Activation steering controls a language model's behavior at inference time. It adds a fixed direction vector, scaled by a strength $\alpha$, to the model's hidden states, and it changes no weights. It is cheap to apply but expensive to tune. The result depends sharply on which layer, which sub-block, how many neuron dimensions, and what strength are used. Too little strength does nothing. Too much produces gibberish, repetition, or refusals. Today the only reliable way to pick a configuration is to generate text under every candidate and grade it, often with a model judge. That makes steering an architecture-search problem that is being solved by brute force. Neural architecture search faced the same situation, where every candidate needed full training. It responded with zero-cost proxies: scores computed from a single forward or backward pass that predict final quality well enough to rank candidates, as in EZNAS.
We propose to build the steering equivalent and to test it honestly. The plan has two steps. First, build a tabular benchmark that exhaustively maps steering configurations to their generation-evaluated outcomes, the steering analogue of NASBench-201. Second, use it to measure whether cheap, generation-free scores can recover the Pareto front between two goals: the target behavior shift, and preservation of fluency and capability. The Pareto front is the set of configurations where no other configuration is better on both. Treating steering tuning as a two-objective problem, rather than tuning one blended score, also addresses a known weakness of loss balancing. A training-time weighting fixes a single compromise. A cheaply computed front instead lets the user pick the trade-off at inference time.
Core Claim
Two teacher-forced quantities, computed without any generation, should rank steering configurations well enough to recover most of the generation-evaluated Pareto front at a small fraction of the cost. Teacher-forced means the model scores given text rather than producing its own. For a configuration $c$ (layer, sub-block, number of steered dimensions, strength), define:
$$B(c)=\mathbb{E}_{(x,y^+,y^-)}\big[\Delta_c \log p(y^+\mid x)-\Delta_c \log p(y^-\mid x)\big],\qquad F(c)=\mathbb{E}_{u}\,\mathrm{KL}\big(p_0(\cdot\mid u)\,\|\,p_c(\cdot\mid u)\big).$$
Here $(x, y^+, y^-)$ are held-out contrastive prompts with a desired and an undesired completion. $\Delta_c$ is the change in log-probability caused by steering with $c$. $u$ ranges over neutral text prefixes. $p_0$ and $p_c$ are the unsteered and steered next-token distributions. $B$ is a cheap estimate of behavioral effect and $F$ is a cheap estimate of the "alignment tax."
The claim is falsifiable. It has two concrete predictions:
- On held-out behaviors and models, the front implied by the $(B, F)$ scores, followed by full re-evaluation of a small shortlist, should reach nearly the same hypervolume as exhaustive search.
- It should beat existing single-number localization heuristics.
Hypervolume is the area of objective space dominated by a set of configurations.
The contrast with prior work is specific:
- DIRECTER already ranks layers by a one-pass cosine-sensitivity test, and AUSteer ranks individual neuron dimensions by "activation momentum" on contrastive pairs. Both are effectively zero-cost proxies. Neither has been validated against exhaustive ground truth, and both collapse the problem to a single objective.
- The sweeps in Patterns and Mechanisms of Contrastive Activation Engineering supply the kind of ground truth we need and document a universal perplexity penalty and out-of-distribution fragility. They propose no cheap predictor.
We supply the missing measurement and a two-objective proxy. The file also contains an unresolved tension: DIRECTER argues that directional change beats norm-based signals for predicting responsiveness, while NAS proxies rely heavily on norms. Our benchmark settles this empirically for steering.
Research Program
The benchmark
The first deliverable is a tabular steering benchmark. The models are Qwen2.5-1.5B and Gemma-2-2B, the small models used by DSAS and PID-Act, plus a subsampled slice on Llama-3.1-8B. For each model and behavior we enumerate configurations over five axes:
- every layer;
- three injection sites (residual stream, attention output, feed-forward output);
- two vector-extraction methods (difference-in-means as in contrastive activation addition, and top principal component as in Improving Reasoning Performance in LLMs via Representation Engineering);
- four sparsity levels (the full block, or the top 4, 16, or 100 dimensions by AUSteer's momentum score);
- about eight strengths, including negative ones, which the reasoning paper found useful.
This yields roughly five thousand configurations per model–behavior pair. That count is our estimate from the grid, not a reported number.
Each configuration is graded by actual generation on three behaviors:
- truthfulness on TruthfulQA, with a judge model;
- toxicity on RealToxicityPrompts and TET, with a toxicity classifier;
- verifiable instruction following on IFEval, which needs no judge.
The cost side of each configuration is graded as Wikipedia perplexity (the utility measure DSAS uses) and GSM8K accuracy (capability damage).
Proxy evaluation
The second deliverable evaluates proxies against this ground truth. The candidates are:
- DIRECTER's cosine sensitivity;
- AUSteer's momentum;
- norm-based scores in the style of NAS proxies;
- $B$ alone;
- $B$ and $F$ used jointly as a two-dimensional predicted front.
We report Kendall rank correlation with generation outcomes. We also report the NAS literature's harder test: correlation within the top decile, where EZNAS reports that proxies break down. The headline metric is front recovery, measured as the hypervolume obtained by re-evaluating the top-$m$ proxy-selected configurations, plotted against GPU-hours spent.
A small, secondary experiment runs an EZNAS-style search over proxy formulas. It must be trained on one behavior and one model and tested on others, because an evolved proxy that only fits its training cell would be worthless.
Do dynamic methods remove the sensitivity?
The third deliverable asks whether dynamic steering flattens the sensitivity surface or merely relocates it. PID-Act, DSAS, and DIRECTER's rejection rule each introduce their own settings: controller gains, detector placement, and a plausibility threshold. For one behavior per model we add a slice that sweeps these settings. We then ask whether outcome variance across the slice is actually smaller than for static steering, and whether the same proxies predict it.
