AI research and inference

Must read

• Axiom Math achieved the first complete computer-checked proof of the Classification of Finite Simple Groups using the Lean proof assistant, utilizing a swarm of dozens of AI agents to manage 108 milestones and 43.1 million lines of code. The process identified 34 deviations from the published record and produced a new mathematical result, Theorem 2.3 ("Preservation of simple tops"), to resolve a gap in the Unbalanced Group Theorem [21].
• Google DeepMind announced Genie 4, a natively multimodal world model that integrates Gemini models to maintain environmental coherence and complex reasoning during real-time interactive simulations. The model uses Gemini Omni for appearance, Gemini 3.8 Flash as a guiding agent to maintain causality, and Gemini 4 Argon to break down user-instructed tasks into interleaved text captions [5].
• Research presented at a NeurIPS 2026 workshop challenges the effectiveness of self-distillation for LLM reasoning, arguing that reported gains often stem from overfitting or non-privileged information rather than true self-improvement. The study found that methods like SDPO induce overfitting on out-of-distribution tasks, while OPSD appears to transfer "thinking-mode behavior" rather than utilizing privileged information [2].
• Update: NVIDIA's OpenShell 0.1.0 provides an open-source runtime to secure AI agents through a three-component architecture that enforces sandboxed execution and formal policies from a trusted supervisor. The system is designed to prevent the credential exposure and observability gaps seen in recent high-profile agent breaches by moving restrictions to the kernel/supervisor level [29]. (published Oct 21) (follows a Sep 29 post)
• Update: DeepSeek-V4.1-Flash-1017, a sparse Mixture-of-Experts model update, achieved significant performance jumps in agentic benchmarks, including an 11-point increase on Terminal-Bench 4.0. The update utilizes a Causal Encoder-Decoder architecture and reinforcement learning with self-feedback to improve end-to-end project development [28]. (published Oct 18) (follows a Oct 17 post)
• SenseNova 6.8 Code2Video is a lightweight open model that generates executable code to define motion, lighting, and composition for video creation instead of using black-box sampling. This approach allows for precise control over elements like camera paths and text, supporting long-horizon sessions of over 1,000 turns [7].

FYI

• OpenBMB released MiniCPM-OCR 1.0 (1.3B) and MiniCPM-OCR Jev (2B), the latter acting as an AI judge to detect errors in formulas and tables with 0.936 AUROC [13].
• LiquidAI released the LFM3 model family, featuring a tri-hybrid architecture of convolutions and attention layers designed for edge devices, including text, vision-language, and omni-modal versions [17].
• An analysis of Diffusion Gemma on Raspberry Pi 5 clusters suggests diffusion LLMs are a poor fit for distributed inference over slow networks due to high synchronization data requirements compared to autoregressive models [6].
• Bria's Fibo 2.0 is an 8B image generation model that uses a structured JSON-based "Visual GenAI Language" instead of text prompts to enable precise, localized edits [30]. (published Oct 23)
• Researchers developed "Minions," a multi-agent system using DeepSeek-V4.1-Flash to orchestrate four Qwen3.5-9B models for binary reverse engineering [14].
• Update: NVIDIA RTX Spark updates include a DLSS 5 model that runs 2x faster than its September version and the Nemotron 3.5 Bolt E4B SLM for on-device AI characters [4]. (follows a Oct 14 post)