acceptodds
Under review as a conference paper at ICLR 2027

Same Task, Different Answer: Persona Conditioning Modulates Shared Task Mechanisms in Language Models

Abstract

Persona conditioning can change a language model's behavior on a task even when the task itself is held fixed, but it remains unclear whether these changes arise from persona-specific mechanisms or from modulation of ordinary task computation. We investigate this question across three instruction-tuned language models, five tasks, and four persona axes. Applying complementary damage and repair path patching to persona-conditioned correct–wrong transitions, we identify attention-head circuits and decompose each into a component shared across axes () and an axis-dependent remainder (). Circuits discovered in opposite causal directions converge preferentially on , and faithfulness, knockout, direct logit attribution, probing, and no-persona validation consistently implicate in task-related computation. Moreover, the bidirectional intersections retain strong task-functional signatures and much of the intervention effect while using substantially fewer heads. As an independent validation, in Qwen-7B SST-2 every persona-derived shared head falls within a separately discovered no-persona task circuit, whereas the remainder heads show much lower containment. The same structure is also behaviorally actionable: a signed intervention selectively improves persona-sensitive accuracy while preserving more than 99% of originally correct outputs. However, how strongly persona information couples to this shared computation depends on the task: the behaviorally stable ARC tasks show greater separation between task and persona information, whereas ETHICS, Safety, and SST-2 show stronger coupling within . Overall, these results show that persona-conditioned behavior can emerge from modulation of shared task-related mechanisms, not only from isolated persona-specific circuitry.

Primary area:
interpretability and explainable AI
Keywords:
Mechanistic Interpretability, Circuit Discovery, Persona Conditioning, Large Language Models, Causal Intervention

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_jzBnOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper studies why changing user persona information can change a model’s answer to the same question. The authors use patching in both directions, from correct to wrong runs and vice versa, to identify relevant attention heads. They separate these into heads shared across persona axes K and the remaining heads R. Experiments suggest that K also supports ordinary task computation, and that persona information interacts with these heads differently across tasks. The authors support this with ablations, probing, logit attribution, and evaluations without persona prompts. They also show that targeted activation interventions can fix some persona-sensitive errors while preserving most originally correct answers.

Critical strengths
  1. The research question is interesting. The definition of problem is clear and easy to follow. The paper measures how persona information affects answer correctness, investigates the underlying mechanisms, and uses steering to fix some of the errors. This makes the findings useful beyond just observing that models are sensitive to persona prompts.
  2. The authors use several experiments to check what the identified heads actually do. The experiments are comprehensive and easy to follow. In particular, showing that these heads still support the task without persona information gives stronger support to the claim that persona effects involve ordinary task computation.
Critical weaknesses
  1. The experiments are limited to relatively simple classification and multiple-choice tasks. It is unclear whether the findings also apply to complex reasoning or open-ended generation, which limits what we can conclude about their usefulness for improving model capabilities.
  2. There is no controlled comparison across model sizes. The overlap between task and persona information could be more pronounced in smaller models. Testing different sizes from the same family, such as Qwen, would help clarify whether the findings hold as models get larger.
Questions for authors
  1. How much do the K sets overlap across tasks? Have you tested whether K discovered for one task also supports another? This would help clarify whether these heads perform task-specific computation or more general answer selection and label prediction.

  2. Since the keep/knockout experiments only intervene at the decision position, how do you distinguish heads involved in solving the task from heads that mainly read out information computed elsewhere? What evidence supports a more specific functional interpretation of K and R?

  3. Do the existing results show any difference between the 3B and 8B models in how separated task and persona information are? Could the observed overlap partly reflect limited model capacity? A controlled comparison across sizes within the same family, such as Qwen, would be a useful additional experiment.

  4. Have the intervention coefficients been evaluated on data separate from the data used to select them? Also, do they transfer to unseen persona descriptions or other tasks? This would help assess how much the repair results depend on setting-specific tuning.

6 comments
AuthorsOpenReview ↗

Response to Reviewer jzBn (Part 1)

We thank the reviewer for their thoughtful evaluation and constructive questions. We address the scope, methodology, and implications of our findings below.

W1. Generalization beyond restricted-output tasks

Thank you for raising this question. Our use of restricted-output tasks is a deliberate experimental choice that enables controlled causal comparisons across persona conditions, with a consistent evaluation metric and a well-defined decision position. Although our experiments cover scientific reasoning, ethical judgment, safety classification, and sentiment analysis, our conclusions are limited to the evaluated tasks and do not establish generalization to complex, open-ended generation.

Extending mechanistic circuit analysis beyond narrowly defined tasks remains an important open research challenge (Sharkey et al., 2025). Recent work has made progress in this direction; for example, Lindsey et al. (2025) demonstrate circuit-tracing approaches for investigating multi-step reasoning.

Open-ended generation introduces additional challenges in constructing controlled counterfactuals, aligning generated sequences, and attributing causal effects across output positions. While path patching is not inherently restricted to single-token outputs, extending our bidirectional discovery and intervention framework to these settings requires additional methodological development.

Our findings therefore demonstrate selective improvements in persona-conditioned task accuracy, rather than general improvements in model reasoning capabilities. We will clarify this scope in Section 7 (Limitations) and identify complex reasoning and open-ended generation as directions for future investigation.

References

  1. Sharkey, L., et al. (2025). Open Problems in Mechanistic Interpretability. arXiv:2501.16496.
  2. Lindsey, J., et al. (2025). On the Biology of a Large Language Model. Transformer Circuits Thread.

