acceptodds
Under review as a conference paper at ICLR 2027

Shared Activation Directions Underlying LLMs’ Tendency toward Social Responsibility

Abstract

Large language model (LLM) responses express value orientations across diverse social contexts. However, endorsing socially responsible choices does not necessarily reflect how frequently people make those choices in practice. We investigate this distinction for socially responsible behavior (SRB), encompassing social contribution and harm reduction, across 39 behavioral items and four LLMs. In persona-conditioned sentence completions without explicit response options, all four models exhibit SRB rates above human references on average, even after behaviorally relevant information is added. We estimate model- and layer-specific shared directions from pre-response activations associated with naturally generated SRB and non-SRB responses. Individual-layer interventions modulate mean SRB rates across items in both directions, with model-dependent response ranges and saturation. Multi-layer coefficients selected using known human reference values reduce the mean absolute gap on held-out respondents for the same items under an output-validity criterion. Excluding the target item from direction estimation while retaining selected coefficients preserves the sign of behavioral changes in eligible nonzero interventions and most improvements over baseline. These findings suggest that SRB over-selection across distinct domains includes a shared, adjustable response tendency, connecting value-related responses to their behavioral prevalence in human data.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Large language models, Social responsibility, Activation steering, Shared representations, Human behavioral alignment, Persona-conditioned simulation

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Di8jOpenReview ↗
Rating
2
Confidence
3(+1)
Informativeness
2
Presentation
3
Soundness
1(−1)
Read review
Paper summary

The paper studies socially responsible behavior (SRB) in persona-conditioned sentence completions across four LLMs and 39 items spanning six domains. It compares response rates with survey-based and external aggregate human references, with 32 behavioral items eligible for reference comparisons before model-specific validity exclusions. For each model and layer, a shared direction is estimated by contrasting pooled pre-response activations associated with generated SRB and non-SRB responses. Individual-layer interventions generally change mean SRB rates in the expected direction. Cell-specific multi-layer coefficients selected using known human reference rates are then evaluated on different respondents answering the same items. Mean absolute reference gaps decrease from 34.42 to 4.27 percentage points for Qwen 3.5 9B, 41.44 to 6.24 for Qwen 3.5 27B, 46.41 to 8.70 for Llama 3.1 8B, and 52.59 to 27.61 for OLMo 3 7B on matched valid cells. Target-item exclusion, validity-threshold analyses, an SFT comparison, and country-specific European extensions provide additional validation. The proposed contribution is shared activation-based control of socially responsible response tendencies and calibration of their aggregate prevalence.

Critical strengths
  1. The study provides substantial evidence of controllable response tendencies across models and items. Four models are evaluated with the same task framework, individual-layer sweeps test both intervention signs, and the results expose saturation and heterogeneous controllability rather than presenting only successful examples. Excluding the target item from direction estimation preserves the sign of every eligible nonzero full-direction effect, while retaining the original coefficients. This supports some shared control across items and provides more evidence than merely demonstrating a single successful steering example (Figure 1; Table 3; Appendix F.4).

  2. The calibration evaluation separates coefficient selection from same-item respondent testing and tracks output validity. On matched valid cells, the selected interventions substantially reduce mean reference gaps, with coefficients fixed before test evaluation. The paper reports invalid-cell counts, residual failures, validity-threshold sensitivity, and an SFT comparison on matched cells. It also explicitly limits the calibration claim to aggregate rates for known items and does not claim individual behavioral fidelity. These design and reporting choices make the bounded calibration result useful despite the broader interpretive concerns.

Critical weaknesses
  1. Several human references do not measure the same quantity as the model completions. Appendix A.2 acknowledges that completion prompts omit recall periods, frequency conditions, and donation thresholds present in the surveys; some OECD references cover nine countries while the personas are US-based; and fraud reporting uses a reference conditional on exposure that the completion does not explicitly specify. Most concretely, WHO donations per 1,000 population include repeat donations and are explicitly not a proportion of people donating, yet are used for numerical comparison with a response proportion. These are differences in the estimand, not ordinary sampling noise. They undermine interpretation of the reported gaps as excess SRB relative to comparable human behavior and weaken the meaning of calibrating outputs to those targets. The main claims need a clearly comparable subset or a revised task/reference construction.

