Future Work

What is implemented, what was tested and dropped, and what remains.

Per-step intensity scheduling — implemented

generate_band_steered takes an optional schedule(step, n_steps) -> [0, 1] multiplier on the user's base strength, applied per generation step via tracer.iter. dim_schedule (linear taper) and start_schedule (steer the opening tokens) keep hard-steered generations fluent while preserving the strength dial. These are the remote-feasible schedules from Scalena, Sarti & Nissim, Multi-property Steering with Dynamic Activation Composition (BlackboxNLP 2024, https://aclanthology.org/2024.blackboxnlp-1.34/).

Full KL-adaptive DAC (their Eq. 6) — tested, not adopted. It sets intensity fully automatically from a KL criterion (no dial) and is a per-token feedback loop, so it is local-only and cannot run on the remote path. The simple, dial-preserving schedules above deliver the same fluency benefit, so DAC is not kept in the core.

Per-position direction extraction — tested, not adopted

The paper extracts a separate direction per generation step. We tested this (K=50 minimal pairs, per-position contrast at the answer tokens): the per-position directions drift from each other but largely as noise, each still tracks the pooled direction, and there is no evidence they steer better than the single pooled vector across a layer band. For a global persona attribute, per-position extraction adds a heavier pipeline for no gain. See the report appendix.

Trait vectors on the Hub

Local persistence is implemented (TraitVectorStore + traits.save_trait_deltas / load_trait_direction / load_trait_band, reading the per-layer mean delta from artifacts/trait_vectors/). Only Hub loading remains: add an HFTraitVectorStore mirroring HFPersonaVectorStore (and a TraitVectorSource union over the two) so trait vectors load from the Hub or disk through one contract.