LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation¶
L1 core-mechanism diagnostic. Projects policy gradients with example-feature SVD; does not train a paper-scale language model.
Paper information¶
| Field | Details |
|---|---|
| Paper | LOCUS: Task-Aware Low-Rank Post-Training for Token-Efficient Language Generation |
| Organization | First-author affiliation: University of Washington |
| First public release | 2026-09-10 |
| Authors' code | No public author repository was found on 2026-09-14. |
| Adapter / method | locus |
| Local reproduction code | src/auto_research/post_training/latest_20260914.py |
Original paper¶
Motivation and method¶
Learn a task-related low-rank post-training subspace. Updating only compact directions shortens responses while retaining task capabilities.
flowchart LR
I[Input and public state] --> M[LOCUS mechanism]
M --> A[Auditable intermediates]
A --> O[Output or update]Key figure from the paper¶
Figure 2: Overview of the paper's method and key components. From the original paper; copyright remains with the authors.
Core computation or operation¶
The local code implements the decisive operation as a deterministic reference kernel and explicitly reports gate, selection, projection, graph-edge, or state-update statistics. Tests cover formula invariants and boundary conditions.
Paper results¶
Pythia outputs shorten by up to 39.84% and Qwen by 14.87%–17.58%, with about 0.24%–0.28% trainable parameters. These are paper results, not directly comparable with the scaled-down local run. No online gains are inferred in the absence of an online A/B test.
Local reproduction¶
See the three-seed metrics. All candidates share the public fixtures, budget, and seeds. The results verify actual execution of core state transitions only.
Reproduction boundaries¶
Local code projects policy gradients with example-feature SVD. No paper-scale language model was trained.