TLAB Mechanistic Interpretability submission.
A lightweight denoising approach for reducing language degradation caused by activation steering, plus an orthogonal correction that preserves the intended steering component.
TL;DR
Standard activation steering modifies a residual-stream activation as
[
\tilde h = h + \alpha v,
]
where v is a steering direction and alpha controls strength. Strong interventions can push the representation away from the model's usual activation manifold, leading to repetition and degeneration.
I train a small residual MLP denoiser on clean GPT-2 activations corrupted with Gaussian noise, without exposing it to the validation steering vectors:
The third method is motivated by a mechanistic observation: the denoiser correction increasingly points against the steering vector as steering becomes stronger. Removing only the parallel correction preserves the intended semantic intervention while keeping the orthogonal cleanup.
Main findings
The denoiser reduces validation reconstruction MSE from 0.4895 to 0.3322 — a 32.1% reduction.
The denoiser was trained on clean activations only; the validation steering directions were not seen during training.
For negative-sentiment steering, the correction becomes increasingly anti-aligned with the steering direction:
cos(δ, v) changes from about -0.34 at r=0.10 to -0.71 at r=0.50.
This motivates the orthogonal correction.
At moderate steering strengths, denoising-based methods improve parts of the concept–quality Pareto frontier.
The effect generalizes from the negative steering direction to a held-out positive-sentiment SAE feature.
The improvement is not universal: extreme steering still causes degeneration, and the full denoiser can alter clean behavior even at r=0.
Experimental setup
Component
Choice
Base model
GPT-2 Small
Intervention point
blocks.5.hook_resid_post
SAE
OpenAI GPT-2 sparse autoencoder
SAE location
resid_post_mlp_v5_32k, layer 5
Activation dimension
768
Denoiser hidden dimension
1536
Denoiser form
x + MLP(LayerNorm(x))
Training activations
50,000
Train / validation split
45,000 / 5,000
Dataset
WikiText-2
Noise
Gaussian, σ = 0.7
Optimizer
AdamW
Learning rate
1e-3
Batch size
256
Epochs
15
Decoding
Greedy, 30 new tokens
Relative steering strength is normalized by activation magnitude:
The identity/noisy-input baseline gives validation MSE ≈ 0.4895.
The best checkpoint reaches validation MSE ≈ 0.3322 at epoch 15.
This is useful as a training sanity check, but reconstruction MSE is not the main success criterion. The main question is whether steering achieves a better trade-off between concept strength and generation quality.
Why orthogonal correction?
For
[
x=h+\alpha v,
\qquad
\delta=D(x)-x,
]
the denoiser correction has a substantial component parallel to v, and that component is negative at stronger steering.
At r=0.25, for example, the correction norm is ≈ 11.72, with ≈ 6.83 in the parallel component and ≈ 9.53 in the perpendicular component.
This suggests that full denoising partly undoes the intended steering. Orthogonal correction removes only the component of the denoiser update parallel to v.
Evaluation
Concept score
For sentiment steering I use the pretrained SST-2 classifier:
Important: Dist-n is treated as a repetition/diversity proxy, not a complete fluency metric. NLL is also imperfect: repetitive text can be highly predictable and therefore have deceptively low NLL.
A local Qwen2.5-3B-Instruct judge was sanity-checked on fluent and deliberately degraded examples, but it was not used for the final aggregate results, because the raw generations were not preserved after a kernel restart. I therefore do not present LLM-judge scores as experimental evidence.
Why there are no TP / FP / TN / FN counts
This experiment does not define a binary classification task. The concept metric is a continuous probability and no classification threshold was preregistered. Reporting true positives / false negatives post hoc would therefore be misleading. The correct comparison is the continuous concept–quality trade-off.
Results: negative sentiment
Selected points:
Method
r
Concept score
Dist-3
Baseline
0.20
0.750
0.531
Orthogonal
0.20
0.878
0.637
Baseline
0.30
0.873
0.598
Orthogonal
0.20
0.878
0.637
Baseline
0.40
0.993
0.441
Orthogonal
0.25
0.997
0.551
Two useful Pareto-style comparisons:
orthogonal r=0.20 vs baseline r=0.30: slightly higher concept score and higher Dist-3.
orthogonal r=0.25 vs baseline r=0.40: higher concept score and substantially higher Dist-3.
At r=0.50, orthogonal correction no longer behaves consistently; extreme steering remains a failure regime.
Results: held-out positive sentiment
The denoiser was not trained on this validation direction.
Selected observations:
r
Baseline (C, Dist-3)
Full (C, Dist-3)
Orthogonal (C, Dist-3)
0.15
(0.498, 0.786)
(0.658, 0.864)
(0.659, 0.891)
0.20
(0.533, 0.839)
(0.659, 0.900)
(0.556, 0.865)
0.25
(0.626, 0.857)
(0.524, 0.820)
(0.874, 0.830)
0.30
(0.873, 0.696)
(0.874, 0.813)
(0.997, 0.714)
The held-out positive direction supports the same high-level conclusion: denoising-based correction can improve the trade-off in the moderate-steering regime, but neither full nor orthogonal correction universally dominates.
Example causal steering behavior
Positive feature 15385:
text
1r = 0.00
2"The experience was a bit different ..."
34r = 0.20
5"The experience was a success ... I was very impressed ..."
67r = 0.25
8"The experience was a success ... It was a great product ..."
910r = 0.50
11"The experience was described ... very good quality ...
12 the quality of the service ... the service of the service ..."
This illustrates the characteristic steering curve: useful concept control at moderate strength, followed by repetition/degeneration at stronger intervention.
Replace <YOUR_USERNAME> with the final public repository owner before submission.
Limitations
GPT-2 Small only.
One fixed intervention layer.
Small validation prompt set.
Greedy decoding only.
Dist-n measures diversity/repetition rather than full linguistic fluency.
NLL can reward predictable repetition.
Full denoising changes model behavior even at r=0.
The orthogonal method does not uniformly dominate full denoising.
Extreme steering remains unstable.
The scientific-discourse vector was causally validated only qualitatively.
No final aggregate LLM-judge evaluation was completed.
Conclusion
A cheap denoiser trained only on noisy clean activations can improve activation steering without learning the held-out steering vector itself.
The most informative mechanistic result is that the learned denoising update increasingly opposes the steering direction at larger intervention strengths. Projecting that correction orthogonally to the steering vector gives a simple way to preserve the intended semantic intervention while still applying denoising in the remaining residual-stream directions.
The result is not a universal fix, but it produces clear Pareto-improving points in the moderate-steering regime and generalizes to a held-out positive sentiment direction.
