AI Archives | Page 2 Of 3 | Free Downloads

FastMix performance comparison chart showing 550x speedup over RegMix

FastMix: How Tencent Finds Optimal Data Mixture via Gradient Descent

By Hunter Io June 25, 2026AI & Machine Learning

FastMix automates data mixture discovery using gradient descent on a single proxy model, achieving 550x speedup over RegMix for LLM training optimization.

AI & Machine Learning

SproutRAG: Attention-Guided Tree Search with Progressive Embeddings — Technical Deep Dive

By Hunter Io June 24, 2026AI & Machine Learning

SproutRAG is an attention-guided hierarchical RAG framework that builds a binary tree over sentence-level chunks using learned inter-sentence attention, achieving 6.1% higher IE with 2.7x fewer tokens than GraphRAG.

AI & Machine Learning Open Source

Sumi-7B: The First Open Diffusion Language Model Trained From Scratch

By Hunter Io June 24, 2026AI & Machine Learning, Open Source

Sumi-7B is the first uniform diffusion language model trained from scratch at 7B scale. This fully open model by Tohoku University can revise any token at any step — a paradigm shift from autoregressive generation.

GateMem benchmark overview diagram showing multi-principal shared memory governance framework

AI & Machine Learning

GateMem: The First Benchmark for Memory Governance in Multi-Principal AI Agents

By Hunter Io June 23, 2026AI & Machine Learning

GateMem reveals a hard truth: today's AI agents can remember, but they can't govern shared memory. A deep dive into the first benchmark for multi-principal memory governance.

Multi-LCB benchmark extending LiveCodeBench to 12 programming languages

Multi-LCB: Extending LiveCodeBench to 12 Languages — Key Findings

By Hunter Io June 22, 2026AI

Multi-LCB extends LiveCodeBench to 12 programming languages. This ICLR 2026 benchmark evaluates 24 LLMs and reveals Python overfitting, language-specific contamination, and significant cross-language performance gaps in code generation.

LoopCoder-v2 Parallel Loop Transformer architecture overview diagram

LoopCoder-v2: Only Loop Once for Efficient Test-Time Computation Scaling

By Hunter Io June 19, 2026AI

LoopCoder-v2 explores Parallel Loop Transformers for test-time computation scaling, showing that R=2 achieves 64.4 on SWE-bench Verified, surpassing models 30x larger via an optimal gain-cost loop count.

SubQ 1.1 Small Review: Subquadratic Sparse Attention Model Delivers 64x Faster Inference

News AI

FastMix: How Tencent Finds Optimal Data Mixture via Gradient Descent

SproutRAG: Attention-Guided Tree Search with Progressive Embeddings — Technical Deep Dive

Sumi-7B: The First Open Diffusion Language Model Trained From Scratch

GateMem: The First Benchmark for Memory Governance in Multi-Principal AI Agents

Multi-LCB: Extending LiveCodeBench to 12 Languages — Key Findings

LoopCoder-v2: Only Loop Once for Efficient Test-Time Computation Scaling

SubQ 1.1 Small Review: Subquadratic Sparse Attention Model Delivers 64x Faster Inference

Nex-AGI Drops Nex-N2: Open-Source Agent Model That Keeps Pace with GPT-5.5

GBrain: The Brain Layer Your AI Agent Has Been Missing — Built by YC’s CEO