Dec 1, 2025

• A New AI Winter Is Coming

  • LLMs are fundamentally flawed, as they generate plausible but often incorrect outputs, leading to a high failure rate in practical applications, with estimates suggesting 60% to 95% of results may be erroneous.

• Program-of-Thought Prompting Outperforms Chain-of-Thought by 15% (2022)

  • Program of Thoughts (PoT) enhances language models by separating reasoning from computation, using Codex to articulate reasoning as a program and an external computer for calculations.

• I Tested the M5 iPad Pro's Neural-Accelerated AI, and the Hype Is Real

  • The M5 iPad Pro demonstrates a 4.4× improvement in time to first token (TTFT) for local AI processing, significantly exceeding Apple's claims, particularly for long prompts of 10,000 and 16,000 tokens.

• Every Token Counts: Generalizing 16M Ultra-Long Context in Large Language Models

  • This research introduces Hierarchical Sparse Attention (HSA), a novel mechanism that enables efficient ultra-long context modeling by ensuring sparsity, random-access flexibility, and length generalization.

• 🤗Transformers v5: Simple model definitions powering the AI ecosystem

  • Transformers v5 introduces a modular design that simplifies model integration, allowing for faster contributions and a cleaner codebase, enhancing the overall user experience in the AI ecosystem.

• At NeurIPS, NVIDIA Advances Open Model Development for Digital and Physical AI

  • NVIDIA introduces the DRIVE Alpamayo-R1, the first open industry-scale reasoning vision language action model for autonomous driving, enhancing vehicle decision-making in complex scenarios through advanced AI reasoning techniques.

• Pose-free 3D Gaussian splatting via shape-ray estimation

  • SHARE introduces a pose-free framework for 3D Gaussian splatting, enabling efficient rendering without the need for precise camera poses, thus addressing common challenges in real-world applications.

• LFM2 Technical Report

  • LFM2 is a family of Liquid Foundation Models optimized for on-device deployment, achieving up to 2x faster prefill and decode on CPUs through a compact hybrid backbone of gated short convolutions and grouped query attention blocks.

• Outcome-based learning vs vector search: 100% vs 3.3% accuracy on adversarial queries (p=0.001)

  • Outcome-based learning achieved a remarkable 100% accuracy on adversarial queries, significantly outperforming vector search's 3.3% accuracy, highlighting the potential of integrating outcome effectiveness into retrieval systems.

• Multi-chain Graph Refinement and Selection for Reliable Reasoning in Large Language Models

  • The Multi-chain Graph Refinement & Selection (MGRS) framework enhances reasoning in Large Language Models (LLMs) by generating diverse trajectories and refining responses through a composite verification strategy, addressing limitations of existing methods like Tree-of-Thought and Graph-of-Thought.

• TIM-PRM: Verifying multimodal reasoning with Tool-Integrated PRM

  • TIM-PRM revolutionizes verification in Multimodal Large Language Models (MLLMs) by transforming it into an active, tool-augmented investigation, addressing vulnerabilities like visual hallucinations and logical inconsistencies that traditional methods fail to resolve.

• ORION: Teaching Language Models to Reason Efficiently in the Language of Thought

  • ORION introduces a framework that leverages the Language of Thought Hypothesis to train models for ultra-compressed reasoning, significantly enhancing efficiency in complex problem-solving.

• Efficient MoE Pre-training at Scale on 1K AMD GPUs with TorchTitan

  • Efficient training of Mixture-of-Experts (MoE) models like DeepSeek-V3 and Llama 4-Scout was achieved using TorchTitan and Primus-Turbo on 1,024 AMD MI325X GPUs, demonstrating a 2.77× speed-up and 96% scaling efficiency.