acceptodds
Under review as a conference paper at ICLR 2027

Discovering Interpretable Acoustic Features in Autoregressive Voice-Cloning TTS Models

Abstract

Voice-cloning TTS models synthesize an unseen speaker's voice from brief speech prompts, but it remains unclear which acoustic properties they represent and use. We introduce a feature-based interpretability framework for autoregressive voice-cloning TTS models, analyzing features from sparse autoencoders (SAEs), PCA, and random directions via sparse probing, steering, and quantitative auto-interpretability. We focus on Qwen3-TTS and extend probing and steering to additional models to test generalization across architectures. These models encode concepts like pitch, energy, noise, and accent, but sparse prediction and steering are effective only for some concepts. Strong probing does not always translate to steering or vice-versa. SAEs are useful for sparse probing of human-labelled concepts and novel-concept discovery, while PCA is competitive for dense probing and steering. Auto-interpretability finds automatically validated SAE features in Qwen3-TTS covering speaker clusters, speaking style, and recording conditions beyond known concepts. Overall, we perform a comprehensive interpretability study of autoregressive voice-cloning TTS models.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
text-to-speech, interpretability, paralinguistics, sparse autoencoders, probing, steering, novel concepts

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_PWcNOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper aims to identify which features of source speech are utilized by zero-shot TTS models for voice cloning, analyzing them through feature extraction methods such as SAE and PCA (Probing + Steering).

The paper proposes a new evaluation pipeline called Auto-Interpretability, which utilizes an Audio LLM (Gemini 3.5 Flash) to evaluate features based on three metrics: Coherence, Label Fit, and Label Stability.

Critical strengths
  • The paper conducts various feature analysis experiments using existing methods such as SAEs, PCA, and Random Directions through sparse probing and steering in zero-shot voice-cloning TTS models. It also proposes an Auto-Interpretability evaluation method using an Audio LLM.
Critical weaknesses
  • The feature extraction analysis relies on previously established methods rather than introducing new feature extraction techniques.
  • Due to API costs, the Auto-Interpretability evaluation is conducted on only one model, Qwen-TTS. Furthermore, the proposed method is essentially a metric-based evaluation relying on the inference results of an existing Audio LLM (Gemini 3.5 Flash).
  • Although the layers used for activation extraction are aligned as closely as possible, they still differ across models. Additionally, a fair comparison would require consistent data splits, but the full feature-fitting split is used for some models, while only a subset is used for others. This raises concerns about the consistency of the experimental setup.
Questions for authors
  • Can the proposed Auto-Interpretability method be reproduced and generalized to other model families beyond Qwen3-TTS?
  • The proposed method uses Gemini 3.5 Flash, a closed-source Audio LLM. Can the evaluation results be reproduced using open-source Audio LLMs, such as Qwen3-Audio, in environments where API access is unavailable?
4 comments

Thank you for your review. We are glad you found our probing and steering analysis across TTS models and our auto-interpretability evaluation worthwhile! Please see our response to each of your comments below.

[W1] About the use of previously established feature extraction methods, we would like to clarify that this is an interpretability analysis paper: we ask what concepts voice-cloning TTS models represent and how they have control over those concepts, rather than aiming to build a new, better feature extractor. Established methods like SAE and PCA are the right choices to answer that question. Most practitioners use them and their behavior on text models is well understood, so our findings reveal facts about TTS models rather than confounding with an untested new method. Thus, we contribute (a) the framework that neatly brings these techniques together for voice-cloning TTS and (b) the findings it yields, which we report in the paper.

[W2, Q1] Yes, the auto-interpretability method can be extended to other model families beyond Qwen3-TTS since the pipeline only needs per-utterance feature activations and selected audio clips, which all models have. We are now running it on CosyVoice2, Fish S1-mini and Higgs TTS 2 using the same setup we used for Qwen3-TTS and will post the results in a follow-up comment once it completes.

[W2, Q2] Yes, we were able to successfully reproduce the auto-interpretability results using an open-source audio LLM judge, Qwen3-Omni! Thank you very much for this suggestion. We reran the Qwen3-TTS layer 19 auto-interpretability evaluation using Qwen3-Omni-30B-A3B-Instruct as labeler/judge (replacing Gemini 3.5 Flash) with the same prompts, feature set, clips and thresholds.

