The Salt - Curated AI

NSA, RLDP, and MaPPO: Techniques for Efficient, Private, and Aligned LLMs

Rethinking LLM Reliability: Calibration, Compression, and the Hidden Costs of Inference-Time Scaling

Is RLVR Just a Conservative Reweighting Process?

Enabling Reasoning with Simple KV Cache Steering

Supervised Fine-Tuning Improves LLM Reasoning at the Cost of Other Skills

Tower+: Translation and General-Purpose Multilingual Models

RLPR: RLVR without Verifiers

Reward Correct CoT for Better Reasoning Models

Magistral: Advancing Reasoning with Efficient GRPO Training

Better Data Recipes for Pre-training LLMs and Training Reasoning Models

Reasoning Models Are More Prone to "Hallucination"

End-to-End FP4 Training for LLMs with Blackwell GPUs