W2/Q3. Model scale and the capacity hypothesis

We compare Llama-3.2-3B-Instruct and Llama-3.1-8B-Instruct. Although they share a lineage, differences in training and architecture prevent a fully controlled scaling comparison.

Behavioral sensitivity. Llama-8B exhibits lower mean persona-conditioned instability than Llama-3B (2.95% vs. 4.71%).

Task–persona organization. We compare task and persona decodability using balanced-accuracy differences (BAK−BARBA_K-BA_R), alongside K/R ratios of persona-token attention, persona-dependent value-content changes, and attention-routing changes. Positive probe differences favor KK; ratios above 1 indicate stronger effects in KK.

Model comparison (Original discovery)

TaskModelTask probe ΔBAPersona probe ΔBAPersona attention K/RContent K/RRouting K/R
ARC-ChallengeLlama-3B+0.168−0.0340.1990.0720.084
Llama-8B+0.166−0.0340.1640.0800.102
ARC-EasyLlama-3B+0.197−0.0270.2550.0760.076
Llama-8B+0.128−0.0530.2760.0990.114
ETHICSLlama-3B+0.042+0.1141.3770.3880.406
Llama-8B+0.046+0.1451.3230.5500.542
SafetyLlama-3B+0.010+0.1862.0060.9620.900
Llama-8B+0.058+0.2231.2730.6470.683
SST-2Llama-3B+0.075+0.1391.2720.8920.837
Llama-8B+0.125+0.1411.4880.5510.597

Results are aggregated across four persona axes. Probe differences measure relative decodability; K/R ratios measure relative persona-related attention, content, and routing effects.

Findings. Both models exhibit similar task-dependent organization: task-answer information is more decodable from KK, while persona information is relatively more decodable from RR on ARC and from KK on ETHICS, Safety, and SST-2. Attention, content, and routing measurements also vary by task.

However, Llama-8B does not consistently exhibit greater separation than Llama-3B. Its content and routing K/R ratios are lower on Safety and SST-2 but higher on ARC and ETHICS. Reverse discovery also shows mixed differences.

Overall, Llama-8B is behaviorally more stable, but mechanistic measurements do not show uniformly greater task–persona segregation. Thus, our findings do not support a simple capacity-only explanation, although capacity effects remain possible. Controlled comparisons across model sizes, such as within the Qwen series, are needed to assess scaling.

AuthorsOpenReview ↗

Response to Reviewer jzBn (Part 2)

Q1. Cross-task overlap and functional transfer

Thank you for raising this question. Although we have not directly evaluated cross-task functional transfer, we examined cross-task structural overlap and the functionality of globally shared heads.

Cross-task structural overlap. For each task tt, we define the bidirectional shared intersection as KK(t)=KO(t)∩KRev(t)KK(t)=K_O(t)\cap K_{\mathrm{Rev}}(t). We then identify the global core G=⋂tKK(t)G=\bigcap_t KK(t), containing heads shared across all five tasks and both discovery directions.

ModelOriginal K JaccardReverse K JaccardGlobal core heads
Llama-3B0.4070.3312
Llama-8B0.3170.2652
Qwen-7B0.2650.2881

Jaccard values are averaged across the 10 task pairs per model.

These results reveal partial cross-task recurrence, but only 1–2 heads per model belong to the global intersection. Thus, KK is not an identical universal circuit across tasks.

Functional comparison. We additionally compare task-specific KK(t)KK(t) with the global core using normalized probability-margin faithfulness and knockout effects across all 40 matched settings.

Head groupFaithfulness ↑Normalized knockout effect ↑
Task-specific KK0.9870.792
Global core0.0860.042

Because the global core contains substantially fewer heads than task-specific KKKK, this comparison establishes the limited functionality of the small universal intersection but does not isolate task specificity from circuit size.

Task-specific KKKK exhibits substantially stronger functional effects than the global core. This indicates that the small universal intersection alone is insufficient to reproduce the functionality of task-specific KKKK, suggesting that additional task-dependent components contribute to performance.

Cross-task transfer. While these analyses establish partial structural recurrence and the functional importance of task-specific KKKK, directly testing whether KK discovered for task tt supports another task t′t' requires cross-task interventions with appropriate size- and layer-matched controls.

Overall, our findings show that KK is not a universal set of identical heads across tasks, and task-specific KKKK has substantially stronger functional contributions than the small globally shared core. These results support task-dependent functional organization while leaving open the possibility that some identified heads also participate in general answer selection or label prediction. We will clarify this distinction in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer jzBn (Part 3)

Q2. Task computation versus decision-position readout

Thank you for raising this distinction. Our faithfulness, knockout, DLA, and independent no-persona validation support KK's task functionality. We conducted additional causal experiments to investigate its role in persona-conditioned behavioral changes.

  1. Causal cross-patching of K and R.

    We performed bidirectional cross-patching between matched correct and wrong persona-conditioned runs, measuring the fraction of the gold-answer margin gap recovered by each component.

    Head groupMedian recovered margin gap ↑
    Original K95.5%
    Reverse K95.5%
    KK intersection91.6%
    Original R16.8%
    Reverse R19.5%

    Evaluation pairs are disjoint from discovery pairs. KK recovers substantially more of the margin gap than RR, despite generally containing fewer heads. This provides direct causal evidence of KK's involvement in persona-conditioned prediction changes, beyond structural overlap or decodability.

  2. Persona-reading-head analysis.

    We independently identify candidate persona-reading heads using persona-token attention (PP) and persona-conditioned activation responsiveness (P′P'), then evaluate their overlap with KK and RR through cross-patching.

