PRISM: Primitive Routing via In-context Skill Mixing for Retentive VLA Adaptation
Source check: Source checked. Not independent scientific confirmation.
Finding: The authors found that low-rank updates trained for individual skills occupied a distinguishable subspace and that a composite-task update concentrated roughly half its energy there, unlike a matched random-rotation control. PRISM then stored skill adapters as reusable primitives and routed among them. Benchmark and robot task results are separate from this representation result.
Technology: Vision-language model-conditioned router with frozen low-rank skill adapters
Publication: Preprint · Test context: retrospective or benchmark. Catalog categories, not an evidence grade.
- AI contribution
- The authors measured whether low-rank updates trained for individual manipulation skills occupy a shared parameter subspace, then proposed PRISM to preserve skill adapters as primitives and combine them with a vision-language-model-conditioned router.
- Test and result
- A composite-task update concentrated roughly half its Frobenius energy in the per-skill subspace, versus near-zero for a matched random-rotation control. RoboCasa365 and an AgileX PIPER setup were used for task evaluation.
- Limitation / still unknown
- The representation result is limited to the reported model and task setup; broader generalization is not established.
Evidence sources and annotation context
Evidence context (descriptions follow the cited sources; test type is a catalog classification):
Evidence sources: https://arxiv.org/abs/2611.08917v1
Source checked · editorial details
The source page was checked against this entry. Independent confirmation is not assessed by this label.
Source type: paper. Paper: not linked in this catalog; this does not mean no paper exists.
Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..
No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.
Editorial assessment: Impact 3/5 · Validation 2/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.