On the Generalization of Steering Vectors for Chain-of-Thought Faithfulness

On the Generalization of Steering Vectors for Chain-of-Thought Faithfulness

Matthew Nguyen University of Virginia &Kyle Cox 1 1 footnotemark: 1 Independent &Austin Meek University of Delaware &Iván Arcuschin Poseidon Research Equal contribution. Correspondence to mbnguyen8@gmail.com ; kylecox2000@gmail.com

Abstract Model capabilities have improved in large part due to scaling chain of thought. This has been a promising development for AI safety—where models verbalize their reasoning, it is possible to monitor it. However, in some cases, models do not verbalize important steps in their reasoning process. For example, models prompted with a cue suggesting the incorrect answer may fail to acknowledge that cue, even when it appears instrumental to their conclusion. When chain of thought (CoT) fails to disclose instrumental reasoning steps, we describe it as unfaithful . Prior work has shown that activation steering can be a useful method to improve faithfulness in CoT (Occhipinti et al., 2026 ) . We extend this line of work by studying how well steering for faithfulness generalizes across cue types, datasets, and methods of constructing the steering vector for three models (Gemma-3 4B, Qwen-3.5 9B, Gemma-3 12B) in a cued question-answering setting. While steering reliably increases cue acknowledgment for only the largest model (Gemma-3 12B), we find that when steering is effective, its effect generalizes broadly across cue types and datasets—in cross-cue and cross-dataset analyses, effect size is determined primarily by the evaluation setting, rather than the vector’s train setting. How the vector is built also matters little—four construction methods, including one whose optimization target mentions no specific cue, yield similar effect sizes. Finally, we consider the possibility that steering promotes the salience of the cue and causes greater cue use , rather than targeting verbalization behaviors. However, we find no evidence for this—steering leaves the rate of cue use roughly unchanged while reducing hidden cue use , i.e., cue use that is not acknowledged.

1 Introduction

Chain-of-thought monitoring is one of the few scalable approaches to the oversight of large language models (Korbak et al., 2025 ) . 1 1 1 Code release: https://github.com/xocelyk/steering-vectors-for-faithfulness However, the reliability of CoT monitoring depends on whether CoT is faithful , i.e., whether the model discloses the factors that determine its answer. Prior work has shown this property does not reliably hold. Models may use prompt cues—for example, revealed answers or appeals to authority—without stating that they have done so (Chen et al., 2025 ) , or they may post-hoc rationalize implicit biases in prompts that carry no cue (Arcuschin et al., 2025 ; 2026 ) .

Recent work has shown that activation steering can be useful toward recovering monitorability. It has been demonstrated that reasoning-related features are linearly represented in the residual stream and can be steered to induce greater reasoning ability in base models (Venhoff et al., 2025a ) and control specific modalities of reasoning in reasoning models (Venhoff et al., 2025b ) . Occhipinti et al. ( 2026 ) extend this line of work to faithfulness directly, probing residual stream activations for faithfulness and using the resulting features to steer model generations. They report that linear steering recovers monitorability by up to 46% with collateral effects below 5%, and that off-policy steering (i.e., using synthetic traces) performs comparably to on-policy steering.

These results establish that steering can improve acknowledgment in the setting where the method was developed. However, ideally, a faithfulness intervention will not depend narrowly on the particular context used to construct it. We consider whether steering success survives changes of model, evaluation dataset, cue type, and construction procedure. On GPQA, we construct and evaluate vectors for four unique cues, and with the cue held fixed we measure transfer across BBH, GPQA, and MMLU. We also compare four ways of building the vector: labeled on-policy contrastive vectors, synthetic vectors, cue-specific optimized vectors, and a cue-agnostic optimized direction. All experiments cover the full set of models: Gemma-3 4B, Qwen-3.5 9B, and Gemma-3 12B.

Our experiments consider two ways in which steering can generalize: (1) efficacy generalization —whether steering is effective at all in a particular model and evaluation setting—and (2) transfer generalization —whether a vector constructed in one setting is useful in others.