    Head group (Original discovery)SettingsMedian headsRepair ↑Degradation ↑
    K excluding P58130.9430.932
    K excluding P′58120.9380.933
    P overlapping K1310.0640.103
    P′ overlapping K2320.0950.081
    R excluding P58250.1650.166
    P overlapping R5020.0020.003

    Values are median normalized causal effects across available settings; empty intersections are excluded.

    KK retains approximately 94% of its cross-patching effect after excluding candidate persona readers, whereas their intersections with KK show much smaller effects. Thus, KK's causal contribution is not confined to these identified readers, although the intersections are smaller and not size-controlled.

  3. Upstream persona-reader → K intervention.

    We mean-ablate upstream persona readers preceding the earliest KK layer, comparing reductions in KK's cross-patching effect against 10 matched random controls per setting.

    Using an upstream-ablated receiver baseline, the additional reduction is negligible across 53 settings (median +0.000310; stronger than random in 30/53). Alternative normalization yields a similar conclusion (−0.000145; 26/53).

    These results provide little evidence that the selected upstream readers form a consistently necessary pathway into KK, although other pathways remain possible.

  4. Direct persona-reader contribution removal.

    We selectively remove complete persona-token contributions, their persona-specific and persona-invariant components, and persona-dependent task-source contributions from identified reader heads.

    Across 60 settings, all four interventions produce zero median change in accuracy and instability, with negligible persona-margin changes (−0.00024 to −0.00049). Random-head controls were also evaluated.

    Together with upstream ablation, these results provide little evidence that the selected direct persona-reading contributions are consistently necessary for persona-conditioned behavioral changes.

  5. Activation-direction nullification.

    We nullify activation directions distinguishing correct and wrong persona-conditioned states within KK and RR, comparing their effects against random-direction controls.

    Nullified componentΔAccuracy (pp)ΔInstability (pp)
    K+17.30+3.60
    R+0.15+0.28
    K random directions≈0≈0

    Nullifying KK directions produces substantially larger effects than RR or random controls, supporting their causal contribution to task predictions. However, the increased instability indicates that this is not necessarily a robustness improvement.

  6. Functional interpretation.

    Together, these experiments support interpreting KK as shared task-functional circuitry involved in task predictions and persona-conditioned behavioral changes. Its contribution is not confined to the identified persona readers, and the tested upstream pathways show no consistent necessity.

    However, decision-position interventions cannot distinguish upstream reasoning from evidence integration or answer readout. Thus, KK may support multiple task-related functions rather than independently performing the complete reasoning process. Similarly, RR is axis-dependent but not necessarily dedicated to persona processing.

We will incorporate these findings and clarify the functional interpretation of KK and RR in the revised manuscript.

Reviewer_9RbpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper combines ideas from mechanistic interpretability with persona conditioning. The core idea of the paper is based on the observation that for certain types of tasks for LLMs, the response (or the "answer") is independent of the persona context provided. Which means there is an objective answer for that query prompt. Empirically, it has been observed that the persona conditioning does affect the response. In fact, certain persona attributes lead to an incorrect answer even if the model knows the correct response (as determined by querying without persona conditioning). The paper proposes the use of circuits mechanism of mechanistic interpretability for understanding this phenomenon and also mitigating this phenomenon. The investigation is done using three small models on five tasks using person conditioning across four axes. The key findings based on attention head circuits are that there is a task dependent common component shared across correct and incorrect responses. Plus there is persona axis dependent remainder that modulates the eventual response. The strength of these components is dependent upon the task as well. The also propose an intervention based on the two components that can mitigate the effect. The overall finding is that in these LLMs, the shared task related mechanism is entangled with the persona circuitry and the outcome is a result of these joint interaction.

Critical strengths

(1) The phenomenon of persona-conditioned response is well known but coupling it with circuits based mechanistic interpretability is novel and interesting.

(2) The analysis which shows that K behaves as a task-related causal core and R changes according to persona and that it modulates the task-directed output in different directions is also very interesting.

(3) While not explicitly stated, these findings can potentially be used to make the LLM responses more robust to persona conditioning where it matters i.e. object questions not dependent on persona. of course this should be done without damaging the ability for cases where one wants the response to change due to the assumed persona.

Critical weaknesses

(1) The paper is very poorly written and it is extremely difficult to read and understand. Variable are introduced suddenly without any definition or description, right since line 161 where {n_B} has been suddenly introduced without telling the reader what it is. This sloppy writing persists throughout the paper and the onus is on the reader to figure out what the authors are saying. Another example is the Section 3.3 which is replete with terms and equations with no explanation. This pattern persists throughout the manuscript. This reviewer had to make two careful passes of the manuscript before the third time where the writing finally started to make sense. This is unacceptably poor writing quality for ICLR.

(2) There are lot of assumptions that are tacitly made without any arguments or proofs. For example, line 187 states that "Strongly negative scores...computation". This is presented like self-evident truth. It is not at all clear why is this statement true. Such assertions are made throughout the paper and are too numerous to catalog.

(3) The models studied are all small models. Do the observed phenomenon scale to larger models?

(4) The five research questions posed in lines 073-091 have never been explicitly answered in the paper. Please directly answer Rq1-RQ5.

Questions for authors

(1) I actually like the scientific investigation of this work but it has not at all been presented in a scientific manner. Please fix the writing issues in the paper.

(2) For fixing assertions made without justification (arguments or data or proof), please read the paper like a skeptic and try to convince a sympathetic skeptic.

(3) In section 5.3 beginning, it is stated that universal K=task and R=persona decomposition is not true. Then how useful is this entire analysis? Does it really lead to more robust models?