  2. The specificity and added value of shared-direction control remain insufficiently established. The pooled class contrast gives different weights to items and personas in the SRB and non-SRB groups, and cells containing only one class still contribute. Appendix E explicitly notes that the contrast can therefore reflect composition as well as behavioral labels. Target-item exclusion reduces direct dependence on that item's activations, but does not remove shared lexical, negation, or cross-item composition effects. Balanced within-item contrasts and generic-direction controls with the same intervention and tuning budget are needed to support an SRB-specific interpretation.

    The calibration results do not resolve this ambiguity: coefficients are selected separately for each item–persona cell against its known human reference rate. The held-out respondents answer the same items, and the primary endpoint is agreement with those aggregate rates. Section 5.1 appropriately states that this does not establish individual behavioral fidelity or preservation of relationships among behaviors. Nevertheless, the evaluation leaves open whether steering preserves useful respondent information, improves subgroup-conditional distributions, or merely changes the overall frequency of affirmative and negative completions. It could improve a marginal rate while worsening the assignment of responses to individuals; this possibility has not been evaluated.

    A reference-rate sampler can match known marginals in expectation without using persona information, illustrating the limited information certified by that endpoint alone. Such a comparator would not explain the activation mechanism, but a simple output-level calibration baseline and pre/post-steering individual and subgroup evaluations would clarify the intervention's added practical value. The reported SFT comparison is useful but does not answer these questions when judged primarily by the same marginal-gap metric. This is a limitation of validation depth and interpretation, while the demonstrated ability to control response rates remains a valid bounded finding.

Questions for authors
  1. Clarification and reanalysis relevant to the current assessment: comparable human references. Do the headline baseline gaps and calibration improvements persist when restricted to items with matched populations, behavioral definitions, recall periods, and denominators? In particular, please exclude the donations-per-population indicator from comparisons with individual response prevalence, and clarify how the general completion wording can be aligned with the original time-bounded survey questions. A measurement-matched analysis would materially affect my evaluation of the behavioral claim.

  2. Additional evidence relevant to the current assessment: direction specificity. What happens when directions are constructed from balanced within-item/persona contrasts, rather than pooled class means? Under the same layer bundle, coefficient-search budget, and validity criterion, how do the resulting directions compare with generic affirmation/negation directions or random-direction controls? These comparisons would help distinguish an SRB-related mechanism from broad response manipulation plus per-cell calibration.

  3. Additional evidence relevant to the current assessment: value beyond marginal calibration. On items with individual human labels, how do balanced accuracy or informedness and subgroup-conditional calibration change after steering? Please compare with a simple output-level calibration baseline given the same known reference rates, and report output validity alongside performance. These results would clarify whether steering preserves or improves respondent-specific information and conditional distributions while reducing aggregate gaps. I recognize that individual fidelity is outside the current claim; the requested evidence would materially affect my assessment of the method's practical contribution.

  4. Clarification of response coding. Was the rule-based parser independently checked against manual judgments on a stratified sample containing baseline, positive-steering, and negative-steering outputs? If so, please report disagreement and error rates by condition. Since the labels determine both the estimated directions and the main evaluation, intervention-dependent coding errors could materially change the conclusions.

8 comments

Thank you for your careful review and comments on controllability and evaluation design.

1. Simplified and restored measurement conditions. To address comparability directly, we restored survey conditions within sentences or in separate explanations for 22 items excluding WHO/OECD external aggregate references. The WHO indicator counts donations, including repeat donations, and thus differs from individual participation prevalence; we will clarify this distinction in the main text. This analysis tests the main findings independently of that indicator by specifying recall periods, frequencies, amounts, and other conditions simplified in §3.2 and Appendix A.3.

Fraud-reporting references and personas already used exposed respondents (FRAUD1=1) with valid reporting responses; restoration makes this sampling condition explicit in prompts. Actual protest-participation examples follow.

ConditionItem and additional explanation
OriginalI [BLANK] in protests about political or social issues.
Conditions restored within the sentenceI [BLANK] in protests about political or social issues during the past 12 months.
Conditions restored in a separate explanationI [BLANK] in protests about political or social issues. For this statement, consider the behavior during the past 12 months.