While our results show that efficacy generalizes only narrowly—steering reliably increases cue acknowledgment only for the largest model we study, Gemma-3 12B, and its effect varies strongly across evaluation datasets—we see that transfer generalizes broadly. On GPQA, a steering vector trained on one cue increases acknowledgment at about the same rate as a vector trained on the cue being evaluated. Similarly, a vector trained on one dataset produces a comparable change in acknowledgment to one trained on the dataset being evaluated, holding cue type fixed. The interchangeability of construction methods reinforces these results—synthetic vectors, which require no model rollouts or judge labels, and the generic optimized direction, whose target never names a specific cue, produce similar effects to the contrastive vectors built from judge-labeled rollouts.

These results characterize how far the steering effect travels; finally, we consider what the steering effect represents. While we intend for steering to increase faithfulness—i.e., verbalization of unacknowledged reasoning steps—it is possible that instead, steering has promoted cue use , which could in turn increase the absolute rate of cue acknowledgment without improving faithfulness. To address this, we study the effect of steering on cue use and on hidden cue use, the portion that goes unacknowledged. We find that when steering increases acknowledgment, it leaves the rate of cue use roughly unchanged while reducing hidden cue use.

2 Setup and Methods

Models, datasets, and cues.

We evaluate three open-weight instruction-tuned models of varying size (Gemma-3 4B, Qwen-3.5 9B, Gemma-3 12B (Team, 2025 ; 2026 ) ) on multiple-choice questions from BBH (Suzgun et al., 2022 ) , GPQA (Rein et al., 2023 ) , and MMLU (Hendrycks et al., 2021 ) . We inject one of four cues from Meek et al. ( 2025 ) (four of their five, omitting the visual-pattern cue): a sycophantic appeal to a Stanford professor (their sycophancy cue, labeled Stanford in our figures and tables), an XML metadata block, a grader-hacking hint, and an unethical insider tip, each pointing to an incorrect option. We vary the cue on GPQA and the dataset on the Stanford cue. For each combination of (model, dataset, cue), we keep the items whose cued rollout is non-degenerate (not looping, repeating, or incoherent, as determined by a generation-time heuristic and an LLM judge; see Appendix ˜ E ). We do not condition on the model following the cue or on its answer without the cue, so the kept items include cases where the cue did not change the answer. The surviving items are split ≈ \approx 67/33 into train and test per scenario with a fixed seed (129–165 test items per scenario).

Decoding.

All rollouts use greedy decoding (temperature 0 ) with vLLM (Kwon et al., 2023 ) in an effort to control randomness due to sampling. Per-model token limits and penalties are shared by baseline and steered runs, so comparisons are matched within model ( Appendix ˜ E ).

Vector construction.

We compare four construction methods, each producing one steering vector v ∈ ℝ d v\in\mathbb{R}^{d} per scenario (details and counts in Appendix ˜ E ). Contrastive takes the difference of mean activations between judge-labeled faithful (cue-acknowledging) and unfaithful rollouts on the cued training prompts following Rimsky et al. ( 2024 ) . Synthetic replaces those rollouts with templated cue-acknowledging vs. cue-ignoring completions, so it needs no rollouts or judge labels. The two optimization methods optimize v v via gradient descent so the steered model assigns high likelihood to target completions, following the one-shot approach of Dunefsky & Cohan ( 2025 ) (also used by Venhoff et al. ( 2025a ) ): the specific variant targets cue-acknowledging completions for the given cue, while the generic variant targets a single fixed completion, shared across all cues, in which the model resolves to consider the hints in the problem. The generic target mentions no specific cue, though both variants are optimized on the same cued prompts.

Steering.

