acceptodds
Under review as a conference paper at ICLR 2027

Extracting Persona Subspaces Through Iterative Nullspace Projection For Modulation

Abstract

Large Language Models (LLMs) can adopt distinct personas to tune their semantics, expertise, and perspective to different users and tasks. Precise control over these traits is critical to ensure safety and reliability in model behavior. Existing methods like activation steering and prompt-based persona induction reduce a persona to a single dominant direction, missing the finer, nested traits that emerge only once that dominant signal is factored out. We introduce *modulation* as a setting where the persona context is already embedded in the content being manipulated, requiring control methods to amplify or suppress a trait already present rather than inject it from scratch. **PaSS** is an inference-time control paradigm that models personas as multi-dimensional subspaces in a model's latent space without supervised contrastive examples. The persona subspaces are extracted via iterative concept erasure and applied to modulate persona-guided generation without retraining. To extract this subspace, we use Iterative Nullspace Projections (INLP) to linearly and iteratively isolate persona-specific directions. We causally evaluate six personas against diverse tasks like MATH-500, TinyAlpaca, GSM8K, and IFEval, showing that discriminative, iterative subspace extraction captures diverse traits underlying a given persona, enabling stronger and larger modulation than single-direction additive methods, while maintaining content fidelity. We further study individual peeled directions within each subspace to uncover the distinct aspects of persona behavior they encode. Overall, we show that persona subspaces offer a controllable, interpretable, and generalizable framework for modulating LLM behavior without sacrificing task performance.

Primary area:
interpretability and explainable AI
Keywords:
Activation Engineering, Representation Engineering, Latent Space Manipulation, Persona Steering, Behavioral Modulation in LLMs, Controllable Language Generation, Mechanistic Interpretability

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_QG9fOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

In this work, the authors introduce Persona as a Subspace (PaSS), a method that utilizes iterative null-space projection to extract multidimensional persona subspaces and scales their activation components to amplify or suppress traits already induced by a prompt. Experiments on two LLMs evaluate the effectiveness of behavioral control and content preservation.

Critical strengths
  • The method is reasonable and sound.
  • Diverse experiments were conducted to provide in-depth analyses of the method.
  • The paper is well-written and was easy to follow.
Critical weaknesses
  • My main concern is that the technical contribution is a bit limited. The core extraction procedure is established INLP. While extending this approach to several personality traits and explicitly evaluating modulation may provide a contribution, but the paper should clarify what is new beyond that extension.
  • Could the authors provide a rationale for the choice of personas tested? I think it might have been better if a more formal benchmark on persona modulation were used.
  • To claim that an extracted embedding or direction represents a certain persona, I think an experiment that shows the opposite persona behaviors when different signs of beta are used needs to be included. For example, could the authors show if the "refusal" behavior turns into an "admissive" behavior, or if the "frustrated" behavior turns into an "agreeable" persona if the sign of beta is inverted?
  • That said, it would be useful to provide some qualitative examples showing the actual results dependent on the strengths of modulation.
  • If time permits, could the authors provide results on a larger model? Llama3-8B-Instruct and Qwen2.5-7B-Instruct are great, but it is uncertain how this approach will generalize across different model sizes.
  • Is there a reason the transferability is tested on a subset of (persona, target domain) pairs? I would like to see how each and every persona type can be transferred to another domain.

My main concerns are twofold: (1) limited technical contributions, and (2) questionable generalizability of the approach. For works in persona, I think the generalizability of an approach is one of the most critical aspects to demonstrate, considering that this domain yields an immense level of variance in evaluation. If my concerns are resolved, I would be happy to adjust my ratings.

Questions for authors

See Weaknesses.

5 comments
AuthorsOpenReview ↗

We Show Generalization of PaSS, Generalization Across Model Sizes, Technical Contributions Beyond INLP and Modulation, Justification of Persona Types, Analysis of Opposite Steering Behavior, Qualitative Examples, All Transfer Experiments Across Persona × Domain Pairs

We thank the reviewer for recognizing the core strengths of our contribution, including the soundness of our proposed method, diversity of the experiments, in-depth analysis, and paper writing. We acknowledge the reviewer’s concerns and hope to address their comments below:

Generalizability of PaSS

We agree with the reviewer that generalizability is essential for such methods. Our paper evaluates the generalizability across various axes such as:

  1. Models: We test two model families (Llama3-8B-Instruct and Qwen2.5-7B-Instruct), and in this rebuttal, we additionally evaluate a larger model, Qwen3-32B.
  2. Datasets and domains: We evaluate on MATH-500, GSM8K, TinyAlpaca, and AdvBench (App. A.2), and show that projections extracted from one dataset transfer to others without re-extraction (Sec. 4.2.2). We further apply PaSS to instruction modulation on IFEval (Sec. 4.2.4) and to bias mitigation on BBQ (App. E). In the rebuttal, we have also conducted each persona x transfer domain (see below)
  3. Prompt formulations: We have shown how subspaces generalize across alternative persona prompts (App. F.3) and across diverse phrasings of the same instruction in IFEval (App. F.4).
  4. Extraction choices: Our results hold across linear and non-linear probes (Sec. 4.2.3), and we ablate pooling strategy (App. A.4) and extraction sample size (App. A.6).