Baselines
Paper defaults (naive reference). The do-nothing reference is the unsteered model, together with a "published default" configuration: a middle layer, the full residual block, and a strength taken from the source paper. This shows how much tuning matters at all.
Random search. Random search over the same space, re-evaluated with full generation, is run at every budget we report. Any proxy that cannot beat random search at equal GPU-hours has no value.
Successive halving (the strongest search competitor). This is multi-fidelity grid search in the spirit of Rethinking Performance Estimation in NAS. It grades all configurations with short generations on a few prompts, then promotes survivors to longer, fuller evaluations. It is the strongest competitor because it uses real generations and needs no modeling assumptions. It gets the same total GPU-hour budget as our method, the same prompts, and the same judges. We tune its promotion schedule on the training behavior only.
Existing localization heuristics (the strongest existing methods). These are DIRECTER's cosine layer ranking and AUSteer's momentum ranking. Each is given an equal-budget strength sweep on top.
Because no paper in the file reports search efficiency for steering configurations, there is no prior search number to beat. The published method results give the scale of what good configurations achieve:
- ITI raises Alpaca's TruthfulQA truthfulness from 32.5% to 65.1% (Inference-Time Intervention).
- DSAS reports nearly 70% toxicity reduction on TET for Qwen 2.5 (1.5B).
- PID-Act reports roughly 8.1x toxicity reduction on Gemma-2-2B.
- DIRECTER reports about 92% task fidelity on GSM8K-Format.
- On proxy quality, the only reference is from NAS: EZNAS-A reaches Kendall τ = 0.65 on NASBench-201 CIFAR-10, versus 0.57 for Synflow and NASWOT.
We found no reported result on Lacuna for proxy-to-generation rank correlation on steering configurations.
What counts as beating the baselines. Our method beats a baseline if the $(B,F)$ shortlist reaches at least 90% of exhaustive-search hypervolume at a quarter or less of successive halving's budget, on behaviors and models it was not developed on.
Ablation. The key ablation replaces the two-objective front with $B$ alone, followed by picking the largest strength whose perplexity stays under a threshold. This isolates whether modeling the fluency cost as a separate objective is what matters.
Why not just…?
Why not just run a small grid search with cheap generations? This is the real threat, and successive halving is exactly that. Prior work does not rule it out. The out-of-distribution failures reported for contrastive activation addition apply to any search tuned in-distribution, proxy-based or not. Two things could still favor a proxy. First, very short generations may not show collapse, which tends to appear later in decoding. Second, generation-based filtering pays judge costs at every stage. The matched-budget hypervolume curves will settle the question. If successive halving matches the proxy, we will report that and recommend combining them, with the proxy used as the first rung of the ladder. The benchmark remains useful either way.
Why not just use DIRECTER's or AUSteer's built-in localization? Both are proxies whose ranking quality has never been checked against exhaustive outcomes. Both also output a single score, while the file documents a two-sided trade-off: ITI's truthfulness–informativeness tension and the universal perplexity penalty. Their correlations appear directly in our tables. If they already recover the front, that is a publishable positive result for them, and our $(B,F)$ proxy becomes unnecessary.
Why not just use dynamic steering and stop tuning static strengths? PID-Act leaves its gains as manual settings and names automatic tuning as open. DIRECTER's threshold and DSAS's detector placement are also hand-set. The dynamic-method slice of the benchmark measures directly whether these methods shrink outcome variance or only move it into new settings.
Why not just learn steering sites differentiably with a sparsity penalty, or fine-tune with LoRA? A differentiable mask reintroduces the problem that Differentiable Neural Network Architecture Search reports: task and sparsity weights $\lambda_1, \lambda_2$ that must be hand-set, plus a gap between the continuous and pruned models. That is multi-objective loss-weight sensitivity moved one level up. Fine-tuning gives up weight-free, toggleable control. We treat the differentiable-mask variant as out of scope rather than claiming it would fail.
Decisive Experiments
Rank quality. We compute within-model rank correlations for each proxy on each behavior. The cosine-versus-norm question from DIRECTER is decided by comparing its cosine score against a norm score of matched cost.
Transfer. Proxies are developed on truthfulness in Qwen2.5-1.5B and frozen. We then report front recovery on toxicity and IFEval, and on Gemma-2-2B and the Llama-3.1-8B slice. The central claim fails if held-out hypervolume drops below that of random search at equal budget.
Top-decile resolution. This tests whether $F$ fixes the NAS-style failure of proxies to distinguish among the best candidates. The prediction is that it will. Near the front, configurations differ mainly in hidden damage to fluency, which a behavior-only score cannot see.
Out-of-distribution robustness. Configurations are selected on in-distribution contrastive pairs and scored on shifted prompt styles. This checks whether proxy-selected configurations are more or less brittle than configurations selected by generation.
Release. The released benchmark lets future steering-search methods be compared without regenerating millions of outputs. Given how much steering practice is currently hand-tuned, that may be the most lasting contribution.
References
- EZNAS: Evolving Zero-Cost Proxies For Neural Architecture Scoring
- Enhancing Instruction Following of LLMs via Activation Steering with Dynamic Rejection
- Fine-Grained Activation Steering: Steering Less, Achieving More
- Patterns and Mechanisms of Contrastive Activation Engineering
- Dynamically Scaled Activation Steering
- Activation Steering with a Feedback Controller
- Improving Reasoning Performance in Large Language Models via Representation Engineering
- TruthfulQA: Measuring How Models Mimic Human Falsehoods
- Rethinking Performance Estimation in Neural Architecture Search
- Inference-Time Intervention: Eliciting Truthful Answers from a Language Model
- Differentiable Neural Network Architecture Search