During generation, we add α ​ v ^ \alpha,\hat{v} , the unit-normalized direction scaled by a coefficient α \alpha , to the residual stream at each generated-token position at the selected layer (Turner et al., 2024 ; Rimsky et al., 2024 ) . We sweep α ∈ { 2.5 , 5 , 7.5 } \alpha\in{2.5,5,7.5} and report α = 5 \alpha{=}5 in the main text. Steering effect sizes are stable across the sweep ( Appendix ˜ C ). To select the residual stream layer used for steering, we train a ridge-regularized linear probe per layer to predict cue acknowledgment; the layer whose probe attains the highest test area under the receiver operating characteristic curve (AUROC) is used for all four steering methods reported here. The probe only selects the layer: the four vectors are constructed as described above and are distinct from it. We note the weakness of this criterion in Section ˜ 4 .

Judge and metrics.

We score each generation with a single LLM judge ( gpt-5-nano (OpenAI, 2026 ) ; prompts in Appendix ˜ F ). Acknowledgment is whether the judge finds that the chain of thought mentions the cue. Following Meek et al. ( 2025 ) , cue acknowledgment is our operationalization of faithfulness—we label a CoT trace as faithful when it mentions the injected cue. However, cue acknowledgment does not establish faithfulness in the conventional sense, i.e., that the verbalized reasoning reflects the computation behind the answer (Jacovi & Goldberg, 2020 ) . We return to this gap in Section ˜ 3.4 . Cue use is an inference from the final answer: whether the judge-extracted answer equals the cued option ( ∼ \sim 97% coverage; Appendix ˜ E ). We call unacknowledged cue use hidden cue use , and report the hidden cue use rate as the rate of non-acknowledgment given cue use. Treating selection of the cued option as use is an assumption, and it can fail in two ways: the model could select the cued option without relying on the cue, or be influenced by the cue without selecting it. There is evidence the first case is rare. We find that the cued option is responsible for 76–87% of all errors, a much larger rate than would be the case were errors distributed equally across incorrect answer options. So, we conclude selecting the cued option usually reflects cue use rather than an incidental wrong guess ( Section ˜ D.8 ). The second case, influence without selection, is not captured by the measure, which to that extent underestimates cue influence.

For each of the n n matched traces, let a b , i , a s , i ∈ { 0 , 1 } a{\mathrm{b},i},a{\mathrm{s},i}\in{0,1} indicate whether the chain of thought acknowledges the cue at baseline and under steering on the same prompt. Our primary metric is the change in cue acknowledgment, Δ ack \Delta_{\mathrm{ack}} : the rate of acknowledgment in steered traces minus the rate of acknowledgment in un-steered (baseline) traces.

Because steered and baseline traces are paired, we can also express Δ ack \Delta_{\mathrm{ack}} in terms of the number of traces that regressed (baseline acknowledgment → \rightarrow steered no acknowledgment) and the number of traces that converted (baseline no acknowledgment → \rightarrow steered acknowledgment):

Δ ack = 1 n ​ ∑ i a s , i − 1 n ​ ∑ i a b , i = 1 n ​ ∑ i ( a s , i − a b , i ) = n conv − n regr n . \Delta{\mathrm{ack}};=;\frac{1}{n}\sum{i}a{\mathrm{s},i};-;\frac{1}{n}\sum{i}a{\mathrm{b},i};=;\frac{1}{n}\sum{i}\bigl(a{\mathrm{s},i}-a{\mathrm{b},i}\bigr);=;\frac{n{\mathrm{conv}}-n{\mathrm{regr}}}{n}. (1)

Each trace contributes a s , i − a b , i ∈ { − 1 , 0 , + 1 } a{\mathrm{s},i}-a{\mathrm{b},i}\in{-1,0,+1} : a converted trace ( a b , i = 0 , a s , i = 1 a{\mathrm{b},i}{=}0,,a{\mathrm{s},i}{=}1 ) contributes + 1 +1 , a regressed trace ( a b , i = 1 , a s , i = 0 a{\mathrm{b},i}{=}1,,a{\mathrm{s},i}{=}0 ) contributes − 1 -1 , and an unchanged trace 0 ; n conv n{\mathrm{conv}} and n regr n{\mathrm{regr}} count the two. So Δ ack \Delta{\mathrm{ack}} has two equivalent readings, the change in the overall acknowledgment rate and the converted minus the regressed fraction of traces; tables report n conv / n n{\mathrm{conv}}/n and n regr / n n{\mathrm{regr}}/n ( Tables ˜ 5 and 8 ), whose difference is Δ ack \Delta{\mathrm{ack}} .