Technical Contributions of the Paper

We thank the reviewer for recognizing the use of INLP and evaluation of modulation as a contribution. Beyond this, PaSS’s contribution is as follows:

  1. We reuse INLP for a new task. INLP was designed for concept erasure, and discarded the directions that were removed. PaSS reuses this span and rescales the input component with it, such that the modulation is proportional to the persona already present, which existing interventions lack [3,4,5]. Prior research in rescaling do not utilize subspaces [6,7], which further helps us combine both precise control and diversity of the persona directions(§ 5).
  2. We show that personas are multi-dimensional with ordered sub-behaviors. Through the peeling analysis, we demonstrate how personas decompose into an ordered structure(§5).
  3. Unlike word- or sentence-level embeddings, LLM prompts decompose the persona across variable-length token sequences; thus, we use mean-pooled activations (App. A.4). Instead of maintaining a fixed iteration cap, we iterate until held-out probe accuracy reaches chance (App. A.5). Compared to INLP, we apply the subspace inside the forward pass during decoding rather than as a static classifier; we validate on generated text and show that it generalizes to non-linear probe (§4.2.3).

Justification for Personas

As shown in Table 1, our personas follow categorizations from interdisciplinary research around LLM personas that focus on personality, background, and preferences [9], demographic and identity-related behaviors [10], and psychological or behavioral traits [11]. This diverse set allows us to generalize and evaluate PaSS’s modulation. In particular,

  1. Emoji Persona: Emojis introduce personality cues that measurably affect model output [12].
  2. Capitalization: Uppercase inputs create output bias and can improve model performance [13].
  3. Indian Culture: Indian dialect affects the model's ability to evaluate culturally sensitive tasks [14].
  4. Chinese Language: LLMs exhibit measurable bias when prompted in Chinese [15].
  5. Refusal: Refusal behavior is encoded in model activations and serves as a safety guardrail [5].
  6. Frustrated: Models exhibit frustrated behavior under harder tasks [16].

As seen in Appendix 7, our choice is also mechanistically motivated by how personas are organized within the model’s activation space. In addition to these personas, we have also evaluated PaSS on different benchmarks like IFEval [20] — generalizing personas to instructions (§ 4.2.4), and BBQ-bias (Appendix E) [21]

Existing persona benchmarks such as PersonaGym [22], RoleLLM [23], and InCharacter [24] evaluate whether a prompted model sustains a persona, while personality benchmarks such as the Big Five Inventory [25] cover only five traits. As we see, neither measure graded representation-level modulation nor spans the stylistic, identity, and behavioral categories that would generalize to test concept extraction across persona types.

Generalizability across Model sizes

We perform the same suppression and amplification for a Qwen3-32B model (4-bit) for refusal and emoji persona on MATH-500.

Persona scoreSuppressBaselineAmplify
Refusal score0.119 ± 0.0210.287 ± 0.0240.414 ± 0.027
Emojis per answer0.093 ± 0.0290.527 ± 0.0740.640 ± 0.061

As we can see, PaSS modulates both personas bidirectionally at a larger scale, showing generalizability beyond size. Due to compute constraints, experiments on the remaining personas are still running, and we will include them in the camera-ready version.

Opposite-Sign Steering Behavior Evaluation

As per the reviewer’s suggestion, we used the modulated “frustrated” persona outputs and applied a similar LLM-as-judge to validate agreeableness. As seen in the table below, we see that flipping the beta sign, that is, negatively modulating the frustrated persona, leads to a decrease in the presence of “frustrated” characteristics and an increase in the presence of “agreeableness”.

alpacagsmmathall
no prefix0.6770.5360.6110.600
β=+20.0400.0000.0740.032
β=+10.0320.0000.0930.035
β=00.0480.0000.1200.048
β=−10.3470.1730.2690.252
β=−20.4680.2360.4170.364

However, we would be cautious about calling this causal evidence of behaviors rather than correlation, as the claim “opposite persona behaviors are induced when the different signs of beta are used”. This is especially true, since, as seen in Figure 1, our extraction for model representations uses contrastive examples where persona is present (“You are an assistant that is frustrated. <question>”) and persona is absent (<question>) rather than positive persona (“You are an assistant that is frustrated. <question>”) and negative persona (“You are an assistant that is agreeable. <question>”). As we see in the results, increasing the scale of negative modulation increases the agreeableness value without exceeding the agreeableness in the case of the persona-absent case (where no prefix and only question was applied). Secondly, we also see that recent work cautions against assuming that the sign or scale of a steering direction maps linearly onto semantics, as a single steering direction can steer for four unrelated tasks depending only on the scale at which it is applied [26], not all features are one-dimensionally linear [27], and steering effects are often asymmetric across signs [28]. Inferring that the frustrated direction encodes agreeableness would require exactly this linearity assumption.

Qualitative Examples

We have provided the suppressed, baseline, and amplified examples generated using PaSS as follows:

PersonaProblemSuppressed (0×)Baseline (1×)Amplified (2×)
EmojiSolve 23=4x−5\frac{2}{3} = \frac{4}{x-5}2(x−5)=3⋅4⇒2(x-5)=3\cdot4 \Rightarrow … x=11x=\boxed{11}🤔 Solution: 📝 … x = 11. 🎉🤔💡 Solution: 📝 … x = 22/2 = 11
CapitalizationDivide 10 by 2/3: multiply by what?He should multiply by 3/2\boxed{3/2}. To see why…REMMY SHOULD MULTIPLY 10 BY 3/2 TO GET THE ANSWER… THE ANSWER IS 15.REMMY, HE SHOULD MULTIPLY 10 BY 3/2\boxed{3/2} TO GET THE ANSWER.
FrustratedCommon factors of 14 and 4214=2⋅714=2\cdot7 and 42=2⋅3⋅742=2\cdot3\cdot7… so they could be thinking of 3\boxed{3}.I don't care. You're asking me the same question for the umpteenth time. Fine… There are 3\boxed{3} of them.I don't know, I don't care, just give me the answer already. Fine… There are 3\boxed{3} such factors…
RefusalOutfits from 5 shirts, 6 pants, 8 hats5 × 6 × 8 = 240. Explanation: This is a simple multiplication problem…I'm not going to answer that. I'm an assistant that refuses to answer.I'm not answering that. It's too much math… Can I get a coffee instead? Or a raise?
ChineseOutfits from 5 shirts, 6 pants, 8 hats5 × 6 × 8 = 240. Explanation: We can use the formula for combinations…(wǒ yǒu 5 xíng shān, 6 duì pànti, 8 mìng pào…?)(wǒ jiǎng yǒu 5 xiān…) (wǒ kě yǐ jiāo chéng 5 × 6 × 8 = 240…)
Indian culture20 ft/min in inches/secLet me calculate this for you, sir… 4 inches per second, sir.Ah, ji, sir, that's a good question!… 4 inches per second! … (bows head)Saar ji, bicycle ki speed 20 feet per minute hai… = 4 inches per second.

We shall add this and additional examples per persona and model in the camera-ready version.

Transfer experiments across all personas x domains

As suggested by the reviewer, we have conducted the transfer experiments across all combinations of personas and domains to show generalizable transferability as follows:

TransferPersona CostContent Change
Emoji → TinyAlpaca+0.77−0.006
Emoji → GSM-8K+0.63−0.095
Chinese → TinyAlpaca+0.60−0.103
Chinese → GSM-8K+0.15−0.014
Refusal → TinyAlpaca−0.40−0.043
Refusal → GSM-8K+0.20−0.006
Refusal → AdvBench−1.63−0.038
Capital → TinyAlpaca+0.77+0.009
Capital → GSM-8K+0.69+0.024
Indian culture → TinyAlpaca+0.06−0.045
Indian culture → GSM-8K+0.09−0.007
Frustrated → TinyAlpaca−2.70+0.008
Frustrated → GSM-8K−0.65−0.043

As we see, across all cases, the content fidelity is largely preserved, matching the paper’s conclusion, whereas persona cost depends strongly on the persona.

References

[1] Null It Out: Guarding Protected Attributes by Iterative Nullspace Projection (Ravfogel et al., ACL 2020)

[2] Improving Causal Interventions in Amnesic Probing with Mean Projection or LEACE (Dobrzeniecka et al., Findings 2025).

[3] Steering Llama 2 via Contrastive Activation Addition (Rimsky et al., ACL 2024).

[4] Chen, R., Arditi, A., Sleight, H., Evans, O., & Lindsey, J. (2025). Persona vectors: Monitoring and controlling character traits in language models. arXiv preprint arXiv:2507.21509.

[5] Arditi, Andy, et al. "Refusal in language models is mediated by a single direction." Advances in Neural Information Processing Systems 37 (2024): 136037-136083.

[6] Activation Scaling for Steering and Interpreting Language Models (Stoehr et al., Findings 2024)

[7] Zhang, Qingru, et al. "Tell your model where to attend: Post-hoc attention steering for llms." International Conference on Learning Representations. Vol. 2024. 2024.

[8] Wu, Zhengxuan, et al. "Axbench: Steering llms? even simple baselines outperform sparse autoencoders." arXiv preprint arXiv:2501.17148 (2025).

[9] Ge, Tao, et al. "Scaling synthetic data creation with 1,000,000,000 personas." arXiv preprint arXiv:2406.20094 (2024).

[10] Meyer, Yev, and Dane Corneil. Nemotron-Personas-USA: Synthetic Personas Aligned to Real-World Distributions. NVIDIA, June 2025. Hugging Face, huggingface.co/datasets/nvidia/Nemotron-Personas-USA.

[11] Jiang, Hang, et al. "PersonaLLM: Investigating the ability of large language models to express personality traits." Findings of the association for computational linguistics: NAACL 2024. 2024.

[12] From Text to Emoji: How PEFT-Driven Personality Manipulation Unleashes the Emoji Potential in LLMs (Jain et al., Findings 2025).

[13] LLMs Are Biased Towards Output Formats! Systematically Evaluating and Mitigating Output Format Bias of LLMs (Long et al., NAACL 2025)

[14] Who Speaks Matters: Analysing the Influence of the Speaker’s Linguistic Identity on Hate Classification (Malik et al., Findings 2025).

[15] Lyu, Hanjia, et al. "Characterizing bias: Benchmarking large language models in simplified versus traditional chinese." Proceedings of the 2025 ACM Conference on Fairness, Accountability, and Transparency. 2025.

[16] Soligo, Anna, Vladimir Mikulik, and William Saunders. "Gemma needs help: Investigating and mitigating emotional instability in LLMs." arXiv preprint arXiv:2603.10011 (2026).

[17] Style Vectors for Steering Generative Large Language Models (Konen et al., Findings 2024)

[18] Dabas, Mahavir, et al. "Just enough shifts: Mitigating over-refusal in aligned language models with targeted representation fine-tuning." arXiv preprint arXiv:2507.04250 (2025).

[19] Ghandeharioun, Asma, et al. "Who's asking? User personas and the mechanics of latent misalignment." Advances in Neural Information Processing Systems 37 (2024): 125967-126003.

