An independent researcher released MAPLE, a structural pruning method for mixture-of-experts LLMs that removes dormant experts, shrinking Gemma 4 26B-A4B from 9.92 GB to 4.75 GB (about 1.48 bits per parameter) so it runs entirely in 6 GB of laptop VRAM at 45.4 tokens per second. The checkpoint is specialized for Python coding and STEM reasoning, and the pipeline is claimed to be domain-agnostic.
Nucleus Sampling
核采样
Random sampling from a candidate set whose cumulative probability exceeds a threshold
Curated Intelligence
A new Google DeepMind paper reveals that LLMs cannot generate true random choices autonomously. Experiments show models favor specific numbers regardless of temperature or sampling settings, but larger models succeed when provided with external random seeds. This finding has direct engineering implications for agent security and decision-making mechanisms relying on randomness.
NVIDIA released the open-source library LogitsProcessorZoo, enabling developers to directly intervene in language model probability distributions to customize generation behavior. Deeply integrated with Hugging Face Transformers, it offers finer control than traditional decoding strategies, suitable for scenarios requiring strict output constraints.
Hugging Face published a technical blog detailing various decoding methods for text generation using the Transformers library. The article compares greedy search, beam search, and sampling strategies to help developers optimize LLM output quality.