(4) One can get the impression that many parts of the work have slightly manifested in these earlier papers:

https://papers.ssrn.com/sol3/papers.cfm?abstract_id=7264360 https://www.mdpi.com/2504-2289/10/4/102 https://arxiv.org/abs/2509.19727 https://www.researchgate.net/publication/415568494_Task_Context_Shapes_Component_Roles_in_Language_Models

Is this impression correct? Please comment.

6 comments
AuthorsOpenReview ↗

Response to Reviewer 9Rbp (Part 1)

Thank you for the detailed review and for recognizing the value of our scientific investigation. We appreciate the concerns regarding presentation and clarity, and will address them in the revised manuscript. Below, we respond to each point, including the specific examples raised.

W1 and Q1. Notation and presentation clarity

Thank you for highlighting these presentation issues. We agree that the original manuscript introduced several concepts and symbols without sufficient explanation. We will address these concerns through the following revisions:

  1. Background and notation. We will add a concise Preliminaries and Notation subsection introducing causal circuit discovery, bidirectional path patching, and faithfulness/knockout evaluation, while defining symbols at their first occurrence. A complete notation table will be provided in the appendix.

  2. Concrete clarification of Section 3.2. We will explicitly define nB=∣B∣n_B=|B| as the number of persona-sensitive items, where BB contains task items answered correctly under at least one persona value and incorrectly under another, before introducing the reliability threshold nB≥25n_B\geq25.

  3. Concrete clarification of Section 3.3. We will explain the probability-margin metric and the interpretation of Original and Reverse patching scores before presenting the circuit-discovery equations.

    The margin MM measures the probability assigned to the reference answer relative to the strongest competing answer. A positive margin indicates that the reference answer is preferred.

    • Original: EO(h)=Mcorrect−MpatchedE_O(h)=M_{\mathrm{correct}}-M_{\mathrm{patched}}. A positive score indicates that inserting the wrong-run activation decreases the gold-answer margin.
    • Reverse: ERev(h)=Mwrong−MpatchedE_{\mathrm{Rev}}(h)=M_{\mathrm{wrong}}-M_{\mathrm{patched}}. A negative score indicates that inserting the correct-run activation increases the gold-answer margin.

    We will also clarify that KK and RR are structural definitions, representing heads shared across persona axes and their axis-dependent remainders. Their functional roles are established empirically rather than assumed. Our findings show that KK supports ordinary task performance and that persona-conditioned behavioral changes can be mediated through these shared task-functional components.

We will incorporate these revisions into Section 3, add the full notation table to the appendix, and ensure that notation and methodological interpretations are clearly explained throughout the manuscript.

AuthorsOpenReview ↗

Response to Reviewer 9Rbp (Part 2)

W2 and Q2. Justification of mathematical and empirical claims

We appreciate the reviewer's concern. We agree that mathematical implications, empirical observations, and mechanistic interpretations should be explicitly distinguished and justified. We address the specific example in line 187 below and identify the evidence supporting our broader claims.

  1. Mathematical justification of Reverse patching.

    For a task input xx with reference answer y⋆y^\star, let Y\mathcal{Y} denote the set of possible task answers. We define the gold-answer probability margin for a model run rr as

    Mr=Pr⁡r(y⋆)−max⁡y∈Y, y≠y⋆Pr⁡r(y),M_r = \Pr_r(y^\star) - \max_{y\in\mathcal{Y},\,y\neq y^\star}\Pr_r(y),

    where Pr⁡r(y)\Pr_r(y) is the probability assigned to answer yy under run rr. A positive margin means that the reference answer has the highest probability; a negative margin means that another answer is preferred.

    Our counterfactual pairs hold the task input and reference answer fixed while varying the persona, producing a correct run and a wrong run. Let zhcorrectz_h^{\mathrm{correct}} and zhwrongz_h^{\mathrm{wrong}} denote attention-head hh's output activations at the decision position in these two runs.

    In Reverse patching, we replace the activation of head hh in the wrong run with its corresponding activation from the correct run. Equation (2) defines the effect as

    ERev(h)=Mwrong−Mpatched(zhcorrect→zhwrong).E_{\mathrm{Rev}}(h) = M_{\mathrm{wrong}} - M_{\mathrm{patched}(z_h^{\mathrm{correct}}\rightarrow z_h^{\mathrm{wrong}})}.

    Because the original wrong run predicts an incorrect answer, Mwrong<0M_{\mathrm{wrong}}<0. If the patched activation improves the gold-answer margin, then

    \quad\Longrightarrow\quad E_{\mathrm{Rev}}(h)<0.$$ Therefore, a more negative Reverse score corresponds to a larger increase in the gold-answer margin. This is an algebraic consequence of the score definition, not an assumed property of the discovered heads. **Importantly, an improved margin does not necessarily mean that the prediction becomes correct.** For example, if $M_{\mathrm{wrong}}=-0.30$ and $M_{\mathrm{patched}}=-0.10$, then $E_{\mathrm{Rev}}(h)=-0.20$. Patching has improved the margin, but the model still prefers an incorrect answer. Accordingly, we will replace the statement at line 187 with: > "Strongly negative Reverse scores identify heads whose correct-run activations, when inserted into the wrong run, increase the gold-answer probability margin. Such patches contribute toward the correct answer but do not necessarily correct the final prediction." We will incorporate this derivation and clarify all variables immediately after Equation (2) in Section 3.3.
  2. Empirical and mechanistic justification.