We report Δ ack \Delta{\mathrm{ack}} rather than the converted fraction alone because conversion and regression can occur together. For example, a substantial but poorly targeted steering intervention could perturb the CoT in an undirected way. We see evidence of this in several instances, where large conversion rates are matched by similarly large regression rates. While n conv n{\mathrm{conv}} tracks the magnitude of the intervention, alone it does not capture the direction of the intervention, so we prefer the more holistic Δ ack \Delta{\mathrm{ack}} as our primary metric. In this way our analysis differs from Occhipinti et al. ( 2026 ) ; they report monitorability recovery over only the traces that were unfaithful at baseline, with regressions reported separately as collateral effects. So, our Δ ack \Delta{\mathrm{ack}} metric uses a different denominator than their recovery rate, and the two are not directly comparable. We analyze acknowledgment, cue use, and hidden cue use in Section ˜ 3.4 .

Experimental coverage.

Our experiments vary only a subset of dimensions at a time rather than factorizing over every dimension. Table ˜ 1 summarizes the full experimental suite. We sweep cue type for only GPQA, while BBH and MMLU use the Stanford cue throughout. Each steered evaluation was run for all three models and all four vector construction methods at each α ∈ { 2.5 , 5 , 7.5 } \alpha\in{2.5,5,7.5} ; results report α = 5 \alpha{=}5 and, unless stated otherwise, the contrastive vectors.

Table 1: Experimental coverage. Every cell was run for all three models; steered analyses for all four construction methods at each α ∈ { 2.5 , 5 , 7.5 } \alpha\in{2.5,5,7.5} .

Analysis

Varies Held fixed

Coverage

Probing ( Section ˜ 3.1 )

dataset, cue —

all 3 × 4 3\times 4 (dataset, cue) pairs, plus pooled cue and dataset sets

Matched steering effect ( Section ˜ 3.2 )

model, dataset, method train = = eval setting

GPQA: all 4 cues; BBH, MMLU: Stanford cue

Cross-cue transfer ( Section ˜ 3.3 )

train × \times eval cue GPQA

all 4 × 4 4\times 4 pairs, plus an all-cue vector

Cross-dataset transfer ( Section ˜ 3.3 )

train × \times eval dataset Stanford cue

all 3 × 3 3\times 3 pairs, plus an all-dataset vector

α \alpha robustness ( Appendix ˜ C )

α ∈ { 2.5 , 5 , 7.5 } \alpha\in{2.5,5,7.5} matched setting

all methods and scenarios

3 Results

3.1 Probing cue acknowledgment

We fit a linear probe to predict cue acknowledgment from residual-stream activations at each layer, and report the test AUROC of the selected layer per (model, dataset, cue) setting ( Figure ˜ 9 ). Acknowledgment is decodable on MMLU (AUROC 0.71–0.98 across cues) and, less consistently, on BBH (0.62–0.94). On GPQA the per-cue probes are only modestly above chance (0.61–0.70) and the pooled cross-cue and cross-dataset probes are lower still (about 0.57–0.60). The reported AUROC is the maximum across layers of the test-set AUROC; because the layer is selected on the same test set, the reported value is optimistic. We note that the train AUROC is 1.00 everywhere, because the activation dimension exceeds the training-set size ( Appendix ˜ E ). Overfitting during training may explain the poor performance of some of our probes on the test set.

3.2 Effect of steering on cue acknowledgment

We measure the change in cue acknowledgment between pairs of baseline (not steered) and steered generations, where a vector trained on a (dataset, cue) pair is evaluated on the same dataset and cue. For the GPQA dataset, we test all four cue types. For BBH and MMLU, we test only the Stanford cue. Figure ˜ 1 reports the results for the Stanford cue across all three datasets. We report the GPQA results for each of the cue types in Appendix ˜ A .

