Skip to content

CURRENT_STATE.md

Latest commit

 

History

History
18 lines (14 loc) · 1.13 KB

CURRENT_STATE.md

File metadata and controls

18 lines (14 loc) · 1.13 KB

Current State: Phase 1 (Token-Level Control)

What We've Actually Built

As of January 2026, we have successfully established the technical feasibility of latent steering.

Negative Results (The Falsified Assumptions)

We proved that simple methods used in Transformers fail in SSMs:

  • Linear Addition: h + h_target has zero sensitivity.
  • PCA Steering: Vectors found in natural transitions cannot force target logits.
  • Contrastive Steering: Delta vectors between concepts (Blue - Red) do not shift the probability manifold.

Positive Results (The Breakthroughs)

  • Functional Mamba: Developed a pure PyTorch implementation of the Mamba step, solving the Autograd in-place update blocker.
  • Controllability: Proved that for any context, there exists a computable $\Delta$ that forces a target token to Rank 1.
  • Trajectory Shaping: Implemented BPTT through time to reduce looping artifacts.

The Semantic Gap

We can now force the model to say "BLUE". We cannot yet force the model to know "The user's name is BLUE". The transition from logit-forcing to fact-injection is the goal of Phase 2.