    Beyond the Reverse patching example, we will clarify the evidence and limitations underlying specific claims in the manuscript:

    SectionSupporting evidenceWhat we will clarify
    4.3 Structural recurrenceOriginal–Reverse Jaccard similarity is higher for KK than RR (0.701 vs. 0.099; Figure 4).Structural overlap establishes recurrence across discovery directions, not functional equivalence.
    5.1 Task functionalityKK achieves higher faithfulness (0.976/0.979 vs. 0.728/0.743) and stronger knockout effects than RR (Table 2).These causal ablations establish task-functional involvement but do not distinguish upstream reasoning from decision-position readout.
    5.2 Task-directed outputDLA shows larger task-directed changes in KK than RR during behavioral transitions (Table 3).DLA characterizes task-directed output contributions but does not independently establish causal mediation.
    5.3 Task-dependent organizationProbing and QK/value analyses reveal task-dependent differences in persona–task organization (Figure 5; Appendix F.1).These observations do not establish that representational segregation causes behavioral stability.

    We will incorporate these clarifications throughout Sections 4–6, distinguishing mathematical implications, causal evidence, and mechanistic interpretations. Together, the findings support our central conclusion that persona-conditioned behavioral changes can be mediated through shared task-functional components, without establishing a unique computational pathway.

AuthorsOpenReview ↗

Response to Reviewer 9Rbp (Part 3)

W3. Generalization across model sizes

We agree that evaluating substantially larger models would strengthen the generalizability of our findings. Our current study evaluates three instruction-tuned models spanning 3B–8B parameters and two model families. Although this does not establish behavior at larger scales, it allows us to examine whether the central findings are consistent across the evaluated models.

Within the Llama 3 family, we compare Llama-3.2-3B-Instruct and Llama-3.1-8B-Instruct. Importantly, Llama-3.2-3B was derived through structured pruning and knowledge distillation from larger Llama models (Meta AI, 2024). This provides a shared model lineage, but not a controlled parameter-scaling experiment, as the models also differ in their training and compression procedures.

Our results show two distinct patterns:

  1. Behavioral sensitivity: Llama-8B exhibits lower mean persona-conditioned instability than Llama-3B (2.95% vs. 4.71%), with statistically significant persona effects in 5/20 versus 10/20 model–task–axis settings, respectively.

  2. Structural recurrence: Despite this behavioral difference, Original–Reverse recurrence remains substantially stronger for KK than RR in all 20/20 settings for each model. Mean Jaccard similarity between Original and Reverse KK is 0.668 for Llama-8B and 0.706 for Llama-3B (Figure 4). The same KK-over-RR recurrence pattern holds for Qwen-7B.

These results show that the shared-component recurrence phenomenon persists across the evaluated model sizes and families, even when their behavioral sensitivity differs. However, they do not establish that persona sensitivity decreases monotonically with model size or that the mechanism necessarily generalizes to substantially larger models. Further evaluations across larger models would be needed to establish whether this pattern persists.

W4. Explicit answers to the research questions

Thank you for this suggestion. Although our results are organized around the five research questions, we agree that their answers should be stated more explicitly. We will make the answer to each research question explicit through a concise concluding synthesis in its corresponding Results subsection, directly linking the conclusion to the supporting evidence.

Our main answers are:

  • RQ1: Persona-conditioned behavior: Persona effects vary across tasks and models, with statistically significant effects in 27/60 settings after multiple-testing correction. ETHICS, Safety, and SST-2 are more sensitive, while ARC remains comparatively stable.

  • RQ2: Selective intervention: Our signed K/RK/R intervention improves persona-sensitive accuracy by +12.00 pp on development data and +8.48 pp on held-out test data, while changing accuracy on originally all-correct development items by only −0.31 pp. Development improvements are significant in 29/33 reliable settings.

  • RQ3: Bidirectional recurrence: Original and Reverse circuits converge substantially more strongly in KK than RR (mean Jaccard 0.701 vs. 0.099), with K>RK > R in all 60 settings. Their intersection provides a more compact intervention target while retaining substantial effectiveness.

  • RQ4: Functional organization: Faithfulness, knockout, and DLA establish KK's stronger task functionality. Despite larger persona-dependent routing and content changes in RR, persona attention and decodability are relatively stronger in RR for stable ARC tasks, but in KK for more persona-sensitive tasks. Cross-patching, nullification, and persona-reader interventions further show that persona-induced answer changes are mediated largely through shared task-functional components, rather than exclusively through separate persona-specific circuitry.

  • RQ5: No-persona task functionality: KK remains task-functional without persona context, achieving faithfulness of 0.965–0.980 versus 0.578–0.639 for matched-random controls. Independent SST-2 task-circuit discovery across three models further confirms that 84.6–100% of persona-derived KK heads belong to independently discovered task circuits, compared with only 8.8–17.6% of RR heads.

We will make these answers explicit in their corresponding Results subsections, followed by the supporting analyses.

Reviewer_pEZ6OpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
1
Soundness
2
Read review
Paper summary

The paper investigates the internal mechanisms of a language model when a persona-based prompt alters the model's prediction on a fixed task. By keeping the task and correct answer constant and changing only the persona, the authors utilize a bidirectional path-patching method to identify the critical attention heads driving this change. In one direction, they patch the internal state of a wrong run into a correct run to disrupt it. In the other, they patch the correct state into the wrong run to repair it. The discovered heads are decomposed into a "shared" group present across all persona types and an "axis-dependent remainder" group specific to individual personas. The primary finding is that the shared group actually comprises the model's ordinary task circuit rather than a separate persona module. Furthermore, applying a small, targeted intervention to these shared parts can repair persona-induced errors without degrading originally correct answers. In one specific evaluation, these persona-derived shared heads were shown to be fully contained within an independently discovered task circuit built without any persona conditioning.

