LLM Research Highlights: December 27, 2025 – January 2, 2026

Bayesian & Cognitive Advances: Transformers achieve ultra-precise Bayesian inference with 10⁻³ to 10⁻⁴ bit accuracy, while CREST boosts reasoning accuracy by 17.5% and cuts token usage by 37.6%. Model Efficiency & Scaling: TG reduces data needs by up to 8% and parameters by 42% compared to GPT-2, and Recursive Language Models handle 100x longer inputs at similar or lower inference costs. State-of-the-Art Performance: Youtu-LLM sets a new bar for sub-2B parameter models with 128k context, DLCM improves zero-shot benchmarks by +2.69%, and ADOPT outperforms all prior prompt optimization methods. Benchmark Breakthroughs: Encyclo-K’s top models hit 62.07% accuracy on complex knowledge queries, while Youtu-Agent accelerates RL training by 40% and scores over 71% on WebWalkerQA and GAIA benchmarks. Novel Insights & Safety: Diffusion Language Models match optimal step complexity in chain-of-thought sampling, and safety analysis reveals a 9.2x disparity between past- and future-tense prompt safety rates. This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit llmsresearch.substack.com (https://llmsresearch.substack.com?utm...)