DecodeShare reveals shared subspace in LLM decode-time decisions

A team of researchers led by Zishan Shao (Duke University, UNC Chapel Hill, and others) has introduced DecodeShare, a protocol that pinpoints a low-dimensional subspace shared across tasks in large language model (LLM) decode-time hidden states. Under KV-cached inference, LLMs rely on a decoded token sequence that builds on stored key-value pairs. The team hypothesized that a task-general structure exists specifically at decode time, distinct from the initial prefill encoding. Using causal intervention experiments, they found that disturbing this decode-shared subspace degrades model decision performance significantly more than disturbing either a prefill-derived or a random subspace under the same budget. This suggests the subspace acts as a high-leverage causal channel for LLM outputs.

The work has immediate practical implications for activation steering—a technique for controlling model behavior by adding vectors to hidden states. The researchers show that common steering directions often overlap with the task-general decode channel, muddying their effect. By projecting out the shared subspace, they can cleanly separate the functional roles of steering components. Critically, they demonstrate that evaluating steering vectors at decode-time (rather than using prefill-based proxies) yields more reliable signals for downstream deployment. Despite its compactness, the shared subspace offers a surprisingly powerful handle for fine-grained model control. The paper and code are available on arXiv.