2 papers
cs.LG2026
WinQ: Accelerating Quantization-Aware Training of Language Models Around Saddle Points
Dongyue Li, Zechun Liu, Kai Yi +6
Quantization-aware training (QAT) is widely adopted to quantize language models by training full-precision weights using gradients from the quantized model. The main bottleneck is…
cs.LG2026
Attention Drift: What Autoregressive Speculative Decoding Models Learn
DoÄaç Eldenk, Payal Mohapatra, Yigitcan Comlek +3
Speculative decoding accelerates LLM inference by drafting future tokens with a small model, but drafter models degrade sharply under template perturbation and long-context inputs.…