Directional similarity. Original and restored directions were re-estimated on the same 22 items and 500 development respondents. Layerwise signed cosine ranges were 0.451–0.526, 0.750–0.796, and 0.377–0.444 for within-sentence restoration, and 0.547–0.607, 0.791–0.869, and 0.657–0.699 for separate explanations, in Qwen/Llama/OLMo order. All were positively aligned, with higher similarity for separate explanations.

Baselines and intervention transfer. We applied the paper's fixed directions, layers, and cell-specific coefficients to 100 separate test respondents; re-estimated directions served only similarity analysis. Comparisons cover human-reference behavioral cells among these 22 items with existing coefficients and pre/post responses, retaining development references and combining persona types. Restored baseline mean signed gaps were +16.22/+33.76/+55.89 pp within sentences and +39.05/+63.04/+58.83 pp with separate explanations, in Qwen/Llama/OLMo order: mean overselection persisted.

The table shows mean absolute gaps before→after intervention (pp) for each prompt. Baseline comparisons require validity under both unmodified and modified baselines; pre/post comparisons require validity under both prompts before and after intervention. Validity means ≤25% unclassified responses. Matched items, personas, and respondents receive equal cell weights (n: common valid cells).

ConditionModelnOriginal prompt: before→after interventionModified prompt: before→after intervention
Conditions restored within the sentenceQwen 3.5 9B2941.44→5.5930.33→24.93
Conditions restored within the sentenceLlama 3.1 8B2558.43→9.3142.10→22.71
Conditions restored within the sentenceOLMo 3 7B2155.09→34.3753.73→42.53
Conditions restored in a separate explanationQwen 3.5 9B2941.44→5.5943.43→31.03
Conditions restored in a separate explanationLlama 3.1 8B3061.58→9.7762.94→23.86
Conditions restored in a separate explanationOLMo 3 7B2053.56→35.8055.98→43.35

The free-form completion format still differs from the original surveys. Nevertheless, after excluding external aggregate items and restoring measurement conditions, all three models retained both overselection and gap reductions without retuning under both restoration methods. Thus, comparisons that reduce the measurement differences you identified preserve the central findings of SRB overselection and shared-direction control.

2. Item–persona composition and balanced directions. To test the composition effects documented in Appendix E.1, we compared pooled estimation (§3.3) with equally weighted cell directions, complementing target-item exclusion (§5.2).

From the existing 500 respondents per cell, we selected item × persona cells with at least 5 samples of each class, averaging within-cell class-mean activation differences equally. Matched pooled uses pooled class-mean differences from exactly the same cells. Retaining unit layerwise norms and the paper's layer bundle, we selected direction/cell-specific α on the same 100 development respondents, minimizing human-reference absolute gaps among candidates with ≤25% unclassified responses (§3.3, Appendix F.2).

Eligible cells numbered 60/38/15 for Qwen/Llama/OLMo; balanced–matched pooled layerwise cosines were 0.491–0.539, 0.717–0.803, and 0.601–0.625. Evaluation covered the existing 48/53/36 human-reference cells, fixing settings and development references on test100 respondents disjoint from the 500 used for estimation (§5.1). The table reports equally weighted mean absolute gaps over cells valid in all three conditions.

ModelCommon valid/evaluated cellsNo intervention (pp)Matched pooled (pp)Balanced direction (pp)
Qwen 3.5 9B47/4834.664.564.97
Llama 3.1 8B50/5345.8511.7220.89
OLMo 3 7B32/3649.1034.0634.57

Valid-cell counts without intervention/matched pooled/balanced were 47/47/48 for Qwen, 53/52/50 for Llama, and 36/34/34 for OLMo. Despite changing pooled weighting induced by class composition, balanced directions reduced gaps on new respondents for all three models.

3. Coefficient-dependent control and comparison directions. The suggested random and generic positive/negative controls ask whether arbitrary activation shifts or general linguistic polarity can explain the effects. We first constructed two unit-norm random directions (seeds 101, 202) and applied the paper's coefficient-response curves (§4.2, Figure 1) and cell–layer Kendall's τ_b analysis relating α to SRB rates (Appendix F.4, Table 12).

This tests consistent SRB changes across coefficients, beyond fitting one coefficient to a human target, which arbitrary directions might also achieve. It is not a final calibration comparison under equal multi-layer tuning budgets: SRB directions used test400 and random directions dev100.

Using the same single-layer α sweeps, we computed τ_b from observations with ≤25% unclassified responses and averaged equally over defined cell–layer pairs. Parentheses give defined/total pair counts.