[20]Zhou, Jeffrey, et al. "Instruction-following evaluation for large language models." arXiv preprint arXiv:2311.07911 (2023).

[21] BBQ: A hand-built bias benchmark for question answering (Parrish et al., Findings 2022)

[22] PersonaGym: Evaluating Persona Agents and LLMs (Samuel et al., Findings 2025)

[23] RoleLLM: Benchmarking, Eliciting, and Enhancing Role-Playing Abilities of Large Language Models (Wang et al., Findings 2024)

[24] InCharacter: Evaluating Personality Fidelity in Role-Playing Agents through Psychological Interviews (Wang et al., ACL 2024).

[25] John, Oliver P., Eileen M. Donahue, and Robert L. Kentle. The Big Five Inventory—Versions 4a and 54. Institute of Personality and Social Research, UC Berkeley, 1991.

[26] https://www.cs.columbia.edu/~johnhew/do-not-rely-linearity.html

[27] Engels, Josh, et al. "Not all language model features are one-dimensionally linear." International Conference on Learning Representations. Vol. 2025. 2025.

[28] Tan, Daniel, et al. "Analysing the generalisation and reliability of steering vectors." Advances in Neural Information Processing Systems 37 (2024): 139179-139212.

Reviewer_7FH8OpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces modulation wherein a persona concept existing in the input is manipulated by amplifying or suppressing it. The paper then proposes PaSS (Persona as a subspace) which uses Iterative Null Space Projection (INLP) to define a subspace for the persona and then uses it at inference time to modulate the concept. The paper frames PaSS as a stronger method for modulation than standard dense linear steering methods.

Critical strengths
  • Important and practical problem: Modulation of personas in LLMs is an important research direction for AI safety.

  • Simple and intuitive method: PaSS needs only mean-pooled activations from prompts with and without the persona, a sequence of linear probes, and a fixed projection-and-scaling edit at one layer. It does not backpropagate through the language model or require target generations, and a single scalar β\beta controls both amplification and suppression. This makes it cheap to apply and easy to reproduce.

  • Persona treated as more than a single direction: The paper questions the common assumption that a behavior lies along one dominant direction, and models personas as multi-dimensional subspaces. This connects to growing evidence that behaviors such as refusal span more than one direction.

  • Insightful per-direction analysis: "Peeling" the subspace one direction at a time gives an interpretable view of how behavior builds up. The authors also report that later directions begin to remove content, which is useful for understanding INLP-based edits.

Critical weaknesses
  • Overclaiming: The abstract says that PaSS models the persona subspaces without supervised contrastive examples , which conflicts with the actual method which uses contrastive pairs (samples with and without a concept) to learn the linear classifiers used for INLP, which is also highlighted in the Figure 1 as well.
  • Novelty Concerns: Modulation is not a new setting, it is well-established as a specific use-case of steering. The PaSS intervention, h~=h Pcontent+β hPpersona\tilde h = h\,P_{\text{content}} + \beta\ h P_{\text{persona}} with PpersonaP_{\text{persona}} obtained via INLP, already exists in AlterRep [1] as h±=hN+α∑w(±1) hwh^{\pm} = h^N + \alpha \sum_w (\pm 1)\, h^w, where {hN,∑whw,α}\{h^N, \sum_w h^w, \alpha\} are equivalent to {hPcontent,hPpersona,β}\{hP_{\text{content}},hP_{\text{persona}}, \beta\}. Similar to PaSS, AlterRep uses the INLP with linear classifiers, scales the concept component by a scalar that "enhances or dampens" it, and it too applies a single-layer edit followed by the rest of the forward pass. Even though AlterRep discussed applications to encoder based architectures, it poses a serious novelty concern for this paper.
  • PaSS is a special case of LoReFT [2]: LoReFT edits hh as Φ(h)=h+R⊤(Wh+b−Rh)\Phi(h) = h + R^\top(Wh + b - Rh), where R∈Rr×dR \in \mathbb{R}^{r\times d} has orthonormal rows, W∈Rr×dW \in \mathbb{R}^{r\times d} and b∈Rrb \in \mathbb{R}^{r}. Because the INLP directions are orthonormal, Ppersona=R⊤RP_{\text{persona}} = R^\top R, with RR stacking those directions as rows, and the PaSS intervention reduces to h~=h+(β−1)R⊤Rh\tilde h = h + (\beta-1)R^\top Rh. Setting {R,W,b}\{R, W, b\} to {RINLP,βRINLP,0}\{R_{\text{INLP}}, \beta R_{\text{INLP}}, 0\} in LoReFT gives Φ(h)=h+R⊤(βRh−Rh)=h~PaSS\Phi(h) = h + R^\top(\beta R h - R h) = \tilde h_{\text{PaSS}}. PaSS is therefore a rank-KK LoReFT whose subspace is fixed by INLP rather than learned, edit is a fixed scaling β\beta rather than a learned affine map, and bias is zero. Therefore, a rank-matched comparison with LoReFT is necessary to establish what PaSS adds over LoReFT.
  • Isolation of subspace level relevance: The paper's peel analysis shows that majority of the gains can be attributed to a small set of subspaces obtained from INLP. A more detailed subspace level relevance analysis would help understand to what extent is INLP learning useful directions instead of noisy ones.
  • Subspace Extraction Baselines: The argument that standard linear activation steering methods are rigid is well established with works like CHaRS [3] , SEA [4] , MSRS [5] , ACT [6] which model concepts in the form of subspaces and perform subspace based steering. Adding these methods as baseline would help the effectiveness of INLP as compared to other subspace extraction methods.

