Researchers introduced DominoTree, a training-free best-first draft tree method for speculative decoding that uses Domino's conditional correction to achieve up to 6.6x speedup over autoregressive dec…
Researchers distilled 2.3 million reasoning traces from Claude Fable 5 into Qwen3-4B, achieving 100% self-consistency, zero output entropy, and zero hallucination variance. The team open-sourced the m…
Researchers introduced BlockPilot, a new speculative decoding method that adapts block sizes to individual inputs, achieving up to 4.20 times speedup on Qwen3-4B. The approach reduces document process…
Hugging Face and Pollen Robotics demonstrated a fully local conversational AI pipeline on the Reachy Mini desktop robot, using Silero VAD v5, Parakeet-TDT 0.6B v3, Gemma 4 or Qwen3-4B LLM, and Qwen3-T…
Researchers compared six offline reinforcement-learning methods for distilling reasoning from large language models into smaller ones, finding that SFT, RFT, and RIFT produce nearly identical weight u…
Researchers introduced cuTile Rust, a tile-based system for safe, idiomatic GPU kernel authoring in Rust that extends Rust's ownership discipline to GPU kernels. On the NVIDIA B200 GPU, cuTile Rust ac…
Researchers propose that publishers precompute key-value (KV) caches for documents, allowing AI agents to buy and load them instead of recomputing prefill from scratch. On Qwen3-4B, reuse is 9-50x che…
NVIDIA researchers introduced X-Token, a logit-distribution-based method for cross-tokenizer knowledge distillation that requires no auxiliary trainable components or architectural changes. The method…
A new study reveals that language model reasoning trajectories during test-time sampling cluster into "reasoning basins," causing majority votes to favor stable but potentially incorrect answers. The …