ModelPaper's SRB direction τ_bRandom seed 101 τ_bRandom seed 202 τ_b
Qwen 3.5 9B0.906 (171/180)-0.194 (147/180)-0.032 (149/180)
Llama 3.1 8B0.647 (136/180)0.305 (99/180)0.028 (104/180)
OLMo 3 7B0.610 (104/200)-0.073 (74/200)-0.070 (72/200)

We compared magnitude using ΔSRB=SRB(α_max)−SRB(α_min), from mean response rates over fixed cells valid throughout each range. Below are minima–maxima across layers for SRB directions and layers/seeds for random directions.

ModelCompared α rangePaper's SRB direction ΔSRB (pp)Random directions ΔSRB (pp)
Qwen 3.5 9B−12~12+32.6~+38.6−4.7~+3.5
Llama 3.1 8B−4~12+26.0~+35.2−1.0~+6.2
OLMo 3 7B−12~12+2.8~+8.9−1.5~+0.8

All three models showed higher mean τ_b and larger ΔSRB for the SRB direction than for the tested random directions. Samples and direction-specific valid cells differ, so this is not a matched-sample effect-size test. It nevertheless demonstrates systematic response changes along the SRB direction beyond fitting one selected coefficient.

Separating this control from generic positive/negative tendencies requires specifying the property—propositional agreement, grammatical negation, or affective polarity—and independently validated data for constructing its direction. Relabeling the same SRB data would not separate the properties. Our current data do not support this independent control, so we did not perform that comparison or claim complete independence from linguistic polarity.

The completed tests support functional relevance from complementary angles: target-item exclusion tests dependence on the target item's activations; balanced contrasts test class-composition effects across cells; random controls test whether arbitrary shifts produce similar control. Together, they support our central contribution: shared directions estimated across behaviors enable systematic modulation of SRB responses.

4. Individual changes accompanying distributional calibration. Section 5.1 targets aggregate calibration, explicitly distinguishing accurate individual reproduction; individual accuracy was not used for coefficient selection. Following your question, we evaluated ACC/BA in existing test400 outputs, keeping interventions fixed and using cases with individual human labels classifiable before and after intervention.

ACC is accuracy across all cases. BA equally weights performer/nonperformer accuracy within item–persona cells, then averages cells equally; class weighting and aggregation therefore differ.

ModelEvaluated cellsCommon valid casesACC before→after (%)BA before→after (%)
Qwen 3.5 9B3614,31557.42→72.6354.42→54.80
Qwen 3.5 27B4618,08146.74→72.1155.04→55.25
Llama 3.1 8B4115,41044.44→69.7452.83→53.76
OLMo 3 7B2810,35144.30→60.4250.75→51.84

ACC increased by 15.21–25.37 pp across the four models. BA gains of 0.21–1.09 pp were smaller and do not establish a large improvement in individual discrimination. Our optimization targets aggregate behavioral distributions; these results show that calibration also improves individual response agreement. We further examined item-level gap reductions and ACC increases on the same common cases.

ModelEvaluated items with individual labelsPearson r between gap reduction and ACC improvementSpearman ρ
Qwen 3.5 9B210.9030.730
Qwen 3.5 27B230.9430.840
Llama 3.1 8B220.9340.853
OLMo 3 7B170.9280.677

Both metrics use the same outputs, so this is not independent evidence of improved use of personal information. Nonetheless, gap reductions and ACC increases co-occurred in all four models: distributional calibration improved response agreement without optimizing individual accuracy.

5. A sampler with identical reference rates. Direction estimation/bidirectional control use no human references (§3.3, §4.2); this sampler supplements the calibration application (§5.1). Using only development human rate p for each item, it assigns test individuals SRB with probability p and non-SRB with 1−p. Persona attributes and test human labels are unused. We report mean ACC across 200 repetitions on identical common cases.

ModelNo-intervention ACC (%)Steering ACC (%)Sampler ACC (%)
Qwen 3.5 9B57.4272.6368.77
Qwen 3.5 27B46.7472.1168.17
Llama 3.1 8B44.4469.7468.02
OLMo 3 7B44.3060.4270.29

The sampler directly draws responses from known rates, matching marginals in expectation; class imbalance alone can yield high ACC. It outperformed OLMo, and the small BA changes also caution against interpreting this comparison as improved use of personal information. Without optimizing individual accuracy, all four models improved over no intervention, and both Qwen models and Llama exceeded the mean ACC of the sampler using identical reference rates. Thus, activation-based distributional control can also improve individual response agreement.