JudgeFeature setCoherenceLabel FitLabel StabilityValidated (of 500)
Qwen3-OmniSAE0.380.570.6346
Random directions0.190.550.626
PCA0.070.500.540
Gemini 3.5 Flash (paper)SAE0.430.630.4192
Random directions0.250.600.3217
PCA0.080.500.221

The ordering between SAE, random directions and PCA remains consistent under both judges, and SAE has the highest Coherence, Label Fit and number of validated features. The controls (shuffled labels and random audio) also stay at chance for both judges. 21 of the 46 SAE features validated by Qwen3-Omni are also passed by Gemini 3.5 Flash, showing agreement. The two judges differ mainly in how many features pass. When we ran the evaluation with an older Gemini model, Gemini 2.5 Flash (not shown in the table above), it validated 55, 4 and 1 SAE, random-direction and PCA features, which roughly matches with Qwen3-Omni. We will add the full results to the paper. Our auto-interpretability results are therefore robust across judges, including an open-source audio LLM judge!

[W3] Regarding the layers question: fundamentally, the models have different depths (24, 28 or 36 backbone layers) so using the same layer index across them would be unwise. Therefore, we match relative depth, using the layer about two-thirds into each backbone. This is the closest alignment possible across models of different depths.

Regarding consistency of data splits, the probe splits and steering examples are already identical across models, but yes, the feature-fitting data differed, with the full split for Qwen3-TTS and CosyVoice2 and a 20% subset for Fish and Higgs (to save on compute costs). This is a valid concern that the reviewer raised. Therefore, we ran experiments that retrained the Qwen3-TTS and CosyVoice2 SAEs and PCAs on the same 20% subset used for Fish and Higgs, and reran the probing and steering experiments. Here are the probing results over the 55 extractor concepts:

ModelFeaturesTop-2 nF1 (SAE / PCA)All-feature nF1 (SAE / PCA)
Qwen3-TTS L19Original0.249 / 0.2240.580 / 0.662
Refit0.240 / 0.2330.576 / 0.662
CosyVoice2 L16Original0.140 / 0.1690.436 / 0.515
Refit0.138 / 0.1720.467 / 0.515

(probing results for human-labelled concepts and steering results are also available, just not shown here)

None of the conclusions changed due to this refit; the best method at every k is the same, no score changes by more than 0.031, and steering conclusions are also unchanged. So the different fitting data does not affect our conclusions, and we will add these refits to the paper.

AuthorsOpenReview ↗

Follow-up on [W2, Q1]: auto-interpretability results on other model families

[W2, Q1] Here, we follow up on our earlier response with the results of applying our auto-interpretability pipeline to other model families. We ran it on all 5 models (Higgs TTS 3, Higgs TTS 2, Fish S2 Pro, Fish S1-mini and CosyVoice2) with the same methodology as for Qwen3-TTS. Here are the results:

ModelFeature setCoherenceLabel FitLabel StabilityValidated (of 500)
Qwen3-TTS (paper)SAE0.430.630.4192
Random directions0.250.600.3217
PCA0.080.500.221
Higgs TTS 3SAE0.450.640.43106
Random directions0.220.570.2912
PCA0.080.490.240
Higgs TTS 2SAE0.360.620.3861
Random directions0.210.570.3410
PCA0.060.510.271
Fish S2 ProSAE0.340.600.3663
Random directions0.220.600.3411
PCA0.060.510.270
Fish S1-miniSAE0.270.580.2821
Random directions0.250.600.3623
PCA0.080.510.261
CosyVoice2SAE0.270.590.3427
Random directions0.240.610.3722
PCA0.080.520.280

Across all six models, PCA scores the lowest, and the shuffled-label and random-audio controls stay at chance (Label Fit 0.49 to 0.51), replicating our Qwen3-TTS findings. However, what differs is how SAEs compare with random directions. SAEs yield far more validated features in four of the six models, Qwen3-TTS, Higgs TTS 3, Higgs TTS 2 and Fish S2 Pro, but are not significantly better than random directions in Fish S1-mini and CosyVoice2.

