light-denoising-for-activation-steering

Light activation denoiser for steering — Llama-3.2-1B, layer 7

A 55M-parameter denoiser that repairs the damage activation steering does to a language model's residual stream, at one forward pass per token — a low-cost stand-in for GLP's flow-matching activation model (0.5–3.3B parameters, 20 passes).

Trained on 8M activations from blocks.7.hook_resid_post of Llama-3.2-1B, the exact hook point GLP uses. Steering directions come from the SAE chanind/sae-llama-3.2-1b-topk-res (d_sae 16384, top-k 10).

Code, full report and reproduction instructions: https://github.com/artemmavrov/light_denoising_for_activation_in_LLM

Files

Quick start

git clone https://github.com/artemmavrov/light_denoising_for_activation_in_LLM
cd light_denoising_for_activation_in_LLM
pip install -r requirements.txt

# fetches this repo, prints steered vs. denoised continuations side by side
PYTHONPATH=src python scripts/quickstart.py --demo

# reduced sweep + paired bootstrap, compares against the published intervals
PYTHONPATH=src python scripts/quickstart.py --repro

Needs a GPU with ≥ 6 GB free (Llama-3.2-1B fp16 + the denoiser). Access to meta-llama/Llama-3.2-1B is gated; the ungated mirror unsloth/Llama-3.2-1B works and is what the config points at.

Using it directly

import torch
from lds.activations import Scaler
from lds.correction import Intervention, unit
from lds.denoisers.io import load_denoiser
from lds.model import load_lm, intervene_layer

bundle   = load_lm("unsloth/Llama-3.2-1B", dtype="float16", layer=7)
scaler   = Scaler.load("scaler.pt", device=str(bundle.device))
denoiser = load_denoiser("F_cond_N8.pt", bundle.d_model, device=str(bundle.device))

v_hat = unit(my_direction).to(bundle.device)      # unit-norm steering direction
alpha = 1.0 * 4.8955                              # r * E||h||

intervention = Intervention(
    v_hat=v_hat, alpha=alpha, scheme="C1",        # C0 = plain steering, C1 = denoised
    denoiser=denoiser, scaler=scaler, x_norm=45.12,
)
with intervene_layer(bundle, intervention):
    out = bundle.model.generate(**inputs, max_new_tokens=48)

scheme selects the correction: C0 plain h + αv, C1 full denoising, C2 partial (s + λΔ), C5 norm-preserving, C6 multi-step.

What it does and when it helps

The denoiser acts as a learned soft clamp: it removes 81% of the steering along v at r = 2, yet its output stays at a constant Mahalanobis distance (37–39) from the clean activation distribution while the steered input drifts from 53 to 122. It does not undo the intervention — it brings an impossible feature level back to a plausible one.

Established, replicated result — less degenerate text at any steering strength:

On 5 DiffMean concepts with objective scorers it wins on both axes: Δconcept = +0.136 [+0.020, +0.274] at r = 2, at NLL 7.31 vs. 8.28.

Where it does not help. The concept gain on SAE features is conditional and, averaged over all 32, negative. The sign is predicted by whether plain steering collapses on that direction: correlation between "retention" and the denoiser's gain is −0.62 [−0.75, −0.42]. Where steering collapses the denoiser wins in 16 of 17 directions; where it holds up, in 5 of 15.

Cheaper alternative for SAE features. Calibrating α so the feature lands at the top of its natural range gives the same concept as this checkpoint (0.01945 vs. 0.01946) at 1.56 nats lower NLL and zero inference cost. The mean calibrated r is 0.52. See report §6 — this is a negative result for the method and is reported as such.

Cost

Caveats

• One model, one layer, one SAE (l0 = 10, 61.8% explained variance).
• The SAE-family concept metric is partly self-referential: the same SAE both supplies the steering direction and scores the generated text.
• The LLM-judge concept criterion was measured and withdrawn (false-positive rate 0.259, position inconsistency 0.38). The fluency axis is validated (r = −0.96 against NLL).
• fp16 throughout (trained on a Tesla T4, which has no bf16).