Reviewer_4S6COpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces an evaluation protocol for investigating the social values embedded in LLMs, with a particular focus on socially responsible behavior. And this paper also demonstrate that interventioning the intermediate layers of LLMs can push the value oriention in LLMs to be more "aligned" with humans' values.

Critical strengths

A new evaluation protocol for the embedded social values in LLMs.

Critical weaknesses

It is difficult to assess the strengths and weaknesses of a paper primarily focused on introducing an evaluation protocol. My main concern is the validity of its conclusions: given that social values operate at the pragmatic level, whereas LLMs are grounded in distributional semantics, to what extent do the reported findings reflect socially responsible behavior?

Questions for authors

to what extent do the reported findings reflect socially responsible behavior?

1 comment

Thank you for the opportunity to clarify our measurement and contribution. We define SRB as actions or behavioral choices that contribute to the community or society, or reduce avoidable harm beyond the actor (Introduction, §2.1). This connects voting, giving, environmental protection, and other behaviors through shared properties; we measure how often persona-conditioned responses indicate performing these behaviors.

I [BLANK] ... elicits first-person behavioral descriptions from the respondent's perspective, rather than abstract agreement with values or judgments of desirability (§3.2). We read generated phrases with their sentences and classify performance/nonperformance; uncertainty or mere intention is not treated as actual performance (Appendices B.1–B.2, Table 7). We directly measure these SRB descriptions and the activation directions that modulate them, without equating them with real-world behavior or intrinsic values. Human references are empirical comparators for this behavioral-response simulation.

Beyond prior reports of value orientations or prosocial biases in individual tasks, we estimate shared directions from pre-response activations associated with naturally generated SRB/non-SRB responses across behaviors, and validate their function through intervention (§2.2–§3.3). Directions require neither human rates nor instructions contrasting desirability. Bidirectional intervention (§4.2), target-item exclusion (§5.2), and additional balanced-direction comparisons establish control across behaviors. Human rates enter when applying this control to distributional calibration.

Under lexical changes, layerwise cosines with the original directions were 0.913–0.921, 0.990–0.992, and 0.833–0.846 for Qwen/Llama/OLMo. Restoring omitted conditions in two ways also reduced mean absolute gaps in all three models using fixed original directions and coefficients (hiuj, response 3; Di8j, response 1). The contribution goes beyond an evaluation protocol: we discover and validate shared activation-based control of SRB response tendencies that transfers across behaviors, wording, and conditions. This directly advances understanding and internal control of biases in text-based human simulation.

We will clarify the SRB definition, the scope of measuring textual behavioral descriptions, and the role of human references in the Introduction and §3.2. In §2.2 and the contribution statement, we will emphasize the distinction between documenting individual behavioral biases and estimating shared directions validated through intervention. These revisions and additional tests will make the discovery and controllability of SRB tendencies across behaviors more explicit. Thank you for your feedback; we would appreciate your considering these clarifications in your assessment.

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

The paper investigates whether four persona-conditioned LLMs’ tendencies to generate texts associated with socially responsible behavior can be up & down-regulated via activation steering. The authors compile a suite of test items along with human rates from the literature and find their LLMs to overrepresent socially responsible behavior (SRB) as compared to humans. Applying an activation steering method, they are able to bring the LLM distribution closer to the baseline.

Critical strengths
  • The paper and appendix include a great amount of detail and documentation. The technical reporting is good.
Critical weaknesses

• The paper does not motivate why (persona-conditioned) LLM SRB rates different from human baselines would be undesirable. I do not understand why too socially responsible LLMs are problematic at all or why persona-conditioning is needed here. It seems the latent thesis of the paper is that the human SRB rates are the only correct option for LLMs and deviations should therefore be minimized, but why? Is it some sort of incorrect “role-playing” if a persona-conditioned LLM does not show similar SRB behavior as the real person the persona was derived from?

