Quality Steering in Language Models

• Quality Steering is an activation control strategy that adjusts hidden states using reference completions to boost factual response accuracy.
• It computes prompt-specific reference activations and uses additive steering plus interpolation across transformer layers to steer model outputs.
• Segmented control over early, middle, and late layers enables tailored tuning of factuality versus fluency, optimizing performance on challenging QA tasks.

Searching arXiv for the target paper and closely related activation steering work for citation support. Fusion Steering is an inference-time activation steering method for LLMs that targets factual question answering by injecting prompt-specific activation deltas across the full transformer stack, without changing model weights. Its defining move is to steer the model toward the internal trajectory associated with a reference completion that contains the question, the ground-truth answer, and a model-generated explanation, thereby treating factuality, relevance, and coherence as properties of internal state geometry rather than only of prompts or logits (Chang et al., 28 May 2025).

1. Conceptual basis

Fusion Steering is framed as a form of activation-level quality steering. The method assumes that, for a given question, the model’s hidden states under a bare prompt can be profitably nudged toward the hidden states it would have exhibited had it already produced the correct answer together with an explanation. In that sense, the steering signal is not a generic behavioral direction such as sentiment or refusal, but an example-specific direction tied to a concrete fact and its semantic neighborhood (Chang et al., 28 May 2025).

This differs from traditional representation steering methods that typically construct a single vector from contrastive examples and inject it at one layer or at a fixed set of layers with a global scale. Fusion Steering instead computes prompt-specific reference activations for each QA instance, applies additive steering plus interpolation across all transformer layers, and offers both a uniform full-layer regime and a segmented regime with separate controls for early, middle, and late layers. The method therefore treats “where” steering is applied as part of the optimization problem rather than as a fixed design choice (Chang et al., 28 May 2025).

The paper operationalizes output quality along multiple axes. During optimization, token overlap with the reference answer is used as a proxy for factual alignment and perplexity as a proxy for fluency. During evaluation, token overlap is combined with an LLM-based rubric score covering correctness, completeness, and relevance, and performance is also reported under the stricter SimpleQA correctness rubric. A plausible implication is that Fusion Steering should be read not merely as a factuality intervention, but as a multi-objective control mechanism over factual alignment and answer quality (Chang et al., 28 May 2025).

2. Reference completions and activation mechanics

For each question, Fusion Steering constructs an enriched reference prompt of the form:

Question: $[prompt]$ Answer and Explanation: $[ground\ truth + explanation]$

The explanation is generated by the model after being given the question and ground-truth answer. The stated purpose is twofold: the answer tokens anchor the specific factual content, while the explanation expands the representation into a richer semantic neighborhood. The resulting reference trajectory is intended to capture what the model’s internal state looks like when it is in a knowledgeable, explanatory mode about the correct fact (Chang et al., 28 May 2025).

Let $\mathcal{T}$ denote the token positions corresponding to the ground-truth answer and explanation. At each layer $l$, the reference activation is the average over those positions: $$h_l \in \mathbb{R}^d,\qquad h_l = \frac{1}{|\mathcal{T}|}\sum_{t \in \mathcal{T}} a_{l,t}.$$

At inference time, the model is given only the question. Let $S$ be the token positions in that question, and let the layerwise mean activation on the bare prompt be

$$\mu_l = \frac{1}{|S|}\sum_{t \in S} a^{\mathrm{original}}_{l,t}.$$

The steering delta at layer $l$ is then defined as

$\Delta_l = \gamma \bigl(h_l - \mu_l\bigr),$

where $\gamma$ is the steering strength. This delta is added to every token activation at that layer: $[ground\ truth + explanation]$0

Fusion Steering then interpolates between original and steered activations using a fusion weight $[ground\ truth + explanation]$1: $[ground\ truth + explanation]$2

In this parameterization, $[ground\ truth + explanation]$3 governs how far the representation is shifted toward the reference trace, while $[ground\ truth + explanation]$4 governs how much of that shifted state is actually used. The method is therefore an interpolated combination of activation patching and additive steering (Chang et al., 28 May 2025).

3. Full-network and segmented control

Fusion Steering is implemented on Gemma-2-2B-IT, whose layers are indexed as $[ground\ truth + explanation]$5. The paper defines two steering regimes. In full-layer steering, a single pair $[ground\ truth + explanation]$6 is used for all layers. In segmented steering, the network is partitioned into three layer groups—early $[ground\ truth + explanation]$7, middle $[ground\ truth + explanation]$8, and late $[ground\ truth + explanation]$9—and each segment receives its own $\mathcal{T}$0. The same update equations apply, but the coefficients become segment-specific (Chang et al., 28 May 2025).

This segmentation is motivated by the claim that adjacent transformer layers often cluster into functional groups. The paper presents segmented steering as a way to push different parts of the model with different intensities, for example more aggressively in early layers and more conservatively in late layers, and to tune that trade-off per prompt. This suggests that the method is not only prompt-specific but also depth-sensitive in its control policy (Chang et al., 28 May 2025).

Parameter selection is done per prompt with Optuna. For full-layer steering, the search space is $\mathcal{T}$1. For segmented steering, the search space contains three such pairs, one for each segment. The paper runs 5 Optuna trials per prompt and keeps the configuration maximizing

$\mathcal{T}$2

with $\mathcal{T}$3. Token overlap is computed against the SimpleQA ground truth, while normalized perplexity is measured by GPT-2. The fixed value $\mathcal{T}$4 makes the optimization explicitly factuality-leaning rather than fluency-leaning (Chang et al., 28 May 2025).

