WRSWorld Research StateContribute
WORLD RESEARCH STATE

Priming: converting pretrained Transformers into hybrid state-space models

Contribution grade

Cross-architecture knowledge transfer · Human-designed research method or resource

Reference score50,000Provisional editorial estimate of this exact contribution; its grade is assessed separately.

What changed

Selected attention layers using substitution-based importance, transferred compatible projections and expanded state capacity, then aligned hidden representations and adapted the full hybrid model with next-token training.

Scientific scope

Selected attention layers using substitution-based importance, transferred compatible projections and expanded state capacity, then aligned hidden representations and adapted the full hybrid model with next-token training. Per-input realization existence is not a common low-dimensional exact realization across all inputs; practical conversion is approximate and the selection score is a proxy. Although the abstract emphasizes layerwise alignment, method Section3.2.4 says end-to-end final-hidden-state MSE is the default after ablation. Reported savings exclude source pretraining; throughput uses8H200 GPUs, BF16 weights and FP32 recurrent state. The three-way SSM comparison is limited to the stated8B/50%-hybrid/Qwen setup.

Evidence

The linked original sources describe the stated contribution and its evidence limits. WRS has not independently rerun the experiments or certified a proof. Per-input realization existence is not a common low-dimensional exact realization across all inputs; practical conversion is approximate and the selection score is a proxy. Although the abstract emphasizes layerwise alignment, method Section3.2.4 says end-to-end final-hidden-state MSE is the default after ablation. Reported savings exclude source pretraining; throughput uses8H200 GPUs, BF16 weights and FP32 recurrent state. The three-way SSM comparison is limited to the stated8B/50%-hybrid/Qwen setup.

Contributors

Provisional WRS assessment, edition 2026-10-16.1. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.

Full scientific record · Browse contribution grades · Request a correction