For CosyVoice2, this is expected given its architecture. Its backbone predicts supervised semantic tokens, which mostly carry what is said and retain only part of how it is said. The final flow-matching decoder, after the backbone, is conditioned again on the reference speech and a speaker embedding, so the backbone is under little optimization pressure to keep detailed acoustic information, such as the speaker's voice. This likely explains why its SAE finds no more interpretable acoustic concepts than random directions. For Fish S1-mini, the cause is less clear, and we leave it to future work.

The kinds of validated features also differ across models. As in Qwen3-TTS, most validated SAE features in Higgs TTS 3 and Higgs TTS 2 are speaker cluster features, which describe voice profiles (86 of 106 and 42 of 61). In Fish S2 Pro, many instead describe laughter (19), specific words such as "probably" and "awesome" (10), and accents (6). Fish S1-mini also has 6 features for specific words or sounds, which are nearly absent in Qwen3-TTS.

Thus, the auto-interpretability method can be applied to other model families, and its main result for Qwen3-TTS, i.e. the SAE's advantage over random directions, holds in Higgs TTS 3, Higgs TTS 2 and Fish S2 Pro i.e. 4/6 models, and does not hold in Fish S1-mini or CosyVoice2. We will add these results to the paper.

AuthorsOpenReview ↗

Follow-up to previous response

Thank you again for your review and feedback! In response to your review, we extended auto-interpretability from 1 to 6 TTS models, added 2 audio LLM judges (including the open-source Qwen3-Omni) and refit 2 models on matched data. The feature-set ranking holds under all 3 judges and our pipeline works across all model families, with SAEs leading in 4 of 6 models.

We hope these clarifications and results address your concerns, and we would be grateful if you would consider raising your score.

Thank you for the authors' cautious answers. I appreciate the detailed experiments and hard work on the requested experiments. Most of my concerns are resolved by the extra experiments. I have raised my score. Good luck.

Reviewer_XiZ1OpenReview ↗
Rating
4
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies the interpretability of residual activations of zero-shot voice cloning text to speech models through the lens of probes such as SAE, PCA, etc. The paper focuses mainly on the Qwen3-TTS model, but also extends it to several other models, which is useful for understanding the generalizability of the probing methods. The paper also studies the steerability of these probed features as well as proposes a framework for LLM based auto-interpretability pipeline. Extensive and broad experiments provide really valuable insight.

Critical strengths
  1. The breadth of the study is highly valuable, from evaluating several models to probing a large set of extractor-driven and human labelled concepts.
  2. Autointerpretability pipeline is interesting and the quantitative validation along with human evaluation grounding is good.
Critical weaknesses
  1. The disconnect between probing and steerability is noted by the authors, however, a deeper analysis of why the disconnect exists would have been very interesting.
  2. While LLM as a judge for the auto-interpretability pipeline is useful, but it still has an inherent risk of hallucination or internal biases towards certain features like age, accents, etc. An analysis/discussion in that direction would have added more value.
Questions for authors

See weaknesses above.

2 comments

Thank you very much for the positive review. We are glad you appreciate the breadth of our study and our auto-interpretability pipeline! Please see our response to each of your comments below.

[W1] We thank you for the suggestion to more deeply analyze the disconnect between probing and steering, which we have now performed. Please see our 'General response on probing vs. steering' for more detail, where we identify three causes of this gap.

[W2] Regarding the risk of LLM judge biases or hallucinations, it helps to distinguish between the different ways the LLM could go wrong. Two of these seem most relevant to our results:

Bad labels. A label can be hallucinated by the LLM, describing a property that is not present in the audio, or stereotyped, where the LLM uses generic accent or age categories that only weakly apply to the audio. Our pipeline handles both using the Label Fit metric: a label passes only if the judge can use it to tell apart new clips where the feature is active from clips where it is not. Hallucinated or stereotyped labels are likely to fail this test: a hallucinated label fits neither set of clips and a stereotyped one fits both, so neither can tell them apart. Consistent with this, our shuffled-label and random-audio controls, where the judge is given another feature's label or random clips, result in Label Fit dropping to chance (0.50). Our human evaluation also suggests little to no accent stereotyping by the LLM: listeners judge labels mentioning an accent to fit the audio as well as other labels (human Label Fit 65% vs. 64%), which would not be the case if the LLM often labelled voices with hallucinated accents.