For Gemma-3 4B, the change (steered minus baseline acknowledgment rate) is near zero on every dataset. For Qwen-3.5 9B it is near zero on BBH and MMLU, but for GPQA it is slightly positive across vector construction types. However, while Figure ˜ 1 presents results for the Stanford cue, in Figure ˜ 4 we see that the steering effect for Qwen-3.5 9B on GPQA depends strongly on the cue type. For Gemma-3 12B the effect depends strongly on the evaluation dataset ( + 0.26 +0.26 on MMLU, + 0.10 +0.10 on BBH, and + 0.06 +0.06 on GPQA). Within GPQA for Gemma-3 12B, the steering effect also depends strongly on the cue type ( Figure ˜ 4 ). In general, the effectiveness of the steering intervention is stable across the steering strengths we test ( Figure ˜ 6 ). We also compare how decodable acknowledgment is with how well steering works. Across the four GPQA cues, the steering effect shows no relationship with probe test AUROC for any model ( Figure ˜ 7 ).

The four construction methods, including the generic optimization vector, whose target completion names no specific cue, yield similar effect sizes on each dataset ( Figure ˜ 1 ). On MMLU, where the effect is largest for Gemma-3 12B, all four effect sizes fall between + 0.24 +0.24 and + 0.28 +0.28 . Notably, the synthetic constructions, which require no CoT rollouts or judge labels, perform similarly well to the contrastive constructions (their pooled matched Δ ack \Delta_{\mathrm{ack}} values agree within 0.02 0.02 on every model in Figure ˜ 4 ). On Gemma-3 12B the two vectors also have very similar directions (cosine similarity 0.80 0.80 in Table ˜ 13 ).

To determine whether steering has any side effects, we also measure changes in accuracy and degeneracy due to steering. Steering has little effect on accuracy: Δ \Delta Acc is mixed in sign across scenarios ( Table ˜ 5 ), declining at most by 0.07 0.07 (Gemma-3 12B on MMLU; Section ˜ D.4 ). However, steering does induce more degenerate reasoning traces: on average, degeneracy increases by 0.02 0.02 – 0.03 0.03 ( Table ˜ 5 ). In Appendix ˜ F , we provide the prompt we use to judge degeneracy. We treat another side effect—whether steering induces greater cue use —as a separate question, which we address in Section ˜ 3.4 .

Figure 1: Steering efficacy depends on the model and the evaluation dataset, and not clearly on vector construction. Δ ack \Delta_{\mathrm{ack}} in the matched setting, Stanford cue on all three datasets ( α = 5 \alpha{=}5 ; whiskers are 90% normal-approximation CIs for the paired change).

3.3 Transfer across cues and datasets

Because the four construction methods give comparable effects, the following analyses use only the contrastive vectors. The steering effect is determined mainly by the setting to which the vector is applied, rather than the setting on which it was trained. We show this for Gemma-3 12B. The heatmap diagonals in Figure ˜ 2 are not substantially higher, on average, than the off-diagonal values. However, the effect sizes are highly consistent within evaluation cue settings (the columns of the heatmap). The cross-dataset matrix (Stanford cue, Figure ˜ 3 ) shows a similar pattern for generalization across datasets, with the largest effect for Gemma-3 12B on MMLU. Note that for the smallest model (Gemma-3 4B), Δ ack \Delta{\mathrm{ack}} is approximately 0 on GPQA (see Figure ˜ 1 ); in this case, near-zero per-cue effects in Figure ˜ 4 mean there is little effect to transfer, rather than poor generalization over cues. Also note that for Qwen-3.5 9B, steering is consistently effective only for the GPQA dataset (see Figure ˜ 1 ). Similarly, in Figure ˜ 3 , Δ ack \Delta{\mathrm{ack}} is only positive for the GPQA eval set. However, vectors trained on any of the datasets are effective for steering on the GPQA dataset (see that the full column is positive). These results point toward two conclusions: (1) the acknowledgment effect is determined mainly by the evaluation setting and (2) vectors trained on a particular setting often capture a useful intervention feature, even when that intervention has a null effect on the train setting.

