Best of Deep Learning — December 2025

1. 1
   Article

   Sebastian Raschka·21w

   The State Of LLMs 2025: Progress, Problems, and Predictions

   A comprehensive 2025 review of large language model developments highlights reinforcement learning with verifiable rewards (RLVR) and the GRPO algorithm as the year's dominant training paradigm, following DeepSeek R1's breakthrough. Key trends include inference-time scaling, tool use integration, and architectural efficiency tweaks like mixture-of-experts and linear attention mechanisms. The analysis addresses benchmarking challenges ("benchmaxxing"), discusses practical LLM usage for coding and writing, and examines the shift toward domain-specific models with proprietary data. Predictions for 2026 emphasize RLVR expansion beyond math/code, increased inference optimization, and the emergence of diffusion models for low-latency tasks.

   176

   1
2. 2
   Article

   MIT News·23w

   Deep-learning model predicts how fruit flies form, cell by cell

   MIT researchers developed a deep-learning model that predicts cell-by-cell development during fruit fly embryo formation with 90% accuracy. The model uses a dual-graph structure representing cells as both point clouds and foam-like bubbles, tracking properties like position, division, and folding minute-by-minute during gastrulation. The approach could eventually predict development in more complex organisms and identify early disease patterns in conditions like asthma and cancer, though high-quality video data remains the primary limitation for broader applications.

   90

   7
3. 3
   Video

   bycloud·24w

   how this tiny model beat ChatGPT on the “AGI” benchmark [HRM & TRM]

   Two novel AI models, HRM (27M parameters) and TRM (7M parameters), challenge the scaling paradigm by outperforming large language models like GPT-4 on the ARC AGI benchmark through recursive reasoning. Instead of processing everything in one pass, these tiny models iteratively refine answers using dual-network architectures with fast and slow update cycles. TRM achieves 40% on ARC AGI with just 7 million parameters by training on actual loop behavior rather than assumed equilibrium states. Empirical results show that smaller models with more recursion outperform larger models with more layers, suggesting that for constrained logical tasks, iterative refinement beats raw parameter scaling.

   39
4. 4
   Article

   Daily Dose of Data Science | Avi Chawla | Substack·22w

   [Hands-on] Deploy and Run LLMs on your Phone!

   Fine-tune and deploy LLMs directly on iOS and Android devices using UnslothAI, TorchAO, and ExecuTorch. The tutorial walks through loading Qwen3-0.6B, preparing reasoning and chat datasets, training with quantization-aware methods, exporting to mobile-ready .pte format, and running the model locally on iPhone at ~25 tokens/second. The resulting model is ~470MB and runs 100% on-device without requiring cloud connectivity.

   37
5. 5
   Article

   Sebastian Raschka·24w

   From Random Forests to RLVR: A Short History of ML/AI Hello Worlds

   A chronological overview traces the evolution of beginner-friendly ML/AI examples from 2013 to 2025. Starting with Random Forests on Iris datasets and XGBoost on Kaggle competitions, it progresses through neural networks (MLPs, AlexNet), transformer models (DistilBERT, Llama 2 with LoRA), and culminates with reasoning models using RLVR on mathematical datasets. Each milestone reflects when methods became mainstream and accessible, often lagging years behind their initial publication due to tooling maturity and community adoption.

   32