BCJR-QAT makes trellis quantization differentiable via BCJR soft decoding at finite temperature, allowing QAT to improve 2-bit LLM perplexity over PTQ with a fused GPU kernel and a drift-budget escape condition.
Maddison, Andriy Mnih, and Yee Whye Teh
2 Pith papers cite this work. Polarity classification is still indexing.
2
Pith papers citing it
citation-role summary
background 1
citation-polarity summary
fields
cs.LG 2years
2026 2verdicts
UNVERDICTED 2roles
background 1polarities
background 1representative citing papers
NBSR frames neural inference as conjugate Dirichlet evidence accumulation over a DAG with Gumbel-Softmax routing, yielding monotonic precision growth and uncertainty-aware early exit.
citing papers explorer
-
BCJR-QAT: A Differentiable Relaxation of Trellis-Coded Weight Quantization
BCJR-QAT makes trellis quantization differentiable via BCJR soft decoding at finite temperature, allowing QAT to improve 2-bit LLM perplexity over PTQ with a fused GPU kernel and a drift-budget escape condition.
-
Neural Bayesian Sequential Routing
NBSR frames neural inference as conjugate Dirichlet evidence accumulation over a DAG with Gumbel-Softmax routing, yielding monotonic precision growth and uncertainty-aware early exit.