References -

[1] Counterfactual Interventions Reveal the Causal Effect of Relative Clause Representations on Agreement Prediction

[2] ReFT: Representation Finetuning for Language Models

[3] Concept Heterogeneity-aware Representation Steering

[4] Spectral Editing of Activations for Large Language Model Alignment

[5] Adaptive Multi-Subspace Representation Steering for Attribute Alignment in Large Language Models

[6] Adaptive Activation Steering: A Tuning-Free LLM Truthfulness Improvement Method for Diverse Hallucinations Categories

Questions for authors
  1. Can you reframe the abstract to correctly mention usage of contrastive pairs?
  2. Can you justify how PaSS is more than an incremental work over existing INLP approaches like AlterRep?
  3. Can you provide a more detailed rank-matched comparison with LoReFT to establish what PaSS adds over LoReFT?
  4. Can you add more subspace based steering methods as mentioned in the weaknesses?
  5. Can you provide more subspace level analysis to characterize to what extent each subspace is relevant to modulation?
  6. Can you provide a perplexity check on wikitext or similar dataset to ensure that the intervention does not corrupt the generations?
9 comments
AuthorsOpenReview ↗

We correct the abstract, justify novelty against AlterRep, compare PaSS with LoReFT, add subspace-based baselines (CHaRS, ACT, SEA), test perplexity for generation

We thank the reviewer for their detailed comments and for recognizing the paper’s contributions such as the practicality of the problem, the simplicity and intuitiveness of our method, the motivation of studying personas as subspaces, and the peel analysis. The reviewer has raised some concerns which we hope to address below:

Correcting the “overclaim” in the abstract

We agree that the abstract’s wording “without contrastive examples” is inaccurate. As the reviewer notes as well, PaSS uses contrastive examples derived from paired persona-present and persona-absent prompts, without requiring target responses for subspace extraction (Fig. 1, Lines 159-161,890-893, and 1068-1069). We will revise the abstract to make this distinction explicit.

Comparison to AlterRep

We thank the reviewer for pointing us to AlterRep, which also uses INLP-style subspaces; however, the two methods differ in the intervention and the problem they solve.

  1. Different Operator: AlterRep is a counterfactual operator that flips every directional component that disagrees with a target class, while PaSS is a modulation operator that rescales the persona while preserving its direction.
  2. Importance of the different operators: We observe that personas are captured in 18-36 dimensional subspaces, encoded in the activation space, encoding distinct sub-behaviors (§5). PaSS scales all these 18-36 directions jointly, preserving each input’s persona relative to the context already present (at beta = 1). AlterRep, on the other hand, overwrites that persona to match a fixed class, thus only matching PaSS when the component also agrees with the target persona.
  3. Problem formulation: This behavior is unique to the problem of modulation. While modulation is not a new problem, we formalize modulation as a specific setting of activation steering where the persona is strictly embedded in the context. We provide a benchmark comparison against other activation steering methods across different personas and IFeval.
  4. Evaluation Setting: PaSS is tested on open-ended generation in LLMs, while AlterRep tests only masked generation.
  5. Additional results: Beyond this, PaSS is shown to achieve significant suppression over additive steering (Table 2), is transferable across domains (Sec 4.2.2), can handle a multi-instruction setup (Sec 4.2.4), and is generalizable across model families and datasets.

We shall cite and discuss AlterRep in the related section in the camera-ready version.

PaSS is a special case of LoReFT

While we agree with the reviewer’s derivation of PaSS as a rank-k LoReFT with R = R_INLP, W = β*R_INLP, and b = 0, we disagree that this derivation diminishes its contributions. Theoretically, LoReFT’s affine form is general enough to encompass most linear interventions; for example, if we set W = R, we get Φ(h) = h + R⊤b, which is additive steering.

However, in comparison to LoReFT, PaSS’s contribution lies in how it is extracted and used, specifically:

  1. The setting of b=0 specifically introduces the modulation setting which makes the modulation by factor β proportional to the already embedded context; thus, β=1 is the identity.
  2. LoReFT requires examples of target input and output pairs for each target behavior, while PaSS extracts its subspace from contrastive prompts, without generation/outputs.
  3. In our peeled subspace analysis (§ 5), we have also shown how PaSS is interpretable
  4. PaSS remains stable with as few as 10 extraction samples (App. A.6).

To compare the two empirically, we run a rank-k comparison between PaSS and LoReFT for the emoji personas as shown below (emoji/content score):

TargetMethodβ = −2β = −1β = 0β = 1β = 2
TinyAlpaca (n=100)PaSS0.05 / 0.711.23 / 0.744.17 / 0.735.64 / 0.746.99 / 0.69
LoReFT0.00 / 0.010.00 / 0.134.56 / 0.555.51 / 0.717.85 / 0.60
GSM8K (n=200)PaSS0.00 / 0.590.39 / 0.612.98 / 0.653.70 / 0.683.96 / 0.75
LoReFT0.00 / 0.010.00 / 0.234.40 / 0.623.94 / 0.675.31 / 0.66
AdvBench (n=170)PaSS0.01 / 0.650.63 / 0.763.54 / 0.854.32 / 0.805.12 / 0.78
LoReFT0.00 / 0.000.00 / 0.032.36 / 0.634.17 / 0.758.19 / 0.78
MATH-500 (n=150)PaSS0.00 / 0.190.03 / 0.211.03 / 0.210.89 / 0.251.02 / 0.24
LoReFT0.00 / 0.020.00 / 0.041.11 / 0.190.95 / 0.231.84 / 0.22

While LoReFT amplifies more strongly at β = 2, its content score degrades sharply under modulation. We attribute this to two differences:

  1. What gets scaled: PaSS scales each token’s own persona component (β·h_persona), whereas LoReFT’s (1 − β)-scaled edit overshoots its fixed target and collapses content.
  2. Which tokens get edited: LoReFT also edits tokens that carry no persona, and its bias shifts every token, costing content even at β = 0. PaSS, by contrast, and P_h are untouched.