4. Evaluation protocol and empirical performance

Evaluation is conducted on SimpleQA, a short-form factual QA benchmark with 4,326 prompts across 10 topic categories. The study samples 500 prompts from the test split, grades Gemma-2-2B-IT baseline answers with a Mistral Small 3.1 factuality classifier, and then selects the 260 prompts labeled INCORRECT for the steering experiments. The use of baseline failures only makes the evaluation a targeted stress test rather than a measure of average-case QA behavior (Chang et al., 28 May 2025).

Two LLM-based graders are used. The first is a strict factuality classifier with outputs CORRECT, INCORRECT, and NOT_ATTEMPTED, used both for subset selection and for strict evaluation. The second is a 1–5 quality rubric covering correctness, relevance, and completeness; it also flags hallucinations and logs notes, though those notes are not included in the composite metric. For reporting, the paper defines

$\mathcal{T}$5

and counts a response as “accurate” when this score is at least $\mathcal{T}$6 (Chang et al., 28 May 2025).

Using Gemma-2-2B-IT with 8-bit quantization, segmented steering produces the strongest gains on this hard subset. The paper also reports that token overlap distributions improve, but that the grade distribution shows both more high-quality responses and more severe failures, indicating a sharper factuality–fluency tension under stronger interventions (Chang et al., 28 May 2025).

Method	Composite accuracy $\mathcal{T}$7	SimpleQA CORRECT
Baseline	3.5% (9/260)	0.0% (0/260)
Full-layer steering	16.2% (42/260)	9.2% (24/260)
Segmented steering	25.4% (66/260)	13.1% (34/260)

Beyond these headline numbers, token overlap at threshold $\mathcal{T}$8 rises from 20.4% for the baseline to 37.3% for full-layer steering and 43.5% for segmented steering. Under the authors’ interpretation, the segmented regime outperforms the full-layer regime because differential control over early, middle, and late layers is more effective than a uniform push across depth (Chang et al., 28 May 2025).

5. Interpretability and relation to subsequent steering work

The paper explicitly states that Fusion Steering is amenable to sparse representations such as Neuronpedia-style sparse autoencoder features and sparse crosscoders. In its present form, the method injects dense reference activation vectors $\mathcal{T}$9, but the authors propose replacing those with projections onto sparse, interpretable feature bases. A plausible implication is that the method could evolve from dense example-level steering into concept-level steering in which only selected factual or anti-hallucinatory features are activated (Chang et al., 28 May 2025).

That future direction aligns with later sparse and fine-grained steering work. Sparse Representation Engineering decomposes dense activations into monosemantic sparse features and edits only task-specific sparse dimensions, reporting superior controllability while maintaining generated content quality on safety, fairness, and truthfulness tasks (He et al., 21 Mar 2025). AUSteer argues that block-level activations entangle beneficial, irrelevant, and harmful features, and that steering selected atomic units rather than full blocks yields more precise and effective control while changing far fewer activations (Feng et al., 4 Feb 2026). These developments do not modify Fusion Steering itself, but they sharpen the interpretability agenda already present in its discussion of sparse feature spaces.

Later work also addressed the adaptivity problem from different angles. ATLAS introduces a lightweight latent verifier that predicts step-level reasoning quality and adaptively chooses among execution, reflection, transition, or no steering at thought boundaries, thereby replacing fixed policies with per-step control (Nguyen et al., 6 Jan 2026). Future Probe Controlled Generation distinguishes detection features from prediction features and selects among candidate next sentences using probes that predict future behavioral likelihood, reporting steering with almost no output quality degradation in several reasoning-model settings (Kortukov et al., 9 Jun 2026). Relative to those approaches, Fusion Steering remains distinctive in being example-specific, answer-conditioned, and full-network, with reference states built from ground-truth answer-plus-explanation traces rather than from generic latent modes (Chang et al., 28 May 2025).

6. Limitations and significance

Fusion Steering has several explicit constraints. It requires ground-truth answers and generated explanations to build reference completions, which limits direct applicability in settings where no reference answer is available at inference time. It also performs Optuna-based tuning per prompt, which is computationally intensive. The experiments use greedy decoding with temperature $l$0, a choice that stabilizes evaluation but also exacerbates brittleness; the paper notes examples of repetitive or incoherent steered outputs such as repeated “7”. The empirical study is further limited to 260 prompts, one baseline-failure subset, one model family, 8-bit quantization, and no out-of-domain or adversarial robustness analysis (Chang et al., 28 May 2025).

The central trade-off is factuality versus fluency. Because the optimization metric is weighted toward token overlap with $l$1, some outputs become more factually aligned while becoming less fluent or less coherent. The grade distribution reflects this: steering increases the number of grade 4–5 answers but also the number of grade 1 failures. The method therefore exposes rather than eliminates the tension between semantic correctness and generation stability (Chang et al., 28 May 2025).

Even with those limitations, Fusion Steering occupies a distinct place in the representation-engineering literature. It demonstrates that per-prompt, full-network activation control can substantially improve factual QA on cases where the baseline model is already known to fail, and that segmented intervention outperforms uniform intervention on the same model and subset. More broadly, it establishes a concrete template for inference-time control in which reference completions define semantically enriched target trajectories and the optimization problem is phrased directly in terms of factual alignment, fluency proxies, and rubric-based answer quality (Chang et al., 28 May 2025).