• The paper compares human behavior to LLM behavior and does so by comparing two rates. I fear that the processes producing those rates are too different to warrant a meaningful comparison, or, at the very least, the paper does not provide evidence to convince me of the opposite. Two examples: The human rates are taken from six different sources which are unlikely to all have the same setup and assumptions. Therefore, it is conceivable that the human rates for specific items might differ under the various approaches, calling into question the validity of the obtained test conglomerate. Thus, I am not sufficiently convinced that exactly those rates should be the anchors for LLM behavior. Second, the test procedures are not standardized across LLMs and humans, so factors like different coding schemes, inclusion/exclusion criteria, item wording etc. all might have disparate impacts on the rates.

• For a paper that sets out to study distributional differences, some design choices are severely weakening the methodology. E.g., it is odd to restrict LLM sampling to greedy only. It is quite possible that LLMs do not always choose the same answer, but rather show a distribution over different alternatives. Also, the prompting scheme is quite restrictive and does not consider multiple templates or paraphrases to rule out known LLM quirks in such setups.

• The steering method appears to modulate the desired metric. But does it modulate only that metric? The paper does not show that the steering does not introduce any side effects or degrade general performance.

Questions for authors

Clarification questions:

  • Why is the alpha range capped at -4 for the llama model but not for the others?
7 comments

Thank you for your careful review and assessment of our technical reporting.

1. Why higher SRB rates than humans are a concern. Our objective is to reproduce participation and nonparticipation distributions in human behavioral simulation. For example, simulating almost everyone as donating or volunteering can underestimate nonparticipation and the circumstances that constrain participation. Human references are empirical criteria for this simulation task, not socially desirable levels or correct answers for every LLM (Appendix A, Ethics Statement).

Supporting social responsibility and practicing it differ: participation depends on time, resources, and others' cooperation (Introduction, §2.1). We therefore constructed common/specialized personas from actual respondents to test whether SRB overselection persists despite providing background and behavioral context (§3.1, Appendix A.3). Controlling both participation and nonparticipation directly serves the purpose of human simulation.

Our core contribution is estimating a shared direction from naturally generated responses and activations without human reference rates, and modulating SRB responses across behavioral items. Calibration to human rates applies this controllability.

2. Sources and comparability of human references. As described in §3.1 and Appendices A.1–A.3, items with individual data use personas constructed from that dataset's respondents and compare against their recorded behaviors. For these items, human–model comparisons are matched within sources; multiple sources cover different behavioral domains.

Free-form completion is designed to reduce option-order and format constraints while eliciting behavioral descriptions from the persona's perspective. Human binarization (Appendix A.2) and contextual model-response coding (§3.2, Appendices B.1–B.2, Table 7) correspond to performing/nonperforming the same behavior; unclassified responses are not treated as non-SRB.

To test whether simplified recall periods, frequencies, and amounts explain the findings, we restored conditions within sentences or in separate explanations for 22 items excluding WHO/OECD external aggregate references. Mean overselection persisted in Qwen 9B, Llama, and OLMo under both variants. With the original directions and coefficients fixed, mean absolute gaps fell from 30.33→24.93, 42.10→22.71, and 53.73→42.53 pp within sentences, and 43.43→31.03, 62.94→23.86, and 55.98→43.35 pp with separate explanations, respectively (Di8j, response 1). Overselection and intervention effects without retuning both survive restoration of measurement conditions: the main findings do not depend solely on their omission.

3. Prompt robustness. The paper tests intervention transfer when excluding the target item from direction estimation (§5.2, Table 3). Following your question, we also changed wording and syntax across all evaluation items. For Qwen 3.5 9B, Llama 3.1 8B, and OLMo 3 7B, we re-estimated directions using the same 500 development respondents and common/specialized conditions, preserving the behavior and blank-completion format.

ConditionExample item
OriginalI [BLANK] in protests about political or social issues.
Lexical changeI [BLANK] in demonstrations about political or social issues.
Syntactic changeIn protests concerning political or social issues, I [BLANK].

Directional similarity. Layerwise signed cosine ranges against the original direction were 0.913–0.921, 0.990–0.992, and 0.833–0.846 for lexical changes, and 0.734–0.786, 0.816–0.841, and 0.409–0.447 for syntactic changes, in Qwen/Llama/OLMo order.

Intervention transfer. We applied the paper's fixed directions, layers, and cell-specific coefficients to 100 separate test respondents; re-estimated directions were used only for similarity. Below are mean absolute gaps before→after intervention (pp), retaining the paper's human references and matching items, personas, and respondents. n counts common cells with unclassified-response rates ≤25% before and after intervention under both prompts.

