Top story

TOP STORY Pretraining Recurrent Networks without Recurrence. A novel pretraining approach enables recurrent architectures to train without sequential recurrence, potentially unlocking faster, more parallelizable training for RNN-class models. arxiv

Research

You Only Index Once: Cross-Layer Sparse Attention with Shared Routing. Proposes shared routing across transformer layers to reduce indexing overhead in sparse attention mechanisms. arxiv

Vortex: Efficient and Programmable Sparse Attention Serving for AI Agents. A serving system that brings efficient, programmable sparse attention specifically to AI agent workloads. arxiv

Agent Memory: Characterization and System Implications of Stateful Long-Horizon Workloads. Characterizes memory access patterns and system-level implications of stateful, long-horizon agent tasks. arxiv

Will the Agent Recuse Itself? Measuring LLM-Agent Compliance with In-Band Access-Deny Signals. Evaluates whether LLM agents appropriately honor in-band access-denial signals as a security and safety measure. arxiv

Tools

Lowfat, Pluggable CLI Filter for LLM Tokens. A CLI tool that filters and compresses context before sending to LLMs, claiming a 91.8% reduction in token usage. GitHub

Ollama. Local inference tool now supports several new frontier models including Kimi-K2.6, GLM-5.1, MiniMax, and DeepSeek. GitHub

Industry

OpenAI, GPT-5.5 Now in Wider Deployment. OpenAI's GPT-5.5 frontier model has rolled out to a broader user base.

OpenAI, New "Dreaming" Memory Architecture for ChatGPT. ChatGPT reportedly received a new memory architecture called "Dreaming," signaling a significant personalization shift.

OpenAI, Frontier Models and Codex Coming to AWS / Amazon Bedrock. OpenAI's frontier models and Codex are becoming available through AWS and Amazon Bedrock.

Community

Goedel-Architect: Streamlining Formal Theorem Proving with Blueprint Generation. Uses blueprint generation and iterative refinement to streamline LLM-assisted formal theorem proving. arxiv

HANDOFF: Humanoid Whole-Body Control via Distilled Complementary Teachers. A distillation-based method for whole-body humanoid robot control using complementary teacher policies. arxiv

Reinforcement Learning Elicits Contextual Learning of Unseen Language Translation. Demonstrates that reinforcement learning can surface contextual translation abilities for unseen languages in LLMs. arxiv