PaperlayerView exact revision on arXiv

Source-linked AI summary

From Passive to Persuasive: Localized Activation Injection for Empathy and Negotiation

Niranjan Chebrolu, Kokil Jaidka, Gerard Christopher Yeo

arXiv:2511.12832v3cs.CLcs.AI

TL;DR

Complex interpersonal behaviors remain difficult to steer because existing methods use global interventions and largely target surface outputs. STAR localizes causal layer–token sites with attribution patching, then injects contrastive vectors there; localized injection consistently outperforms global steering and instruction priming across emotional dialogue and negotiation, with larger multi-turn effects.

  • Problem

    Existing steering methods lack causal localization, distinguish insufficiently between steering goals, focus on single-turn settings, and target surface outputs rather than contextually grounded interpersonal traits.

    Source: §1 · p. 1

  • Method

    STAR uses attribution patching to identify behaviorally causal layer–token positions and injects contrastive activation vectors at those localized sites during inference.

    Source: §1 · p. 1, §3.2 · p. 2

  • Results

    Localized injection consistently outperforms global steering and instruction priming across emotional dialogue and negotiation in single- and multi-turn settings, with affiliative politeness gains about 20× larger across turns.

    Source: §1 · p. 1, §5 · p. 8

  • Takeaways & Limitations

    Complex interpersonal behaviors can be represented as localized, approximately linear directions that are identifiable, transferable, and precisely injectable.

    Source: §5 · p. 8

    Results use a single model family, human-evaluation samples are too small for corrected significance, and steering vectors require manually authored seed pairs.

    Source: §5 · p. 9

Abstract

from arXiv · show

Complex social behaviors, such as empathy and strategic politeness, are widely assumed to resist the directional decomposition that makes activation steering effective for coarse attributes like sentiment or toxicity. We present STAR: Steering via Attribution and Representation, which tests this assumption by using attribution patching to identify the layer--token positions where each behavioral trait causally originates, then injecting contrastive activation vectors at precisely those locations. Evaluated on emotional dialogue and negotiation in both single- and multi-turn settings, localized injection consistently outperforms global steering and instruction priming; human evaluation confirms that gains reflect genuine improvements in perceived quality rather than lexical surface change. Our results suggest that complex interpersonal behaviors are encoded as localized, approximately linear directions in LLM activation space, and that behavioral alignment is fundamentally a localization problem.

1 Introduction

STAR addresses failures of global, surface-oriented behavioral steering by localizing where complex traits emerge and injecting contrastive vectors there. Across emotional dialogue and negotiation, localized injection outperforms global steering and instruction priming, with effects that compound across turns.

Motivation

  • Existing steering methods apply vectors globally, distinguish neither corrective alignment from expressive enhancement nor multi-turn persistence, and target surface outputs rather than contextual interpersonal traits.These gaps motivate causal localization for socially sensitive behaviors.Source: §1 · p. 1

Contribution

  • STAR uses attribution patching to identify causal layer–token locations, then injects contrastive activation vectors precisely at those positions.Localization directly parameterizes injection rather than serving as post-hoc analysis.Source: §1 · p. 1, Contributions · p. 1

Results

  • Localized injection consistently outperforms global steering and instruction priming across emotional dialogue and negotiation in both single- and multi-turn settings.Affiliative politeness gains are approximately 20× larger across turns (∆≈+0.46 versus ∆≈+0.021).Source: §1 · p. 1, Contributions · p. 2

Evaluation

  • The evaluation spans affective attunement in BOLT SMS emotional dialogue and strategic influence in Craigslist Bargain negotiation.The two domains represent complementary poles of social behavior.Source: §1 · p. 1

Implication

  • The findings suggest that complex interpersonal stances are recoverable as precise, localized directions in residual stream space, making localization central to behavioral alignment.Human evaluation indicates that automatic gains reflect perceived quality rather than lexical inflation.Source: §1 · p. 1, Contributions · p. 2

2 Related Work

Prior activation-steering work typically applies vectors globally and focuses on surface attributes, while STAR formulates localized, goal-sensitive steering for complex social behaviors. Its problem formulation restricts intervention to behaviorally relevant final-token states.