ConditionModelnOriginal prompt: before→after interventionModified prompt: before→after intervention
Lexical changeQwen 3.5 9B4633.80→5.3931.03→12.78
Lexical changeLlama 3.1 8B4044.81→9.4843.77→15.75
Lexical changeOLMo 3 7B2848.13→23.3148.82→28.43
Syntactic changeQwen 3.5 9B3329.87→5.6933.07→25.69
Syntactic changeLlama 3.1 8B3944.50→9.9348.08→26.25
Syntactic changeOLMo 3 7B1836.83→25.9639.74→37.01

Lexical changes retained high directional similarity in all three models; syntactic changes generally weakened similarity and calibration effects. OLMo had the lowest syntactic-change cosine and the smallest gap reduction (39.74→37.01 pp). This shows model-dependent stability and transfer under rephrasing, without establishing the cause of those differences. Crucially, all three models reduced mean absolute gaps under both lexical and syntactic changes without retuning the original directions, layers, or coefficients.

4. Decoding and repeated-response variability. Appendix C.1 specifies temperature 0.7, top-p 0.9 for direction-estimation labels, and greedy decoding for coefficient selection, testing, and single-layer sweeps to focus on activation-intervention effects.

We additionally generated seven repetitions without and with fixed intervention for 100 test respondents per cell under the same stochastic settings. Each repetition averages cell-level SRB/(SRB+non-SRB) and absolute human-reference gaps with equal cell weights; we report mean ± sample SD across repetitions. A fixed common set had unclassified-response rates ≤25% in all repetitions and conditions.

ModelCommon valid/total cellsSRB rate before→after (%)Mean absolute gap before→after (pp)
Qwen 3.5 9B43/4861.35 ± 0.76 → 51.49 ± 0.2527.77 ± 0.88 → 13.65 ± 0.28
Llama 3.1 8B35/5387.39 ± 0.24 → 58.20 ± 0.6042.11 ± 0.43 → 13.85 ± 0.43
OLMo 3 7B21/3695.97 ± 0.50 → 89.67 ± 0.3748.71 ± 0.18 → 41.96 ± 0.35

Across all outputs, including excluded cells, mean unclassified rates rose from 4.45→5.31%, 4.94→15.80%, and 5.74→21.02% for Qwen/Llama/OLMo. We therefore distinguish valid-cell improvements from output-validity costs. SDs of mean SRB rates (0.24–0.76 pp) and absolute gaps (0.18–0.88 pp) were smaller than the 6.75–28.26 pp gap changes. Every repetition reduced both metrics on common valid cells in all three models, showing that fixed interventions transfer beyond greedy evaluation to stochastic generation.

5. General task performance. We use existing activation steering to discover and functionally validate a shared SRB direction. Beyond output validity (§3.2, Table 2(b), Appendix G.4), we evaluated general tasks using combinations nearest the median and 75th percentile, plus the maximum, of the L2 norm distribution of nonzero coefficient combinations, weighted by actual cell usage. Directions have unit norm per layer; α follows the listed layer order (median / 75th percentile / maximum norm).

  • Qwen 3.5 9B, layers (21, 22, 23): (−12, −4, −4) / (−10, −8, −8) / (12, 20, 12)
  • Llama 3.1 8B, layers (15, 17, 18): (0.5, −8, 0) / (−8, −8, −4) / (−19.75, 15.875, −4)
  • OLMo 3 7B, layers (27, 28, 29, 30, 31): (0, 0, 16, −8, 8) / (0, 12, 4, −12, 20) / (−24, 12, 12, −12, 24)

Layers, signs, and coefficient allocation also differ, so differences cannot be attributed to magnitude alone. We evaluated 140 MMLU-Pro items (fixed sampling by domain) and 285 MMLU-Redux items under identical items, prompts, and generation budgets, counting extraction failures as incorrect.

