An arXiv preprint dated 2026-10-07 introduces ResidualQuant, a KV cache quantization technique for looped Transformers. It leverages similarity of KV states across loops to…
A new arXiv preprint introduces SIGMA, a pipeline where models use only a written Model Spec to synthesize alignment dilemmas, then train themselves via supervised fine-tuning…
An arXiv preprint dated October 6 2026 introduces OSFP4 a quantization scheme that jointly optimizes diagonal smoothing entries and block scales to minimize squared…
An arXiv preprint introduces MASC, a multi-agent self-calibration system using latent construct alignment to maintain psychological coherence in simulated counseling clients.…
Ai2 has published its byteifying method for converting subword models to byte-level operation in Nature. New Bwen 8B and Blama 8B checkpoints, derived from Qwen 3 and Llama 3…
arXiv preprint dated October 6 2026 introduces MedCORE a vision-language model that decomposes diagnosis into clinical criteria localizes evidence and models inter-criteria…
New preprint argues that probe scores like R² of 0.6 lack fixed meaning without baselines. It introduces floor from simple inputs, ceiling from full input, and headroom between…
arXiv preprint dated October 5 2026 presents the first systematic measurements of complete training runs for a 3B-parameter model on a mobile device, covering memory, time…
An arXiv preprint from October 2026 tests open-source generative LLMs on ContractNLI and NLI4Wills datasets for legal Natural Language Inference. Gemma-4 26B reaches 81.2…
A new arXiv preprint introduces Sequential Local Operator Alignment, a training-free approach that merges fine-tuned transformers by aligning operators along the partially…