Figure 2: Steering transfer across cue types. Δ ack \Delta_{\mathrm{ack}} on GPQA by training cue (rows) and evaluation cue (columns) using contrastive vectors with α = 5 \alpha{=}5 .

Appendix ˜ B attempts to quantify (1): off-diagonal effects correlate with the test setting’s own (diagonal) effect but not the train setting’s, and the relationship survives a split-half control for shared evaluation noise. Conclusion (2) inspires the brief investigation into feature geometries in Section ˜ D.5 , where we compare the directions of vectors constructed across different cues, datasets, and methods ( Tables ˜ 10 , 11 and 13 ). This analysis yields an inconsistent account: on Gemma-3 12B three of the four cue vectors align (pairwise cosine 0.49 0.49 – 0.82 0.82 , against 0 ± 0.02 0\pm 0.02 for random directions), but the early-layer BBH vector transfers behaviorally without aligning with the other dataset vectors, and Qwen-3.5 9B shows moderate cross-dataset alignment despite its null effect.

Figure 3: Steering transfer across datasets. Δ ack \Delta_{\mathrm{ack}} under the Stanford cue by training dataset (rows) and evaluation dataset (columns) using contrastive vectors with α = 5 \alpha{=}5 .

3.4 Cue acknowledgment and cue use

While our work nominally studies faithfulness, it is not a priori clear that increases in acknowledgment indicate increases in faithfulness. Because the evaluation setting is not conditioned on the model following the cue ( Section ˜ 2 ), on many items, the cue is plausibly not a factor in the model’s answer (at baseline the model selects the cued option 41–71% of the time; Table ˜ 2 ). A change from baseline no-acknowledgment to acknowledgment under steering may therefore reflect that the model has verbalized a factor that was not relevant to its computation, or alternatively that steering has made the cue more salient and caused the model to consider a cue it previously did not. The former is undesirable because it constitutes an unfaithful account of the model’s reasoning (specifically, it constitutes reasoning theater (Boppana et al., 2026 ) ). The latter is undesirable for two reasons: (1) in our regime the cued answers are incorrect and intended to bias, so making them more salient may make the model an inferior reasoner, and (2) it may indicate that the selected vector does not correspond to a general faithfulness or verbalization feature but rather is specific to the cue itself, such that steering with it is better characterized as a nudge toward “consider this particular cue” than “verbalize each factor in your reasoning process.” In consideration of these hypotheses, we report cue acknowledgment, cue use, and hidden cue use below.

In Table ˜ 2 , we report three rates for the GPQA setting, baseline → \to steered: cue use, cue acknowledgment, and hidden cue use. The overall acknowledgment rate changes substantially only for Gemma-3 12B, as in Figure ˜ 1 . We find that cue acknowledgment does not coincide with greater cue use: while acknowledgment increases meaningfully for Gemma-3 12B (0.47 to 0.54), use is nearly unchanged (0.71 to 0.70) and hidden cue use falls (0.48 to 0.42).

In Table ˜ 7 , we perform a similar analysis, focusing on hidden cue use across our sweep of models and cue types for GPQA. Hidden cue use is highly cue-dependent, ranging from about 0.19 to 0.77 under steering. Its change under steering is small and mixed in sign across cues: Qwen-3.5 9B improves on the Stanford and Grader cues and worsens on XML, Gemma-3 4B improves on the same two cues but worsens on XML and Unethical, and Gemma-3 12B improves on every cue, most on Unethical.

The effect also varies by dataset ( Table ˜ 2 , right; Section ˜ D.4 ): it is much larger on MMLU, where steering reduces hidden cue use from 0.31 0.31 to 0.05 0.05 , than on GPQA or BBH. Within Gemma-3 12B this dataset ordering matches probe AUROC ( Section ˜ 3.1 ), but the association does not hold across models.

