself to answer hard long-context questions? We RL fine-tuned a small model to behave as a native RLM. On evidence selection across scientific papers, our 4B RLM matches Sonnet 4.6 in quality
alphaXiv
@askalphaxiv
4h
"Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding" Across 1,867 repos, CLAUDE.md-style files grow +226% because adding instructions is cheap, but deleting them gets riskier as their rationale is forgotten. The older an instruction gets, the less
alphaXiv
@askalphaxiv
7h
"Dion3: Full-Stack Orthogonal Updates" Dion3 makes Muon practical at scale. Instead of running Newton-Schulz on the full momentum matrix, it orthogonalizes a selected row subset, reformulates the iteration over the smaller Gram matrix, exploits symmetry with custom CuteDSL
alphaXiv
@askalphaxiv
15h
"Model Discovery Agent: LLM-assisted Bayesian experiment design for data-efficient discovery of mechanistic world models" This paper's Model Discovery Agent combines LLM hypothesis generation with Bayesian experimental design for mechanistic world-model discovery. How it works
alphaXiv
@askalphaxiv
17h
"From RLVR to RLSVR: Task Transformation Induces Self-Verifiable Rewards for Open-Ended LLM Self-Improvement" RLSVR extends RLVR beyond math and code by making rewards verifiable through task transformation. The researchers used SpyRL to instantiate this with