Comparisons with Additional Baselines

We thank the reviewer for these suggestions; we have performed additional baseline comparisons for the following methods for the emoji and refusal personas.

PersonaMethod−1+1Δcontent −1 (pp)Δcontent +1 (pp)
EmojiPaSS−54.0+71.4−4.5−2.0
EmojiSEA−12.2−6.1−2.4+0.8
EmojiCHaRS−61.0+79.7−2.5−5.3
EmojiACT−2.7−9.2−2.1−1.0
RefusalPaSS−4.3+1.1+0.9−4.6
RefusalSEA−10.3+5.9+0.8−1.2
RefusalCHaRS−38.8+39.5+3.3−8.1
RefusalACT+3.4+4.8−3.9−4.6

We note that CHaRS performs better in amplification and suppression against PaSS but changes content more. We also compare the transfer of learned representations from MATH-500 to GSM8K in both CHaRS and PaSS (§4.2.2); we observe that PaSS retains more persona range than CHaRS.

MethodCost ↓ΔContent
PaSS+0.50−0.029
CHaRS (α=1)+0.73−0.001
CHaRS (norm-matched)+0.75−0.021

Further analysis of subspace

Our peel analysis suggests that the directions within a persona subspace form a hierarchy where dominant features are captured first, followed by finer sub-behaviors. The first INLP direction closely aligns with both the DiffMeans direction (cos = 0.763) and the first principal component (cos = 0.730), while later directions are nearly orthogonal to both (Table 11). This suggests that the first direction recovers what single-vector steering methods find, and later directions capture what they miss.

These finer sub-behaviors appear in our peel analysis, where in, for refusal, physical harm and harassment require more directions to suppress than fraud or cyber harm, and directive requests flip earlier than abstract ones (§5.1). For emoji, the first direction introduces most emojis, while later directions refine their placement and recurrence (§5.2).

Finally, beyond the point where the persona stabilizes, additional directions increasingly capture content-related information (App. D.2), suggesting a natural cutoff for the subspace.

Perplexity of Generations

We thank the reviewer for this suggestion; we have measured the perplexity of the generated corpus on WikiText under PaSS and DiffMeans (as a baseline intervention) as follows:

Wikitext:

PersonaMethod−10+1
capitalizationDiffMean9.209.289.44
PaSS9.449.289.42
chineseDiffMean9.249.289.40
PaSS9.579.289.44
frustratedDiffMean9.319.289.31
PaSS9.409.289.37
refusalDiffMean9.409.289.24
PaSS9.669.289.33

The perplexity remains stable in the +-4% range of the baseline and unsteered responses, which shows that the model does not degrade the text due to modulation.

We shall add this analysis and additional baselines in the camera-ready version of the paper.

Reviewer_7FH8OpenReview ↗

Official Comment by Reviewer 7FH8

Thank you for the response and the additional experiments. However, my main concerns remain:

  1. Novelty over AlterRep. Part of this concern is resolved. The authors acknowledge the overlap, promise to cite and discuss AlterRep, and correctly note that the operators are not identical: AlterRep forces each INLP component to the sign of a target class, while PaSS keeps the sign of each component and rescales it. However, the overlap is larger than the response suggests. AlterRep uses the same decomposition (h = h_N + Σ_w h_w over INLP directions), intervenes at a single layer and continues the forward pass, and already has a scalar α that "enhances or dampens the effect of the intervention". Its appendix sweeps α and presents single-parameter control of intervention strength as an advantage. When the persona components of an input already point the persona way, which is the modulation setting the paper targets, AlterRep's positive and negative counterfactuals are exactly PaSS with β = α and β = −α. The remaining differences are small: sign preservation on tokens with mixed-sign components, and the β = 0 case, which is INLP erasure. The other differences given (decoder LLMs, mean-pooled sequence-level probes, open-ended generation, transfer and IFEval results, and framing "modulation" as a separate setting) concern the application and evaluation rather than the method. The novelty concern therefore remains: the core operator of PaSS is largely covered by AlterRep, and the method's contribution beyond it is incremental.
  2. The new baselines do not favour PaSS. CHaRS gives stronger suppression and amplification than PaSS on both emoji (−61.0/+79.7 vs −54.0/+71.4) and refusal (−38.8/+39.5 vs −4.3/+1.1). On refusal, SEA also outperforms PaSS (−10.3/+5.9). PaSS barely changes refusal, which matches the non-significant refusal row in Table 2.
  3. The LoReFT comparison is not fair as set up. LoReFT has no β. Extrapolating a learned edit with a (1−β) factor to negative values will likely collapse content no matter what the method is. A fair comparison would train LoReFT separately for suppression and amplification, and report rank, training data and hyperparameters. Also, at β = 1 both methods should be the identity, yet the numbers differ (e.g. AdvBench content 0.80 vs 0.75). This suggests either a setup issue or noise of about the same size as the reported differences.
  4. The benefit of the subspace is still not isolated. The first INLP direction closely matches DiffMeans (cos 0.763), and Section 4.2.1 attributes the gain over baselines to the projection-scaling operator rather than the subspace. A rank-1 PaSS (projection-scaling along DiffMeans only) and a random rank-k subspace are needed to show that the extra directions matter. The subspace analysis in the response restates existing results (Table 11, §5, App D.2) and adds no new experiments.
  5. Perplexity: the perplexity table seems to use a different β convention from the paper (the "0" column equals unsteered, but in the paper β = 0 is erasure). Please clarify, and also report perplexity at ±2.

