| title | On-Policy Morphing into Hybrid Attention Models |
|---|---|
| authors | Disen Lan, Xuanda Wang, Jianbin Zheng, Huafeng Kuang, Wei Tang, Yu Cheng |
| date | 2026-11-18 |
| venue | arXiv (preprint, not peer reviewed) |
| id | arXiv:2611.24809 |
| affiliation | ByteDance Seed |
| upvotes | 12 (Hugging Face Daily Papers, 2026-11-21) |
| links | abs pdf hf |
| cite | |
| related | Training Long-Context Vision-Language Models Effectively with Generalization Beyond 128K Context Continuous Latent Diffusion Language Model HarnessDev: Can LLMs Create and Evolve Their Own Agent Harness? DanceOPD: On-Policy Generative Field Distillation GenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modeling |
| abstract | Hybrid attention models combine full and linear attention to achieve a better quality--efficiency trade-off. Converting pretrained Transformers into such models requires identifying which layers should retain full attention under a prescribed budget. We identify two off-policy mismatches in FlashMorph's layer selection: (i) continuous branch interpolation and teacher-forced hidden states differ from discrete hybrid computation; and (ii) frozen model weights fail to account for parameter adaptation during subsequent recovery. We introduce FlashMorph2, an on-policy layer-selection method that (I) evaluates sampled discrete hybrid architectures end to end using their own hidden states, and (II) jointly optimizes layer-selection scores and model weights under the recovery objective to reflect post-adaptation performance. Experiments across Qwen3 model scales and linear-attention variants show that FlashMorph2 consistently outperforms existing layer-selection methods, improving average scores over FlashMorph by +2.42 points on RULER, +1.08 on LongBench, and +0.76 on recall-intensive tasks, while completing layer selection in only 0.72--3.57 GPU hours across 0.6B--8B models, demonstrating high efficiency and scalability. |
bibtex
@misc{lan2026policy,
title={On-Policy Morphing into Hybrid Attention Models},
author={Disen Lan and Xuanda Wang and Jianbin Zheng and Huafeng Kuang and Wei Tang and Yu Cheng},
year={2026},
eprint={2611.24809},
archivePrefix={arXiv},
url={https://arxiv.org/abs/2611.24809}
}