Search publications
“For n ≥ 2, let cuv be the binary multiplicity of the unordered pair uv and set c′uv = min{cuv, k}, auv = ⌊ 2n2 c′uv / k ⌋. (6.6)” · Show in paper
“Represent the independent weights as τe = Φ(Ue), Ue = ∑_{h≥1} 2^{−h} , where all are independent Bernoulli variables of parameter 1/2.” · Show in paper
“We will keep one copy of g∗ throughout: first remove the product masks, then eliminate the other linear factors, and finally identify the sampling law of the surviving cube root.” · Show in paper
“Allow at most Dbits trials in each uniform-integer call; if all fail, abort the run and return zero. The cap is O(1 + log M). Couple this procedure with the unrestricted-trial one using the same random bits up to an abort.” · Show in paper
“Finding 1: QATFactory converges stably across model families, parameter counts from 8B to 230B, and NVFP4, MXFP4, and Q4_K. In vLLM and llama.cpp, the exported QAD checkpoints achieve higher average benchmark accuracy and lower mean KL divergence than the RTN and GPTQ post-training quantization baselines.” · Show in paper
QATFactory: A Versatile, Deployment-Aligned Framework for Quantization-aware Training and Distillation of LLMsAll reported training curves, benchmark accuracies, and KL divergence metrics in Figures 3-4 and Tables 2-3 directly support this claim.
“For slot n, let the donor D be x when n is even and y when n is odd, and set C = P( D), B = P( D), A = P( D) − B.” · Show in paper
“A channel is a uniformly random matrix A = (a0, F) ∈ FH×h0_P conditioned on the last r = h0 − 1 columns F being linearly independent.” · Show in paper
“For each decode layer j ∈ {1, ..., Ld}, BridgeKV first finds the prefill layer at the same relative depth, cj = 1 + floor((j-1)(Lp-1)/max(1, Ld-1)), and takes the window of W consecutive prefill layers centered on cj, shifted inward at the first and last layers so that it always contains W valid layers” · Show in paper