Frontier lab papers,
read for you, every day.

Reviews

• May 17, 2026

  Agentic Modeling, World Models & Safety Adaptation

  Three Tier 2 frontier-lab papers: Google’s LiSA conservative policy induction for adaptive guardrails, Microsoft Research’s Orchard open-source agentic modeling framework (67.5% on SWE-bench Verified at 30B-A3B size), and NVIDIA’s SANA-WM 2.6B minute-scale world model with hybrid linear attention.

• May 16, 2026

  Multimodal Foundations & Context Optimization

  Three Tier 2 frontier-lab papers plus one Tier 3 OpenAI engineering note: Alibaba’s Qwen-Image-VAE-2.0 high-compression VAE suite (f16c128 beats FLUX.1-dev on text legibility despite 2× compression), ByteDance Seed’s MMProLong 5B-token long-context VLM recipe (32K→128K, generalizes to 512K), Google DeepMind’s BeamSearch-IS context optimizer that makes external tools actually useful, and OpenAI’s Codex mobile relay.

• May 15, 2026

  AI Geopolitics & Codex Sandboxing

  Two Tier 3 frontier-lab posts: Anthropic’s “2028: Two scenarios for global AI leadership” policy essay on US-China compute and distillation policy, and OpenAI’s engineering note on the Windows-native Codex sandbox (CodexSandboxOffline/Online split, DPAPI, firewall, four-layer execution).

• May 14, 2026

  Flow Map Distillation

  NVIDIA + NUS Show Lab introduce AnyFlow (arXiv:2605.13724): the first any-step video diffusion distillation framework based on flow maps, with Flow Map Backward Simulation for on-policy distillation that preserves test-time scaling from 1.3B to 14B.

• May 13, 2026

  Qwen-Image-2.0 Launch

  Alibaba’s Qwen team ships a unified 7B image-generation-and-editing foundation model (arXiv:2605.10730): Qwen3-VL encoder + MMDiT, native 2K, 1K-token prompts, #1 on AI Arena in both T2I and editing.

• May 12, 2026

  RL, Byte-Level & Long Video

  Three Tier 2 frontier-lab papers: Tencent Hunyuan’s Listwise Policy Optimization, Meta FAIR + Stanford + UW’s Fast Byte Latent Transformer, and Google’s A²RD long-video diffusion architecture.

• May 10, 2026

  The April 2026 frontier-model wave

  Cold-start backfill: eight Tier 1 flagship launches in 19 days — DeepSeek-V4, Qwen3.5-Omni, GPT-5.5, Claude Opus 4.7, Kimi K2.6, MiniMax M2.7, Grok 4.3, Mistral Medium 3.5.

About

This repository hosts daily reviews of newly released papers and technical reports from frontier AI labs. Each review is generated by a scheduled agent that scans arXiv, Semantic Scholar, Hugging Face Daily Papers, X, the Emergent Mind newsletter, and other sources, then writes a structured summary covering methodology, evaluation, and results.

Only papers from a curated list of frontier labs are included. Reviews are deduplicated against prior entries — each paper appears at most once.