DecodeShare: Tracing the Shared Subspace of LLM Decode-Time Decisions

Zishan Shao, Lixun Zhang, Kangning Cui, Yixiao Wang, Ting Jiang, Hancheng Ye, Qinsi Wang, Zhixu Du, Yuzhe Fu, Fan Yang, Danyang Zhuo, Yiran Chen, Hai Helen Li
Proceedings of the 43rd International Conference on Machine Learning, PMLR 306:110338-110373, 2026.

Abstract

Large language models (LLMs) handle many tasks with one set of parameters, but under KV-cached inference it is unclear what task-general structure, if any, is used at $\textit{decode time}$ rather than during $\textit{prefill}$. We propose $\textbf{DecodeShare}$, a protocol that identifies a low-dimensional subspace consistently shared across tasks in decode-time hidden states, and then tests its causal role by removing that subspace only during decoding. In our experiments, disturbing the discovered shared subspace degrades decision performance far more than disturbing either a prefill-derived or random subspace under the same intervention budget. We further show this decode-shared subspace has practical consequences for activation steering: common steering directions can overlap the task-general decode channel. Projecting out this shared subspace directly separates the functional roles of the two components, while evaluating steering vectors at decode-time yields more reliable signal for downstream deployment than prefill-based proxies. Despite its compactness, the shared subspace can serve as a high-leverage causal channel at decode time. Code is available at: https://github.com/Zishan-Shao/decodeshare.git.

Cite this Paper

BibTeX
@InProceedings{pmlr-v306-shao26j, title = {{D}ecode{S}hare: Tracing the Shared Subspace of {LLM} Decode-Time Decisions}, author = {Shao, Zishan and Zhang, Lixun and Cui, Kangning and Wang, Yixiao and Jiang, Ting and Ye, Hancheng and Wang, Qinsi and Du, Zhixu and Fu, Yuzhe and Yang, Fan and Zhuo, Danyang and Chen, Yiran and Li, Hai Helen}, booktitle = {Proceedings of the 43rd International Conference on Machine Learning}, pages = {110338--110373}, year = {2026}, editor = {Zhang, Tong and Dudik, Miroslav and Jaggi, Martin and Agarwal, Alekh and Li, Sharon and Schuurmans, Dale and Zhu, Jerry and Berkenkamp, Felix and Dong, Hanze and Bietti, Alberto}, volume = {306}, series = {Proceedings of Machine Learning Research}, month = {06--11 Jul}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v306/main/assets/shao26j/shao26j.pdf}, url = {https://proceedings.mlr.press/v306/shao26j.html}, abstract = {Large language models (LLMs) handle many tasks with one set of parameters, but under KV-cached inference it is unclear what task-general structure, if any, is used at $\textit{decode time}$ rather than during $\textit{prefill}$. We propose $\textbf{DecodeShare}$, a protocol that identifies a low-dimensional subspace consistently shared across tasks in decode-time hidden states, and then tests its causal role by removing that subspace only during decoding. In our experiments, disturbing the discovered shared subspace degrades decision performance far more than disturbing either a prefill-derived or random subspace under the same intervention budget. We further show this decode-shared subspace has practical consequences for activation steering: common steering directions can overlap the task-general decode channel. Projecting out this shared subspace directly separates the functional roles of the two components, while evaluating steering vectors at decode-time yields more reliable signal for downstream deployment than prefill-based proxies. Despite its compactness, the shared subspace can serve as a high-leverage causal channel at decode time. Code is available at: https://github.com/Zishan-Shao/decodeshare.git.} }
Endnote
%0 Conference Paper %T DecodeShare: Tracing the Shared Subspace of LLM Decode-Time Decisions %A Zishan Shao %A Lixun Zhang %A Kangning Cui %A Yixiao Wang %A Ting Jiang %A Hancheng Ye %A Qinsi Wang %A Zhixu Du %A Yuzhe Fu %A Fan Yang %A Danyang Zhuo %A Yiran Chen %A Hai Helen Li %B Proceedings of the 43rd International Conference on Machine Learning %C Proceedings of Machine Learning Research %D 2026 %E Tong Zhang %E Miroslav Dudik %E Martin Jaggi %E Alekh Agarwal %E Sharon Li %E Dale Schuurmans %E Jerry Zhu %E Felix Berkenkamp %E Hanze Dong %E Alberto Bietti %F pmlr-v306-shao26j %I PMLR %P 110338--110373 %U https://proceedings.mlr.press/v306/shao26j.html %V 306 %X Large language models (LLMs) handle many tasks with one set of parameters, but under KV-cached inference it is unclear what task-general structure, if any, is used at $\textit{decode time}$ rather than during $\textit{prefill}$. We propose $\textbf{DecodeShare}$, a protocol that identifies a low-dimensional subspace consistently shared across tasks in decode-time hidden states, and then tests its causal role by removing that subspace only during decoding. In our experiments, disturbing the discovered shared subspace degrades decision performance far more than disturbing either a prefill-derived or random subspace under the same intervention budget. We further show this decode-shared subspace has practical consequences for activation steering: common steering directions can overlap the task-general decode channel. Projecting out this shared subspace directly separates the functional roles of the two components, while evaluating steering vectors at decode-time yields more reliable signal for downstream deployment than prefill-based proxies. Despite its compactness, the shared subspace can serve as a high-leverage causal channel at decode time. Code is available at: https://github.com/Zishan-Shao/decodeshare.git.
APA
Shao, Z., Zhang, L., Cui, K., Wang, Y., Jiang, T., Ye, H., Wang, Q., Du, Z., Fu, Y., Yang, F., Zhuo, D., Chen, Y. & Li, H.H.. (2026). DecodeShare: Tracing the Shared Subspace of LLM Decode-Time Decisions. Proceedings of the 43rd International Conference on Machine Learning, in Proceedings of Machine Learning Research 306:110338-110373 Available from https://proceedings.mlr.press/v306/shao26j.html.

Related Material