3.1 Framework Overview

  • STAR distinguishes corrective alignment from expressive enhancement and treats localization as necessary for contextual grounding and interpersonal calibration.The framework targets empathy, disclosure, counter-offering, and strategic politeness rather than only surface-level outputs.Source: §3.1 · p. 2, §3.2 · p. 2

3.2 Problem Formulation

  • During inference, STAR modifies only the final k hidden states using a behavior vector scaled by α, rather than steering every token position.The intervention is defined over the final positions of the sequence.Source: §3.2 · p. 2

2 Related Work

  • Existing activation-based methods manipulate internal representations, but commonly inject vectors uniformly, evaluate only single-turn behavior, and leave complex social traits out of scope.The related-work framing contrasts causal internal steering with methods that optimize superficial reward cues.Source: §3.2 · p. 2, §2 · p. 2

3 Method

STAR identifies causally influential layer–token locations with attribution patching, constructs behavioral contrastive vectors, and injects them at those locations. The method is evaluated across two datasets, two interaction lengths, and baselines that isolate localization from generic activation steering.

3.3 Stage 1: Attribution-Based Layer Selection

  • Attribution patching compares aligned and misaligned completions and replaces hidden states across layers and token positions to estimate each component’s causal contribution.The resulting attribution map yields the target layer and number of final positions for intervention.Source: §3.3 · pp. 2–3

3.4 Stage 2: Behavioral Exemplar Construction and Injection

  • STAR constructs the steering vector from paired target-aligned and neutral behavioral exemplars at the causally selected layer.The vector captures the dominant behavioral shift before localized injection during generation.Source: §3.4 · p. 3
  • The steering strength is selected by a validation Pareto sweep, yielding α=2.0 for emotional dialogue and α=2.5 for negotiation.Efficiency collapses after α=2.0→2.5 in emotion and one step later in negotiation.Source: §3.4 · p. 3

3.5 Datasets

  • The evaluation covers empathy and disclosure in 11,002 single-turn BOLT SMS dialogues plus 1,102 filtered multi-turn dialogues, and negotiation in 5,155 single-turn Craigslist Bargain dialogues plus 515 multi-turn dialogues.Multi-turn subsets require exchanges of at least five turns, with an additional seller-concession criterion for negotiation.Source: §3.5 · p. 3

3.6 Experimental Conditions

  • Single-turn tests measure immediate effects, while multi-turn configurations test persistence, induction, and reversal through UU, US, SU, and SS schedules.Llama-3.1-8B receives the steering vector; Mistral-7B-Instruct serves as conversational partner in specified settings.Source: §3.6 · p. 4

3.7 Baselines

  • Baselines include unsteered generation, zero-shot instruction priming, and global activation steering using the same vector without attribution-guided localization.The global condition isolates the contribution of localized injection.Source: §3.7 · p. 4

4 Results

STAR identifies task-specific intervention loci through attribution patching, then evaluates localized steering across emotional dialogue and negotiation. Localized injection improves automatic and human-evaluated quality, while multi-turn effects are larger but qualitative failures and nonsignificant multi-turn comparisons bound the evidence.

4.1 Identifying the Intervention Locus

  • Attribution patching identified selective, task-specific loci: affective tasks concentrated in the final 15 prompt tokens at layers 2–3, whereas negotiation integrated the full prompt at layers 2 and 4.Affective tasks showed dominant attribution gaps of +1.78 to +2.69 over the runner-up and +2.63 to +3.60 over the median; negotiation’s dominant signal came from mid-layer MLP components.Source: §4.1 · p. 4, Table 2 · p. 4

4.2 Human Evaluation

  • STAR led single-turn human evaluation on every empathy dimension and significantly beat prompt priming and global steering on multiple dimensions, while no multi-turn comparison survived correction.Single-turn gains over the unsteered baseline reached ∆=+1.12 for Perceived Naturalness and Emotional Quality; multi-turn means remained suggestive rather than statistically conclusive.Source: §4.2 · p. 5, Table 4 · p. 5

4.3 Automatic evaluation

  • Automatic evaluation showed the largest gains across turns: sentiment rose to ≈0.78 and affiliative politeness to 0.64–0.68, while empathy changed negligibly and softening markers decreased slightly.In single-turn emotional dialogue, STAR also produced significant gains in sentiment, empathy, affiliative politeness, and softening markers.Source: §4.3 · pp. 6–7, Figure 1 · p. 6

4.4 Qualitative Analysis and Diagnostics

  • Qualitative examples showed context-grounded, emotionally calibrated outputs, but low-scoring responses exhibited self-repetition, context mismatch, and register over-inflation.Removing injection in multi-turn settings was associated with lower Content Quality than sustained steering, although the difference was not significant.Source: §4.4 · p. 7, Table 5 · p. 7

5 Discussion and Conclusion

The discussion argues that complex interpersonal behaviors can be represented as localized, approximately linear activation directions, while emphasizing task-specific locations, multi-turn effects, failure modes, and scope limits. It situates STAR within prior alignment, activation-steering, linear-representation, and socially grounded dialogue work.