Critical strengths
  • The counterfactuals are well constructed, since only the persona sentence changes while the task and the gold answer are held fixed.
  • The main claim is supported by several independent tests, including faithfulness, knockout, logit attribution, probing, and a separate no-persona experiment.
Critical weaknesses
  • My biggest problem with this paper is that it is very hard to follow, and substantial work is needed to streamline it. For example, the K and R components are defined structurally in Section 3, but their functional meaning is only explained much later (Sections 5–6), so the reader has to carry undefined terms for most of the paper. The research questions are also answered piecemeal across sections, so it is hard to track which section answers which question. In addition, there is no section that properly gives the reader the needed background or sets out the notation conventions used in the paper. Finally, there is no end-to-end worked example: the paper shows one SST-2 sentence flipping under a persona, but never walks through a single example through the whole pipeline, from a persona pair to the discovered circuit to the intervention result.

  • The paper claims that the heads discovered via persona flips are part of the model's ordinary task machinery, but it shows this for only one model and one task. I would suggest running the same check on one or two more model–task pairs.

  • Based on my understanding, the ETHICS dataset is constructed by majority vote among humans, so there is no single true correct answer, and the reasonable answer can change depending on the background and perspective of the person answering. A persona-induced shift toward a minority human view may therefore be a reasonable difference of opinion rather than a failure of the model. The paper's intro admits that "effects are not necessarily undesirable," but then treats such effects as errors to repair anyway, which is questionable for this task.

Questions for authors

N/A

3 comments
AuthorsOpenReview ↗

Response to Reviewer pEZ6 (Part 1)

Thank you for the positive assessment of the counterfactual design and the multi-test evidence, and for the concrete suggestions on presentation and validation. We agree that several aspects of the original presentation made the method harder to follow than necessary. We address each weakness below.

W1. Presentation, notation, and end-to-end explanation

Thank you for these suggestions. We will organize the presentation to introduce the underlying concepts, structural definitions, and research questions before presenting their empirical interpretations.

To improve clarity within the page limit, we will make three revisions:

  • Background and notation. We will add a concise preliminaries subsection introducing causal circuit discovery, bidirectional path patching, and faithfulness/knockout evaluation. All symbols will be defined at their first occurrence, with a complete notation table in the appendix. In particular, we will clarify the evaluation subsets (B,U,W)(B, U, W), the reliability threshold (nB=∣B∣)(n_B=|B|), and the definitions and interpretations of Original/Reverse patching scores and the shared (K)(K) and remainder (R)(R) components.

  • Structural definitions and RQ organization. In Section 3.3, we will introduce KK as the cross-axis shared component and RR as the axis-dependent remainder, together with their hypothesized functional roles. These are structural definitions, not predetermined functional labels. Our results reveal that KK is more task-functional, while persona influence can operate through KK itself rather than exclusively through a separate persona-specific component. We will make the answer to each research question explicit through a concise concluding synthesis in its corresponding Results subsection, directly linking the conclusion to the supporting evidence.

    • RQ1: Persona-conditioned behavior: Persona effects vary across tasks and models, with statistically significant effects in 27/60 settings after multiple-testing correction. ETHICS, Safety, and SST-2 are more sensitive, while ARC remains comparatively stable.
    • RQ2: Selective intervention: Our signed K/R intervention improves persona-sensitive accuracy by +12.00 pp on development data and +8.48 pp on held-out test data, while changing accuracy on originally all-correct development items by only −0.31 pp. Development improvements are significant in 29/33 reliable settings.
    • RQ3: Bidirectional recurrence: Original and Reverse circuits converge substantially more strongly in KK than RR (mean Jaccard 0.701 vs. 0.099), with K > R in all 60 settings. Their intersection provides a more compact intervention target while retaining substantial effectiveness.
    • RQ4: Functional organization: Faithfulness, knockout, and DLA establish KK's stronger task functionality. Despite larger persona-dependent routing and content changes in RR, persona attention and decodability are relatively stronger in RR for stable ARC tasks, but in KK for more persona-sensitive tasks. Cross-patching, nullification, and persona-reader interventions further show that persona-induced answer changes are mediated largely through shared task-functional components, rather than exclusively through separate persona-specific circuitry.
    • RQ5: No-persona task functionality: K remains task-functional without persona context, achieving faithfulness of 0.965–0.980 versus 0.578–0.639 for matched-random controls. Independent SST-2 task-circuit discovery across three models further confirms that 84.6–100% of persona-derived K heads belong to independently discovered task circuits, compared with only 8.8–17.6% of R heads.
  • End-to-end worked example. We illustrate the pipeline using Qwen-7B on SST-2:

    • Behavioral transition: For "that loves its characters and communicates something rather beautiful about human nature" (gold: Positive), the model predicts Positive under a joyful persona (margin +0.975) but Negative under sad persona (−0.872), with the task input unchanged.
    • Circuit discovery: Original and Reverse patching identify heads contributing to correct-to-wrong and wrong-to-correct margin changes, respectively.
    • K/RK/R decomposition: Original and Reverse discovery identify 13 and 10 shared K heads, respectively, with 10 heads in their intersection (KKKK).
    • Functional validation: Faithfulness and knockout experiments establish K's stronger task-functional contribution relative to RR.
    • Intervention: On Qwen-7B SST-2/Emotion, the circuit-guided intervention improves persona-sensitive accuracy by +9.97 pp (355 repaired, 20 harmed), preserving 95.83% of originally all-correct sample states.

    The example illustrates a verified persona-conditioned transition; circuit sizes and intervention results are aggregated over the corresponding task or task–axis setting, not measured on this individual sentence.