Bias towards certain concepts, where the judge detects easy-to-hear concepts more reliably than subtle ones, which limits coverage for subtler concepts because features corresponding to them are less likely to pass validation. This is a real limitation on how completely we can recover all the concepts present in the features. However, it does not affect whether the labels that do pass are correct, and the comparison between feature sets (SAE, PCA and random directions) holds with a completely different judge: an open-weight judge, Qwen3-Omni, with its own perceptual strengths and biases, reproduces the same ranking (see our response to Reviewer PWcN). Overall, with current audio LLMs, the labels that pass are reliable, the comparison across feature sets is robust across judges, and the validated labels already go beyond the qualitative labels in prior audio SAE work.

We will add these discussions to the paper.

Reviewer_XiZ1OpenReview ↗

Response to rebuttal

Thanks for the detailed responses. I think including the probing vs steering analysis in the paper would be really good. I also read through the other experiments reported in the rebuttals, especially with the open weight audiollm judge and the corresponding comparison between that and a frontier model is also valuable. I stand by my positive rating, good luck!

Reviewer_pbAvOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper analyzes residual-stream activations in six autoregressive voice-cloning TTS models using sparse autoencoders (SAEs), PCA, random projections, and untrained SAEs. Logistic probes predict 55 extractor-derived and 21 human-labeled concepts; activation-steering experiments use feature directions selected for the extractor-derived concepts. Sparse predictability does not consistently predict steering effectiveness, and simple baselines remain competitive. For Qwen3-TTS, Gemini generates feature descriptions and evaluates their coherence, held-out predictive fit, and stability; 684 SAE features pass these automatic checks, with human evaluation on a subset.

Critical strengths
  1. The six-model study includes strong simple baselines and failed steering cases, making the comparison useful beyond selected success stories.
  2. The distinction between predicting a property and controlling it is supported by concrete examples and weak correlations across concepts.
Critical weaknesses
  1. The claimed SAE advantage needs qualification. Random directions and untrained SAEs perform similarly on top-2 human-concept prediction at Qwen3-TTS layer 19 (Fig. 14). Pooling, candidate counts, and feature scaling during selection differ across methods, so the advantage over PCA does not isolate sparse feature learning.
  2. The 684 features pass checks largely judged by the same model that writes their descriptions. Human label fit is 64.4% for passing features, and 72 of the initial 92 describe speaker clusters. This supports some useful descriptions, but leaves their reliability and distinctness unclear.
  3. The experiments do cover many models and concepts, howevr the analysis remains mostly descriptive. For example, why are some properties easy to predict but difficult to steer, and why does steering work in some architectures but fail in others? What these differences reveal about the representations, and what conclusions should readers draw when choosing an interpretation or steering method?
Questions for authors
  1. Do the main comparisons hold with matched pooling and candidate counts, and with features standardized before selection?
  2. Were human-evaluation clips separate from Gemini’s examples? How much did raters agree, and how many passing descriptions overlap?
  3. Do the extractor-concept results hold on corpora with reliable speaker IDs? Batch-local and utterance IDs do not guarantee separate speakers across splits, I believe.
3 comments

Thank you for your review. We are glad you value our experiments that distinguish prediction of a concept from its control, and our honest reporting of any steering failures. Please see our response to each of your comments below.

[W1, Q1] Yes, the main comparisons hold with standardized features, matched feature set counts and matched pooling. In fact, standardizing makes the SAE advantage clearer. Thank you for this suggestion!

To standardize, we rank features by their mean difference (between positive and negative examples of a given concept) divided by the feature's standard deviation, rather than by the raw mean difference. To match feature set counts, we subsample the SAE, untrained SAE and random-direction features to the number of PCA components. Here are the top-2 nF1 scores on human-labelled concepts, averaged over six models:

Feature selectionSAEUntrained SAERandomPCAModels where SAE is best
Original (paper)0.1080.0590.0510.0674/6
Standardized0.1640.0620.0600.0706/6
Standardized, matched counts0.1340.0510.0680.0706/6

(extractor concepts not shown here)

With standardized selection, the SAE is the best of the four methods in all six models, and it stays the best when feature set counts are also matched. The tie between SAE and untrained SAE you pointed out at Qwen3-TTS layer 19 also disappears: the SAE scores 0.184 against 0.075 for the untrained SAE. On extractor concepts, the paper's conclusion also holds under both changes.

