FastGRPO speeds up GRPO training by 2.35×‑2.72× using concurrency‑aware speculative decoding and online draft learning, keeping reasoning quality stable. Read more: https://getnews.me/fastgrpo-speeding-policy-optimization-with-concurrency-aware-decoding/ #fastgrpo #speculativdecoding #rl

Training‑free speculative decoding lifts LLaMA 3 scores by 3.3 points and speeds generation 2.23×, with draft token acceptance up to 2.39 tokens. Read more: https://getnews.me/training-free-speculative-decoding-enhances-llama-3-speed-and-accuracy/ #llama3 #speculativdecoding

SelfJudge trains verification judges from the model’s own outputs, removing the need for human‑annotated data. Benchmarks show faster inference with higher accuracy. Read more: https://getnews.me/selfjudge-boosts-speculative-decoding-for-faster-llm-inference/ #selfjudge #speculativdecoding

Researchers found a side‑channel in speculative decoding that can fingerprint queries with up to 100% accuracy on REST and leak data at over 25 tokens per second. Read more: https://getnews.me/side-channel-risks-in-speculative-decoding-for-large-language-models/ #speculativdecoding #sidechannel

DiffuSpec swaps the usual draft model for a pretrained diffusion language model, a training‑free method that can cut inference time by up to three times, per benchmark results. Read more: https://getnews.me/diffuspec-boosts-speculative-decoding-using-diffusion-models/ #diffuspec #speculativdecoding