5 Discussion and Conclusion

  • Attribution gaps of +1.78 to +2.69 and transfer across held-out dialogues support approximately linear behavioral representations recoverable from small diagnostic sets.Reported shifts include sentiment (∆≈+0.45 multi-turn) and affiliative politeness (∆≈+0.49).Source: §5 · p. 8
  • Empathic traits localize in the final 15 prompt tokens at layers 2–3, whereas negotiation spans the full prompt at layers 2–4.The discussion links this dissociation to immediate affective cues versus discourse-level context integration.Source: §5 · p. 8
  • Localized injection amplifies affiliative politeness across turns, with multi-turn gains ≈20× larger than single-turn gains, while first-turn-only steering reaches Perceived Naturalness 3.88.Sustained steering is also associated with a modest empathy dip (∆<0.001), attributed to stylistic saturation.Source: §5 · p. 8
  • The linear approximation breaks down through verbatim repetition, context mismatch, and register over-inflation when fixed directions are activated without behavioral appropriateness.Proposed remedies are context-conditioned exemplars or adaptive coefficient schedules.Source: §5 · p. 8
  • STAR is presented as an interpretable, lightweight approach for identifying and injecting localized directions without sacrificing responsiveness to extended social interaction.The conclusion frames localization as central to steering complex interpersonal stances.Source: §5 · p. 8

A Ethical Considerations

  • The paper reports results for one steered model family, limited human-evaluation sample sizes, offline public datasets, and manually authored exemplars that may limit generalizability, significance, scalability, and cultural neutrality.It also identifies dual-use risks, including manipulation and deceptive rapport-building, requiring auditing and constraints beyond the study.Source: §5 · p. 9, A Ethical Considerations · p. 10

B. Related Work

  • Prior work established linear activation directions mainly for static or coarse attributes, while STAR extends targeted steering to dynamic, relational behaviors such as empathy and negotiation.Related work describes activation interventions, residual-stream steering, and the challenge of context-sensitive social traits.Source: §B.3 · p. 11, §B.2 · p. 11, §B.4 · p. 11

C Attribution Patching: Methodology and Full Heatmaps

STAR uses attribution patching to identify causally important layer–token locations and then applies localized steering there. Across four diagnostic tasks, attribution concentrates in early layers but separates emotional from negotiation token-position patterns, directly motivating task-specific injection.

C Attribution Patching: Methodology and Full Heatmaps

  • Attribution patching identifies an intervention locus ℓ∗ by aggregating head-level causal scores across 32 layers and selecting the layer with the highest positive attribution.The score compares desired and undesired token logits after replacing activations with clean-run counterparts.Source: C Attribution Patching: Methodology… · p. 12
  • Negotiation heatmaps retain the same early-locus structure, but Counter-Offer has larger attribution magnitudes than Strategic Politeness, requiring per-task colorbars.The reported ranges are approximately ±4.8 and ±2.0, respectively.Source: Figure 3 · p. 13
  • Across all four tasks, ℓ∗ lies in an early layer, ℓ∗ ∈ {2, 3, 4}, with concentrated positive early-head attribution and diffuse opposing late-layer MLP attribution.This pattern is used to parameterize STAR’s localized interventions.Source: C Attribution Patching: Methodology… · p. 13
  • Emotional tasks concentrate attribution in the final 15 prompt tokens and shallow components, whereas negotiation distributes attribution across full-prompt clusters with mid-layer MLP involvement.The dissociation motivates running attribution patching separately by task rather than assuming a shared injection point.Source: C Attribution Patching: Methodology… · p. 13

D Steering Vector Derivation and Application Details

  • STAR constructs and applies steering vectors through behavioral exemplar construction, vector computation, and coefficient selection, using target layers ℓ∗ = 2 for Empathic Support, ℓ∗ = 3 for Empathic Disclosure, and ℓ∗ ∈ {2, 4} for negotiation tasks.Source: D Steering Vector Derivation and… · p. 13

D.1 Designing Steering Probes

STAR uses controlled single-token dialogue probes and attribution patching to localize behavioral traits before constructing and applying steering vectors. Emotional and negotiation behaviors occupy different layer–token loci, while exemplar differences define the injected direction.