ModelBenchmarkNo interventionMedian representative75th-percentile representativeMaximum norm
Qwen 3.5 9BMMLU-Pro75.0072.1474.2969.29
Qwen 3.5 9BMMLU-Redux78.9578.9578.2578.25
Llama 3.1 8BMMLU-Pro42.8630.002.145.00
Llama 3.1 8BMMLU-Redux68.4268.7747.0243.16
OLMo 3 7BMMLU-Pro45.0047.1452.1450.00
OLMo 3 7BMMLU-Redux63.5164.5663.5164.21

Intervention effects varied by model, coefficient combination, and task. Qwen's median/75th-percentile combinations changed performance little, and the tested OLMo combinations showed no decline. Llama's median combination reduced Pro accuracy by 12.86 pp; at the 75th percentile, 139/140 items reached the length limit and 116 failed extraction. SRB controllability and preservation of general capabilities therefore require separate evaluation. These results show that behavioral calibration need not incur large general-performance losses, while identifying model- and setting-specific considerations for using the discovered shared direction.

6. Llama's α lower bound of −4. This is the displayed lower bound for the single-layer sweep in Figure 1, not the search bound for multi-layer calibration. At tested single-layer α<−4, every evaluated cell exceeded 25% unclassified responses, so the displayed range follows the criterion of common valid cells throughout the sweep (Appendices F.3, F.5). This distinction also explains coefficients below −4 in the multi-layer combinations above. Model-specific displayed ranges apply the same output-validity criterion. We will clarify this in §4.2 and Appendix F.3.

Thank you again for your careful review and specific comments. If the explanations and additional validation above address your concerns, we would be grateful if you would take them into account in your assessment.

Comments on the paper

1 comment

We sincerely thank all three reviewers for their careful reviews and constructive suggestions. We have uploaded the revised PDF, with additions and revisions highlighted in blue. Below we summarize the main changes and where the analyses discussed in our replies appear.

1. Motivation and contribution (hiuj, Di8j, 4S6C). We clarified that our contribution is to discover a shared internal direction across socially responsible behaviors and validate its behavioral relevance, rather than to develop a new activation-steering technique or an individual behavior predictor. We expanded the rationale for the task using prior work on value evaluation and human behavioral simulation, distinguished direction estimation without human reference rates from subsequent calibration using those rates, and made the implications for studying value-related internal tendencies and behavioral choices more concrete.

2. Prompt wording and measurement conditions (hiuj, Di8j; Appendix G). We added lexical and syntactic prompt changes and restored survey conditions, applying the original directions and intervention settings without retuning. We distinguished cosine similarity from functional transfer and reported alpha sweeps under changed prompts to examine coefficient-dependent responses alongside improvements at fixed coefficients. We also clarified which conditions, such as prior scam exposure, were already reflected in persona construction and which measurement conditions were additionally restored.

3. Functional relevance and linguistic polarity (Di8j, hiuj; Appendices F–G). Random-direction controls compare the magnitude of SRB changes and Kendall τ-b across intervention strengths, rather than final calibration performance alone. We also applied the original directions, layers, and coefficients to seven candidate items whose semantic polarity was reversed while preserving sentence structure as closely as possible. We report mean SRB changes together with evaluated and retained cell counts and uncodable-response rates; OLMo was excluded from this analysis's primary summary because of poor output validity. This is presented as a behavioral transfer evaluation under modified prompts, distinct from a control that estimates new directions.

4. Individual agreement and age-group distributions (Di8j; Appendix H). We clarified the supplementary role of individual-level evaluation in examining changes accompanying distributional calibration. We distinguished ACC from BA and reported the relationship between item-level SRB-gap reductions and improvements in individual agreement. For age groups, we replaced the main error-summary table with actual human-reference, baseline, and intervention SRB rates, making the movement toward human references visible for each age group and model. We distinguish improvements in within-group rate levels from preservation of between-group differences, retaining the relevant error analysis and output-validity results.

5. Decoding variability and general performance (hiuj, Di8j). We repeated stochastic generation seven times per prompt and report the mean and standard deviation of estimates computed separately from each repetition's 100 respondents. For general performance, we specify the actual intervention coefficient combinations and describe costs that vary across models and settings. These supplementary evaluations distinguish the direction's functional relevance from the conditions for applying it to other tasks.

Throughout the revision, we distinguish the original findings, the questions addressed by additional evaluations, and the scope supported by the results. In particular, we separate geometric similarity from functional transfer, distributional calibration from individual discrimination, and improvements in group-specific rates from preservation of between-group differences.

We appreciate the opportunity to make the study's purpose and findings clearer, and would be glad to answer further questions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.