For pooling, we recomputed every method with both mean and max pooling at Qwen3-TTS layer 19, using standardized selection. On human-labelled concepts, the SAE remains the best of the four methods under both poolings, with 1, 2, 5 and 10 features. On extractor concepts, PCA's lead with all features holds when each method uses its better pooling.

So any SAE advantages are not an artifact of feature scaling, feature set counts or pooling. We will add these findings to the paper.

[W2, Q2] You raise concerns about the reliability of our validated features, since Gemini both writes and checks their labels, and about their distinctness.

Regarding reliability, our results do not depend only on Gemini checking its own labels, since human raters independently reported a Label Fit of 64.4% for validated vs 51.0% for nonvalidated features (chance is 50%).

Answering your first two questions in Q2: yes, the human-evaluation clips were separate from the ones used by Gemini to write the labels. Interrater agreement was modest. Two raters gave the same answer on 66.5% of clips for validated features (Krippendorff's alpha 0.25) against 56.9% for nonvalidated features (alpha 0.06).

Distinctness is a valid concern, because 334 of the 684 descriptions have another description that says essentially the same thing (cosine similarity >= 0.9 with a sentence-embedding model). These are mostly descriptions of voice profiles (timbre, accent and delivery), which are hard to differentiate in words. The features themselves are not duplicates, because only 14 of the 601 pairs of similar descriptions are active on mostly the same utterances. We will clarify this carefully in the paper.

[W3] Thank you for these questions. Please see our 'General response on probing vs. steering', which identifies three causes of the gap between probing and steering. We find that a good concept detector direction may sometimes be a bad concept steering direction, and we provide analyses for this.

For choosing a method, we will add concrete guidance to the paper:

  1. To check whether a concept is encoded, PCA is enough. To find a few features that predict a concept, use SAEs.
  2. To steer a concept with a single vector, use DiffMean if you have examples with and without the concept. If DiffMean fails, SAE and PCA features almost always fail too. Without labelled examples, e.g. for a feature found by auto-interpretability, use it to steer and check the generated speech directly, rather than relying on its description.

[Q3] Yes, the results hold. Emilia and MSP-Podcast lack reliable speaker IDs, so we removed them, split the remaining six datasets by speaker, and reran probing: the main extractor-concept conclusions continue to hold in all six models even on corpora with reliable speaker IDs.

Thank you for the detailed rebuttal and additional experiments. I appreciate the authors' efforts to address my concerns, particularly regarding the controlled comparisons (W1) and speaker-disjoint evaluations. I do believe that these clarifications strengthen the paper; my general concerns about interpretability, the reliability of feature descriptions, and the usefulness of insights remain. Therefore, I will maintain my original scores.

AuthorsOpenReview ↗

Follow-up to previous response

Thank you for taking the time to read our rebuttal, and for appreciating that our controlled comparisons (W1) and speaker-disjoint evaluations strengthened the paper. Below, we briefly summarize how our response addresses the three general concerns you mention: the reliability of feature descriptions, interpretability, and the usefulness of insights.

Reliability of feature descriptions. Our validated descriptions do not rely on Gemini checking itself: human raters, using held-out clips, found that validated labels fit better than nonvalidated ones (64.4% vs. 51.0%, chance 50%), and only 14 of the 601 pairs of similar descriptions are duplicate features. An independent open-weight judge, Qwen3-Omni, reproduces the same feature-set ranking.

Interpretability. We read this concern as two points from your review. First, in our general response, we explained in detail why probing and steering diverge, and found three causes: some concepts, such as timing, cannot be steered with a single vector; probing-selected SAE features can be misaligned with the concept's steering direction; and some models, such as CosyVoice2, resist steering altogether. Second, the many speaker cluster features (72 of the initial 92) are expected, since voice cloning needs substantial timbre capacity and the judge hears timbre easily. Importantly, we also validate many non-speaker concepts, such as laughter and accents: 147 in Qwen3-TTS, and most validated features in Fish S2 Pro, which even include specific words like "probably" and "awesome".