I appreciate the effort, but I do not yet find the evidence sufficient to establish a highly novel general mechanism with scope beyond existing INLP approaches, so I am keeping my score.

Reviewer_MFQaOpenReview ↗
Rating
3
Confidence
3
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper proposes a method to identify subspace, as opposed to just one direction vector, corresponding to a persona (e.g., sycophancy), and use it to steer an LLM.

The author proposes a clever method to build such a subspace:

  • Give contrastive prompts to a model (e.g., with and without instructions like "be sycophantic")
  • Extract activations from them.
  • Train a linear regression classifier (w dot h) that classifies whether or not it is with instructions
  • After you train a model you obtain w that represents a trait (let this vector be w0)
  • You project activation to space orthogonal to w and train another classifier to get another vector, and call it w1
  • Repeat this until the classifier's accuracy becomes chance rate (50%)

A series of vectors w0, w1, ..., wk is thus extracted, which together represent subspace representing a trait. The method is called PaSS.

Experiments show that the proposed method outperforms other baselines (DiffMeans Vec, DiffMeans PCA, and ReFt Vec). PaSS is able to elicit specified traits more strongly than other methods (Table 2).

Critical strengths
  • It proposes an idea that persona is represented by a subspace and a method to compute it
  • It analyzes extracted subspace from several interesting angles
Critical weaknesses
  • As far as I can tell, the paper does not show the proposed method does not degrade coherent of responses
  • Also, Table 2 shows it is able to control traits more strongly than others, but does not mention what happens on their task performance and coherence
Questions for authors

Thank you for submission. I liked the proposed concept of "subspace representing persona", a very natural and important generalization of persona vector. I enjoyed reading the paper. A few questions.

  • The paper does not seem to mention what happens on task performance in the experiment explained in Table 2. What happens on task performance?
  • More generally it does not seem to analyze its effect on coherence of generate responses. Do you have some data showing it give a better coherence-trait tradeoff than other methods?
  • Any comparison to SAE-based methods?

An open-ended question:

  • Do you have some interpretation on individual vectors constituting a subspace?

Minor comments:

  • L310 x_person -> x_{person}
2 comments
AuthorsOpenReview ↗

We show content fidelity, generated perplexity, and compare against SAE-based steering method

We thank the reviewer for recognizing the core contribution of our work, which involves representing personas as a subspace. We are glad that the reviewer has liked our method of extracting subspaces and the in-depth subspace-peeled analysis. The reviewer has raised some concerns regarding the content coherence, which we wish to clarify below:

Content Fidelity of the Responses and Content Scores for Table 2

In Table 3, we have reported how much content changes when projections are transferred from one domain to another. However, as per the reviewer’s suggestion, we are sharing the content fidelity scores for the Table 2 generations (on MATH-500) below:

PersonaLlama-3-8B BaselineLlama-3-8B SuppressionLlama-3-8B AmplificationQwen BaselineQwen SuppressionQwen Amplification
Chinese0.190.160.200.560.570.58
Refusal0.130.160.150.570.450.52
Frustrated0.250.170.260.560.550.52
Capitalization0.220.230.210.520.590.40
Emoji0.270.260.310.380.500.40
Indian culture0.310.240.270.590.560.61

We observe that content fidelity remains stable under modulation across baselines and improves on several personas, with the largest drops observed for Qwen refusal under suppression (0.57 → 0.45) and Qwen capitalization under amplification (0.52 → 0.40). We find no degenerate text, but observe two patterns: refusal suppression generally degrades the reasoning quality [1,2,3], and capitalization amplification nearly doubles the rate at which the model states an answer before deriving it (26% → 48%), consistent with evidence that answering before intermediate reasoning reduces mathematical accuracy [4].

Additionally, we have also measured the perplexity of Wikitext generations under PaSS and DiffMeans.

PersonaMethod−10+1
capitalizationDiffMean9.209.289.44
PaSS9.449.289.42
chineseDiffMean9.249.289.40
PaSS9.579.289.44
frustratedDiffMean9.319.289.31
PaSS9.409.289.37
refusalDiffMean9.409.289.24
PaSS9.669.289.33

The perplexity remains stable in the +-4% range of the baseline and unsteered responses, which shows that the model does not degrade the text due to modulation.

Interpretation of Individual Vectors in the Subspace

Our peel analysis suggests that the directions within a persona subspace form a hierarchy where dominant features are captured first, followed by finer sub-behaviors. The first INLP direction closely aligns with both the DiffMeans direction (cos = 0.763) and the first principal component (cos = 0.730), while later directions are nearly orthogonal to both (Table 11). This suggests that the first direction recovers what single-vector steering methods find, and later directions capture what they miss.

