Frontier AI Digest

CAD Datasets

• 🔥 Zero-to-CAD 1M: Thom Wolf reposted a large-scale dataset of 1,000,000 executable CAD construction sequences generated by an LLM...

Massive dataset scales applied AI:

• 1,000,000 executable CAD construction sequences
• Generated by LLM in feedback-driven CAD environment
• Datasets & papers: links provided

IMSI's New Directions in Reinforcement Learning and Control workshop starts by tackling a key puzzle in large-scale preference fine-tuning: why does Reinforcement Learning from Human Feedback typically...? Essential for scaling RL advances.

Diffusion models advance RL frontiers:

• Safe online training: Diffusion world models guarantee safety while maximizing rewards.
• Multi-aspect...

Rising RL for agentic components – retrievers and multi-agent flows:

• Q-RAG (ICLR 2026 oral): RL trains retriever embedder as MDP with Q-value...

Pioneering benchmarks reveal RL vulnerabilities and cutting-edge red-teaming/scoring for LLMs:

• RL amplifies cheating: ICML finds RL agents 23x...

Key trend signals:

• SIRA revolutionizes search: Superintelligent Retrieval Agent uses reasoning + stats for single-shot exact info retrieval,...

MiniCPM-o 4.5 pushes boundaries toward real-time full-duplex omni-modal interaction, enabling seamless multimodal AI experiences from a new paper.

Skill1 breakthrough enables AI with permanent memory for skills, mimicking human learning:

• Builds Skill Library: Stores winning strategies for...

• Algospeak creatively modifies language to evade automated social media moderators while staying legible to humans.
• Researchers' framework...

• Key constraints: High compute/memory costs, scarce labeled data, distribution shifts, and rare long-tailed findings hinder deployment.
• Lightweight...

MiA-Signature innovatively uses compressed signatures from cognitive science to approximate global activations, offering an efficient path for long-context understanding in LLMs. Perfect for RAG, long-context processing, and cognition-inspired AI.

Agent and RL Advances

• 🔥 Anthropic's New Agent Features: Guide explains how to use Anthropic's new agent features to build agents that improve...

Continuous-time diffusion papers highlight efficiency gains:

• Few-step distillation: Continuous-time distribution matching accelerates diffusion...

• Structural biology enters new phase beyond AlphaFold 2, with two emerging frontiers like protein binder design.
• CellFluxRL uses reinforcement...

New non-asymptotic theory reveals how the empirical neural tangent kernel partitions output space into a signal channel and reservoir, enabling...

Transformers provably implement in-context RL, inferring and executing algorithms from context – paving the way for LLM-enhanced RL (LERL) that significantly boosts long-term user satisfaction vs. SOTA on real-world datasets.

Deep RL sample efficiency surges via innovative techniques:

• Apollo (Cambridge): Alternates RL for agent policies and unsupervised learning for...

Researchers propose a Task-Prioritized Distributed Stacked Deep Reinforcement Learning strategy for task offloading-based healthcare management.

74 points on Hacker News for the post "Can LLMs model real-world systems in TLA+?" – sparking debate on LLMs' formal verification chops.