Usefulness of insights. We provide actionable guidance based on our results. To check whether a concept is encoded, PCA is enough; to find a few features that predict it, use SAEs. To steer with a single vector, try DiffMean first: where it fails, SAE and PCA features also fail. Features discovered without labels should be used to steer and checked on the generated speech, not trusted from their descriptions.

To recap, in response to your review, we ran 4 probing robustness checks and 3 new analyses (rater agreement, description overlap, and why probing and steering diverge) and added guidance for choosing a method. Across all reviews, we also extended auto-interpretability to 6 models and 3 judges, refit 2 models, and posted 6 new results tables.

If any part of these concerns remains unclear, we would be glad to clarify it. We hope these points address your concerns, and we would be grateful if you would consider increasing your score. Thank you again for your time and feedback.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General response on probing vs. steering

We thank all reviewers for their helpful comments. We are glad to see that they found our study broad, spanning several TTS models and concepts (Reviewers PWcN and XiZ1), with strong simple baselines and a clear distinction between prediction and control (Reviewer pbAv); and that Reviewer XiZ1 found our auto-interpretability evaluation interesting.

Reviewers pbAv and XiZ1 were both interested in understanding why strong probing does not always translate into effective steering, and we discuss the gap between probing and steering in detail here, answering pbAv’s W3 and XiZ1’s W1.

Here, as in the paper, steering means adding a single steering vector to the residual stream during generation. A concept that can be probed well can nevertheless fail to get steered because either (a) the selected features are bad steering vectors, or (b) the concept cannot be steered at all using a single steering vector. To tell these two apart, we compare each steering vector against each concept’s supervised difference-of-means (DiffMean) direction: if steering fails even with DiffMean, the problem lies with the concept or the model, while if it succeeds, the problem lies with the feature method. Based on this analysis, we find three individual causes for the probing-steering disconnect:

1. Some properties can be predicted but not steered with a single vector. At Qwen3-TTS layer 19, consider these concept groups:

Concept groupAll-feature nF1 (SAE / PCA)Target Change (SAE / PCA / DiffMean)
Energy0.59 / 0.670.49 / 0.31 / 0.38
Pitch0.52 / 0.620.25 / 0.25 / 0.17
Timing (duration, speaking rate, pauses)0.62 / 0.640.01 / 0.06 / 0.06

Timing is predicted as well as energy and pitch, yet no single steering vector steers it, not even DiffMean. This result holds for every model. Adding a fixed direction at each inference step can steer frame-level acoustics such as loudness and pitch, but it seems like it cannot affect utterance-level decisions such as when to pause. Timing is of course controllable by other means; it is just not achievable with this kind of single-vector intervention.

2. Probing selects SAE features that detect a concept, but they are not always aligned with the concept's DiffMean direction. The steering vector is built from the selected features' decoder directions, which for many concepts point only partly along DiffMean:

Qwen3-TTS L19Fish S1-miniFish S2 ProHiggs TTS 2
SAE alignment = cos(SAE vector, DiffMean)0.490.410.430.29
Spearman(SAE alignment, SAE Target Change)0.560.470.510.33
Spearman(SAE top-2 nF1, SAE Target Change)0.20−0.070.000.12

(Spearman correlations are over 55 extractor concepts)

SAE alignment, the median cosine similarity of the SAE vector with the DiffMean vector, can predict which concepts are steerable by SAEs, while probing accuracy barely does i.e. a good concept detector direction is sometimes not useful as a concept steering direction. This also better explains failures like the SAE pitch steering failure in Higgs TTS 2; its selected pitch features turned out to be nearly orthogonal to the pitch DiffMean direction, despite being top-ranked by probing.

3. Some models resist single-vector steering altogether. In CosyVoice2, no steering vector changes any concept, not even DiffMean. Its backbone only predicts semantic tokens, and a separate flow-matching decoder generates the final speech re-conditioned on the reference speech, so it likely re-imposes the reference's pitch, energy and timbre regardless of the steering.

Together, these show that successful single-vector steering needs three things: (a) concepts amenable to change via a single, fixed added residual vector, (b) a good, correctly aligned steering vector, and (c) a model that can retain the applied change through to the audio. We will add these analyses to the paper.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.