    We will incorporate this walkthrough into the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer pEZ6 (Part 2)

W2. Independent no-persona task-circuit validation

Thank you for this suggestion. We extend independent no-persona SST-2 task-circuit discovery from Qwen-7B to Llama-3B and Llama-8B, covering all three evaluated models. We retain SST-2 because its sentiment counterfactuals allow controlled label changes while preserving sentence structure and tokenized length; constructing equivalent counterfactuals for ARC is more challenging.

Across all three models, 84.6–100% of shared K heads are contained within independently discovered task circuits, compared with 8.8–17.6% of R heads. The task circuits also retain nearly full-model accuracy when isolated (faithfulness 0.994–1.010), substantially outperforming random controls (0.376–0.527), and exhibit knockout effects.

Together with our existing no-persona faithfulness and knockout experiments across all 15 model–task pairs, these results provide converging evidence that persona-derived shared heads participate in ordinary task computation rather than exclusively persona-specific mechanisms.

ModelTask headsK Original (%)K Reverse (%)KK (%)R Original (%)R Reverse (%)Faith. (Circuit/Random) ↑KO (Circuit/Random) ↓
Qwen-7B32100%100%100%10.0%17.6%1.010 / 0.5250.525 / 0.996
Llama-3B2191.7%92.3%90.9%12.5%9.0%0.994 / 0.5270.549 / 1.005
Llama-8B4287.5%85.7%84.6%8.8%17.0%1.000 / 0.3760.855 / 1.000

All K/R percentage columns report containment: the percentage of persona-derived heads contained in the independently discovered task circuit. R values are averaged across four persona axes.

We will update Section 6 and the appendix to include this three-model validation.

W3. ETHICS labels and the interpretation of correctness

Thank you for raising this important distinction. We agree that persona-induced changes in ETHICS may reflect legitimate differences in moral judgment rather than objective errors. Throughout our paper, "correct" means agreement with the benchmark reference label, not necessarily an objectively unique truth.

ETHICS Commonsense Morality uses consensus-cleaned annotations, retaining examples with at least four of five agreeing annotators (Hendrycks et al., 2021). Other tasks also rely on human judgments: SST-2 derives from the Stanford Sentiment Treebank, whose phrases were independently rated by three annotators (Socher et al., 2013). Our Safety task draws on human-authored Dolly 15K instructions and curated Do-Not-Answer prompts (Conover et al., 2023; Wang et al., 2024). In contrast, ARC consists of multiple-choice science questions with keyed answers (Clark et al., 2018).

Our mechanistic analysis holds the input and reference label fixed while varying only the persona, investigating the internal mechanisms underlying prediction changes. Our intervention therefore aims to restore benchmark-label agreement, not enforce an objectively correct moral viewpoint. We will explicitly define correctness as benchmark-label agreement in Section 3 and clarify in the Limitations section that persona-induced differences in subjective judgments, particularly ETHICS, should not necessarily be interpreted as errors.

Importantly, our findings do not depend on ETHICS. Excluding it, Original–Reverse recurrence remains substantially stronger for K than R (mean Jaccard 0.672 vs. 0.101), and intervention improvements remain McNemar-significant in 19/21 reliable settings.

These results indicate that the identified mechanisms and intervention effects extend beyond subjective moral judgments.

References

  • Hendrycks, D., Burns, C., Basart, S., Critch, A., Li, J., Song, D., & Steinhardt, J. (2021). Aligning AI With Shared Human Values. International Conference on Learning Representations (ICLR).
  • Socher, R., Perelygin, A., Wu, J., Chuang, J., Manning, C. D., Ng, A., & Potts, C. (2013). Recursive Deep Models for Semantic Compositionality Over a Sentiment Treebank. Proceedings of the 2013 Conference on Empirical Methods in Natural Language Processing (EMNLP), 1631–1642.
  • Conover, M., Hayes, M., Mathur, A., Xie, J., Wan, J., Shah, S., Ghodsi, A., Wendell, P., Zaharia, M., & Xin, R. (2023). Free Dolly: Introducing the World's First Truly Open Instruction-Tuned LLM. Databricks Blog.
  • Wang, Y., Li, H., Han, X., Nakov, P., & Baldwin, T. (2024). Do-Not-Answer: Evaluating Safeguards in LLMs. Findings of the Association for Computational Linguistics: EACL 2024, 896–911.
  • Clark, P., Cowhey, I., Etzioni, O., Khot, T., Sabharwal, A., Schoenick, C., & Tafjord, O. (2018). Think You Have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge. arXiv preprint arXiv:1803.05457.

Dear Reviewer,

Thank you again for your thoughtful review. As the rebuttal period is approaching its deadline, we would like to kindly follow up on our responses and additional experimental results.

We hope these clarifications and new analyses have addressed your concerns. We would greatly appreciate any further feedback and would be happy to address any remaining questions before the discussion closes.

Thank you for your time and consideration.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Consolidated Author Response for the AC and SAC (Part 1)

We thank the Area Chair, Senior Area Chair, and reviewers for their careful evaluation. This comment consolidates the paper's contribution and the new evidence provided during rebuttal; Part 2 summarizes our responses to the remaining concerns.

1. Central contribution

We investigate why changing only a persona prompt can change a model's answer to the same task item. Across three instruction-tuned models, five tasks, and four persona axes, complementary Original (damage) and Reverse (repair) circuit discovery identifies cross-axis shared attention heads (KK) and axis-dependent remainders (RR).

Our central finding is that persona-conditioned prediction changes can be mediated through shared task-functional mechanisms, rather than exclusively separate persona-specific circuitry. K/RK/R is a structural decomposition, not an assumed KK = task, RR = persona separation. ARC shows relatively greater task–persona segregation, while ETHICS, Safety, and SST-2 show stronger coupling within KK.