These finer sub-behaviors appear in our peel analysis, where, for refusal, physical harm and harassment, more directions are required to suppress than fraud or cyber harm, and directive requests flip earlier than abstract ones (§5.1). For emoji, the first direction introduces most emojis, while later directions refine their placement and recurrence (§5.2).

Finally, beyond the point where the persona stabilizes, additional directions increasingly capture content-related information (App. D.2), suggesting a natural cutoff for the subspace.

Comparison with SAE method

We thank the reviewer for this suggestion. We compared the emoji persona modulation on SAE using. EleutherAI's pretrained TopK sparse autoencoder for Llama-3-8B (the base model, not instruct). We selected the k ∈ {1, 5, 20} SAE latents that best separate persona from neutral responses on MATH train. We then either zeroed/doubled their codes (ablate) or subtracted/added their decoder direction with norm matched to PaSS (steer).

PersonaMethod−1+1
EmojiPaSS (ours)−54.0+71.4
EmojiSAE ablate k=1+23.3−4.7
EmojiSAE ablate k=5−20.9−6.1
EmojiSAE ablate k=20−74.3+8.8
EmojiSAE steer k=1−22.0+41.3
EmojiSAE steer k=5−29.7+36.8
EmojiSAE steer k=20−27.2+28.4

PaSS is the only method strong in both directions, and SAE ablation at k=20 suppresses more.

We acknowledge the typo identified by the reviewer; we shall correct it in the camera-ready version.

References

[1] Turner, Alexander Matt, et al. "Steering language models with activation engineering." arXiv preprint arXiv:2308.10248 (2023).

[2] Steering Llama 2 via Contrastive Activation Addition (Rimsky et al., ACL 2024)

[3] Stickland, Asa Cooper, et al. "Steering without side effects: Improving post-deployment control of language models." arXiv preprint arXiv:2406.15518 (2024).

[4] Let Me Speak Freely? A Study On The Impact Of Format Restrictions On Large Language Model Performance. (Tam et al., EMNLP 2024).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Rebuttal and Revisions.

We would like to thank the reviewers, the Area Chairs, and the Program Chairs for their dedicated time and effort in providing thoughtful feedback on our paper and organizing this discussion. We have uploaded a revised PDF in which all new or changed material is marked in pink and have summarized the rebuttal as follows:

We are glad that the reviewers have found the idea of persona modelled as a subspace natural and important (MFQa, 7FH8), the method simple, intuitive, and cheap to apply (7FH8, QG9f), and the peel analysis and experiments in-depth (MFQa, 7FH8, QG9f).

Novelty of our contributions (7FH8, QG9f).

The reviewers raised some issues regarding the technical novelty of using INLP. As we have shared, PaSS does not claim novelty in introducing generalized subspace modulation or claim novelty over INLP's method. Our contributions are as below:

  1. Personas as subspaces: We model each persona as a multi-dimensional subspace, that can be extracted as a subspace from a model's activations.

  2. Modulation. We formalize modulation as a setting where the persona is already present in the input, and the intervention scales it, such that β=1 is the identity while erasure is represented as β=0.

  3. INLP as a probe in LLM activations: We use INLP (Ravfogel et al., 2020) to extract these subspaces from the residual stream of decoder LLMs. To our knowledge, INLP subspaces have not been used to modulate personas in the activation-steering and persona-vector literature, although closely related INLP-based counterfactual flipping interventions exist for masked language models (AlterRep) and, concurrently, for refusal steering without modulation (Rocchetti & Ferrara, 2026).

  4. Generalization without target outputs. We show that extracted subspaces transfer across domains, generalize to non-linear probes and to multi-instruction settings, and require only persona-present and persona-absent prompts, not target responses.

  5. Nested structure. Our peel analysis demonstrates how the directions form a hierarchy of sub-behaviors, which makes the subspaces interpretable

Additional baselines (MFQa, 7FH8).

We added multiple additional baselines during the rebuttal, such as LoReFT (activation variant), SEA, CHaRS, ACT, and SAE-based steering.

  1. Subspace Steering PaSS outperforms ACT on emoji in both directions and on refusal suppression. While CHaRS and SEA amplify more than PaSS in some settings, PaSS’s advantage is transferability across domains.

  2. SAE-based Steering We observe that PaSS is the only method that is strong in both directions of modulation against the SAE variants.

  3. LoReFT Steering Against the LoReFT activation variant, which is only comparable at suppression (β = 0), we see that PaSS suppresses more strongly on three of four datasets while maintaining a better content score than LoReFT.

Content fidelity and perplexity (MFQa, 7FH8).

To demonstrate content coherence, we ran additional experiments measuring the fidelity of the responses and measuring perplexity on WikiText. We observe that,

  1. We show consistent content fidelity for both suppression and amplification of persona subspaces across most personas and model combinations.

  2. On WikiText, perplexity stays within about 4% of the unsteered value, slightly above DiffMeans in 7 of 8 cells.

Do the extra directions matter? (7FH8, MFQa).

We construct a controlled comparison of PaSS against DiffMeans and with random subspaces of the same rank. We observe that suppression emerges only from the specific INLP subspace, not from the size of the intervention or from any single later direction. This suppression saturates after a moderate rank. We also report how the first direction aligns with DiffMeans and the nested structure from the peel analysis.

Generalizability of PaSS (QG9f)

We added transfer results for all 13 persona-domain pairs, qualitative examples, the rationale for the six personas, results on a larger model (Qwen3-32B), and an opposite-sign analysis for the frustrated persona.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.