GRAFT lets one backbone absorb five vision foundation models in sequence, with each new capability costing a single distillation

It proposes GRAFT, a continual multi-teacher distillation framework that lets a unified backbone progressively acquire capabilities from an open-ended sequence of foundation models. Existing approaches assume a fixed teacher set and require repeating expensive joint distillation over the entire teacher set when a new teacher is added; GRAFT instead treats the previously distilled model as a teacher to preserve learned capabilities while the student learns from both the previous model and the incoming teacher. At the abstract level, the framework design and motivation are stated, with the cost of acquiring each new capability described as a single distillation rather than a full re-distillation; no datasets, metric values, or ablations are provided.

It introduces Teacher Specific Readout Tokens that grant each teacher an independent read-out of the shared encoder. To address incompatible representation geometries among heterogeneous teachers, it uses teacher-specific read-outs rather than a shared read-out interface for all teachers. The abstract lists this as a method component and describes its role in reconciling representation geometry differences; no quantitative comparison of this component is given.

It introduces a Geometry Agnostic Relational Loss that aligns a vision-language teacher by matching image-text similarity structures rather than raw feature values. The alignment target for the vision-language teacher shifts from raw feature-value matching to relational-structure matching, sidestepping geometric incompatibility. The abstract states that the loss aligns image-text similarity structures; no loss values or alignment metrics are reported.

It provides the GRAFT model, a single continually extensible backbone that unifies five domains with strong performance across all of them. Relative to maintaining separate specialized foundation models, GRAFT consolidates image understanding, 2D dense prediction, 3D human pose estimation, 3D vision, and vision-language capabilities into one backbone. The abstract claims strong performance across all five domains but lists no benchmark names, numbers, or per-domain comparisons against the specialized teachers.

The work targets settings where capabilities from multiple vision foundation models must be consolidated into a single backbone and teachers arrive over time, for example unified modeling of image understanding, 2D dense prediction, 3D human pose estimation, 3D vision, and vision-language. As described in the abstract, the design goal is for each new capability to cost a single distillation rather than a full joint re-distillation over all teachers, so it suits research and engineering environments where the teacher set is open-ended and incremental extension is more common than a one-time fixed set. The abstract does not state applicable model scale, data scale, or compute conditions, which would need to be checked in the original text.

The abstract provides no benchmark names, performance numbers, teacher count or ordering, distillation rounds, or compute cost, so the magnitude of the claimed strong performance and the degree to which old capabilities are retained when new ones are added cannot be judged. The individual contribution of Teacher Specific Readout Tokens and the Geometry Agnostic Relational Loss, and the effect of teacher arrival order, would need to be examined in the body. Based on the current text alone, readers should treat the conclusions as the authors' reported design goals and overall performance rather than quantified comparative findings.