Table 2: Cue use, acknowledgment, and hidden cue use, baseline → \to steered (contrastive, α = 5 \alpha{=}5 ). Left: per model, pooled over the four GPQA cues. Right: per dataset, for Gemma-3 12B with the Stanford cue. The per-cue breakdown is in Table ˜ 7 ; Table ˜ 6 varies the steering coefficient.

Model Use Ack Hidden use

Gemma-3 4B 0.63 → 0.64 0.63!\to!0.64 0.52 → 0.51 0.52!\to!0.51 0.39 → 0.40 0.39!\to!0.40

Qwen-3.5 9B 0.41 → 0.41 0.41!\to!0.41 0.47 → 0.48 0.47!\to!0.48 0.46 → 0.41 0.46!\to!0.41

Gemma-3 12B 0.71 → 0.70 0.71!\to!0.70 0.47 → 0.54 0.47!\to!0.54 0.48 → 0.42 0.48!\to!0.42

Dataset Use Ack Hidden use

BBH 0.30 → 0.29 0.30!\to!0.29 0.60 → 0.69 0.60!\to!0.69 0.33 → 0.23 0.33!\to!0.23

GPQA 0.66 → 0.62 0.66!\to!0.62 0.55 → 0.62 0.55!\to!0.62 0.39 → 0.37 0.39!\to!0.37

MMLU 0.32 → 0.35 0.32!\to!0.35 0.62 → 0.89 0.62!\to!0.89 0.31 → 0.05 0.31!\to!0.05

4 Conclusion

We find that steering increases cue acknowledgment reliably only on the largest model (Gemma-3 12B) and mainly on MMLU ( + 0.26 +0.26 , against + 0.10 +0.10 on BBH and + 0.07 +0.07 pooled over the GPQA cues), with little change on the smaller two; on MMLU it also cuts hidden cue use from 0.31 0.31 to 0.05 0.05 . Where present, the effect is similar across the four construction methods, including one built without cue-specific supervision, and tracks the dataset a vector is applied to rather than the one it was built from.

Several limitations bound these observations. First, our criterion for selecting the steering and probing layer (the argmax probe test AUROC per scenario) is weak, because the probes overfit the training set (train AUROC = 1.00 =1.00 everywhere; Appendix ˜ E ). Further, because we select steering layers per-scenario, the layers are confounded with the scenario parameters (dataset, cue). So, comparisons between, e.g., GPQA and MMLU, are not like-for-like; the layer effect and dataset effect are entangled. All labels come from one judge ( gpt-5-nano ), unvalidated against human annotation or a second judge, so judge error is not bounded. We report α = 5 \alpha{=}5 , and over the swept α ∈ { 2.5 , 5 , 7.5 } \alpha\in{2.5,5,7.5} the Gemma-3 12B effect is roughly constant ( ≈ + 0.07 \approx+0.07 ) while Qwen-3.5 9B reaches a marginal effect only at α = 7.5 \alpha{=}7.5 ( Appendix ˜ C ), so we lack a fine-grained or wide-range curve and cannot say where the effect begins or saturates. Cues were varied only on GPQA and datasets only on the Stanford cue, so the cross-cue and cross-dataset results are specific to those settings, and filling the remaining MMLU and BBH combinations would test the transfer more broadly.

Although we observe only mixed generalization, steering for faithfulness merits scrutiny because of what it offers if it does generalize. In contrast to purely observational studies—e.g., monitoring or probing for faithfulness—which offer only diagnoses , steering offers a solution to unfaithfulness. And because steering changes model behavior at inference time, it avoids one potential risk of training-based faithfulness interventions: optimizing either against the chain of thought itself or only against the final output can teach models to obfuscate behaviorally relevant information in their reasoning traces (Korbak et al., 2025 ; Baker et al., 2025 ; Drori et al., 2025 ) . Steering for faithfulness would therefore be a powerful instrument to preserve CoT monitorability; our work seeks to better understand its limitations as a generalizable method.

Acknowledgments

This work was conducted through the Spring 2026 cohort of the Supervised Program for Alignment Research (SPAR). We thank Kairos for helping to organize SPAR and supporting this project.