Topology-Aware Risk Allocation and Controlled Ternary Quantization for Heterogeneous Language Models
Anonymous research manuscript
TRACE-Q is a post-training ternary quantization framework for heterogeneous language models. It models how quantization error can propagate through different architectural components, then allocates a quantization budget with topology-aware risk control.
- Identifies architecture-specific quantization risk across attention, recurrent state, expert routing, and multimodal components.
- Uses controlled ternary quantization to balance compression efficiency and model quality.
- Targets practical, structure-aware deployment of heterogeneous language models.
This repository preserves an anonymous internal-review version of the manuscript. It is not a submission. Any internal evaluation values in the paper are not empirical measurements.
Copyright remains with the manuscript authors. No license is granted for reuse, redistribution, or derivative works without permission.