2. Additional evidence from rebuttal

In response to the reviewers, we conducted the following additional experiments and analyses:

Reviewer concernAdditional evidenceMain finding
Ordinary task functionality (pEZ6, 9Rbp)Independent no-persona SST-2 task-circuit discovery extended to all three models84.6–100% of KK heads are contained in independently discovered task circuits, compared with 8.8–17.6% of RR heads; faithfulness and knockout validate the task circuits' functionality.
Causal role of K/RK/R (9Rbp, jzBn)Bidirectional cross-patching and activation-direction nullificationKK recovers 95.5% of the correct–wrong margin gap, versus 16.8–19.5% for RR; the compact KKKK intersection retains 91.6%. Nullification independently supports causal involvement.
Upstream mechanisms and readout (jzBn)Persona-reader partitioning, upstream ablation, and selective contribution removalKK retains most of its causal effect without identified direct persona readers; the tested readers show no consistent necessity as an upstream pathway.
Intervention generalization (9Rbp, jzBn)Held-out evaluation with frozen coefficients and independently tuned random controls+8.48 pp persona-sensitive accuracy improvement on held-out examples, versus +4.36 pp for the strongest random control, preserving 98.92% of originally all-correct sample states.

The held-out evaluation covers 24 ETHICS/Safety settings, with coefficients selected exclusively on development data. The original development improvement was +12.00 pp across 33 reliable settings, significant in 29/33. Independently tuned and held-out random controls currently use one replicate.

These additional results strengthen the evidence for shared task-functional mechanisms and selective intervention within the evaluated settings, without establishing a unique computational pathway.

AuthorsOpenReview ↗

Consolidated Author Response for the AC and SAC (Part 2)

3. Responses to the Remaining Reviewer Concerns

We appreciate the reviewers' suggestions on presentation, interpretation, and scope. We have addressed them in our detailed rebuttal responses and will incorporate the corresponding clarifications into the revised manuscript.

  • Clarity, notation, and scientific justification (pEZ6, 9Rbp). While the original manuscript presents the relevant definitions and supporting analyses, we recognize that several explanations could have been clearer and more explicit. In our rebuttal, we therefore provided additional mathematical derivations, including the interpretation of Reverse patching scores: a negative score indicates an increase in the gold-answer margin, but does not necessarily imply a corrected prediction. We will incorporate these derivations, introduce clearer notation, add a dedicated preliminaries subsection and an illustrative walkthrough, and explicitly summarize RQ1–RQ5, with closer connections between claims and supporting evidence.

  • ETHICS and the interpretation of correctness (pEZ6). We agree that persona-induced differences in ethical judgments should not necessarily be regarded as errors, since reasonable perspectives may differ. Here, "correctness" refers to agreement with benchmark reference labels, rather than an objectively unique moral judgment. ETHICS uses consensus-cleaned annotations, but this distinction remains important. Importantly, our central findings are not dependent on ETHICS: after excluding it, Original-Reverse recurrence remains substantially stronger for KK than RR (Jaccard 0.672 vs. 0.101), and intervention improvements remain significant in 19/21 reliable settings.

  • Cross-task functionality and model capacity (9Rbp, jzBn). We appreciate the suggestions on task specificity and scaling. Our additional cross-task analysis reveals partial overlap, with only 1–2 heads per model shared across all five tasks, suggesting that the identified circuits are not universally identical. Although direct cross-task transfer remains untested, our results suggest task-dependent organization. We also find that Llama-8B exhibits lower behavioral instability than Llama-3B (2.95% vs. 4.71%), although representational analyses do not show uniformly greater task–persona segregation.

  • Open-ended generation and generalization (jzBn). We agree that complex reasoning and open-ended generation are valuable future directions. Our focus on restricted-output tasks is a deliberate methodological choice that enables controlled persona counterfactuals, consistent evaluation, and aligned causal interventions. Although these experiments do not establish generalization to open-ended or multi-token generation, they provide a controlled setting for mechanistic analysis. Similarly, our held-out evaluations demonstrate improvements on unseen examples within the evaluated settings, while transfer to unseen persona descriptions, new tasks, and substantially larger models remains an important direction for future work.

  • Relationship to suggested prior work (9Rbp). We thank the reviewer for highlighting these relevant studies. Li (2026) investigates task-dependent persona effects on responses; Gozzi and Fallucchi (2026) examine emotional framing and model performance; Sun et al. (2025) explore personality modulation through weight-space model merging; and concurrent work by Mohapatra and Bassett (2026) investigates context-dependent component roles across graph tasks. While these studies provide complementary perspectives, our work specifically investigates persona-induced correct–wrong transitions with the task held fixed, using bidirectional causal circuit discovery, cross-axis decomposition, independent task-circuit validation, and selective intervention.

4. Planned Manuscript Revisions

We appreciate the reviewers' recommendations to improve the presentation and scientific discussion. In the revised manuscript, we will incorporate the additional experimental results reported in our rebuttal, improve notation and mathematical explanations, provide an illustrative end-to-end example, explicitly summarize the answers to RQ1–RQ5, expand the related work, and clarify the scope and limitations of our findings.

Overall, we believe that the additional experiments and clarifications substantially strengthen the evidence supporting our central contribution: persona-conditioned behavioral changes can involve shared task-functional mechanisms rather than exclusively separate persona-specific circuitry. Although generalization and precise computational roles remain open questions, we hope our rebuttal conveys the significance of the findings and our careful consideration of the reviewers' concerns.

We sincerely thank the Area Chair, Senior Area Chair, and reviewers for their time, thoughtful feedback, and consideration of our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.