Function-vector anchoring
Function-vector anchoring is a proposed structural preservation constraint in the post-V-KTO candidate list.
Project status: Research proposal/design in the inspected project sources; no implementation or completed local efficacy run identified. This entry describes the source audit of 14 September 2026; historical measurements retain their original dates.
Mechanism
Anchor selected internal representations associated with useful functions, aiming to preserve skills rather than only the sampled token distributions. The next-arms document calls this a Function-Vector KL anchor; the exact mathematical object and its relationship to activation matching need specification.
Implementation and controls
The historical proposal points to a small extension of the KL implementation and frames it as a preservation lever rather than a direct gain mechanism. A concrete design must identify layers, contexts, the function-vector estimator, the reference state, and a loss whose units and masking are defined.
Evidence and evaluation
The cited project document records this candidate and its intended experiment. It does not provide a completed local result for this method. Published-paper results mentioned by that document are background, not Trainfer measurements.
Limitations and interpretation
The named concept is less specified than the current token-distribution KL or CCD hidden-state matching. It should not be advertised as an available kl_anchor scope. Anchoring one representation can preserve some functions while constraining useful adaptation elsewhere.
Sources
- agi: lile/docs/research/NEXT_ARMS_jun02.md — historical revision
3842fd8875ca. - trainfer: trainfer/objectives/kl.py — checkout audited
1c6391f3773b.