D.1 Designing Steering Probes

  • Single-token, 50-token probes isolate preference between semantic poles while enabling comparable full-context attribution across prompts.Probe contexts use Person A/Person B framing and domain-grounded conversational or marketplace scenarios.Source: §D.1 · pp. 13–15
  • Empathic Support concentrates in the final 15 tokens with embed and early MLP components leading, while Empathic Disclosure has a higher peak magnitude in the same region.The heatmaps use separate colorbars, with approximate ranges of ±3.3 for Support and ±4.3 for Disclosure.Source: Figure 4 · p. 14
  • Attribution patching identifies a dominant layer–position pair for each behavioral trait, with affective tasks concentrated in the final 15 tokens and negotiation spanning the full prompt.Affective loci occur at early layers, whereas negotiation uses deeper, mid-layer components and distributed token clusters.Source: Figure 4 · p. 14, Figure 5 · p. 15
  • Counter-Offer Generation spans three prompt clusters with dominant mid-layer MLP attribution, while Strategic Politeness shows lower peaks across the full token range.The negotiation loci contrast with the boundary-gated concentration observed for emotional tasks.Source: Figure 5 · p. 15

D.2 Behavioral Exemplar Construction

  • Behavioral exemplars use positive and negative conversational examples, with at least 11 examples per polarity class after augmentation and manual filtering.Texts are token-normalized so mean activation differences reflect behavioral content rather than sequence length.Source: §D.1 · p. 15, §D.2 · p. 16

D.3 Vector Computation

  • The steering vector is computed as the mean positive activation tensor minus the mean negative tensor, then added as α·vsteer at the localized target positions.At inference, injection occurs at the target layer and final 15 positions identified by attribution.Source: §D.3 · p. 16

D.4 Steering Coefficient Selection

STAR selects domain-specific steering coefficients through held-out validation sweeps that balance behavioral improvement against fluency or coherence costs. The resulting operating points are α=2.0 for emotion and α=2.5 for negotiation.

D.4 Steering Coefficient Selection

  • Validation sweeps evaluate emotional affect with SST-2 positive-sentiment probability and fluency with completion-only perplexity, using marginal efficiency to choose coefficients.The emotional sweep covers α ∈ {0.5, 1.0, 1.5, 2.0, 2.5, 3.0, 3.5, 4.0} on N=100 held-out BOLT SMS dialogues.Source: §D.4 · p. 16
  • 0.920 ± 0.007 affect at α=2.0, with perplexity 55 versus 0.500 affect and perplexity 18 unsteered, was selected as the last Pareto-efficient emotional operating point.At α=2.0 → 2.5, affect improves only 0.04 while perplexity rises 64%; beyond α=2.5, affect declines as perplexity increases.Source: §D.4 · p. 16, Figure 6 · p. 17
  • 0.78 agreement rate at α=2.5 is the negotiation sweep peak, and α=2.5 is selected before repetition and semantic coherence degrade more sharply.The negotiation sweep used N=100 held-out Craigslist Bargains dialogues.Source: §D.4 · pp. 16–17

E Experimental Design and Evaluation

STAR’s experimental design tests whether attribution-guided localized steering yields coherent behavior across both single-turn and extended multi-turn contexts. It compares localized intervention with instruction-level and unlocalized steering baselines.

E Experimental Design and Evaluation

  • The experiments are designed to evaluate behavioral coherence across single-turn and extended multi-turn contexts while comparing localized injection against instruction-level and unlocalized steering.Shared implementation details and domain-specific choices are separated across the experimental design.Source: E Experimental Design and Evaluation · p. 17

E.1 Emotional Support and Disclosure

The emotional evaluation compares steered and unsteered responses in single-turn and multi-turn dialogue settings. It measures whether behavioral changes persist, decay, reverse, or accumulate across repeated interactions.

E.1 Emotional Support and Disclosure

  • Single-turn evaluation generates paired unsteered and steered responses from BOLT SMS dialogue histories, focusing on linguistic and emotional features of the response.Steering is applied at the final 15 token positions of the target layer.Source: §E.1 · p. 17
  • Multi-turn evaluation uses a three-exchange loop with Mistral generating Person A responses and Llama generating Person B responses at two possible steering points.The experiment uses the final 10% of eligible dialogues, totaling 1,102 examples.Source: §E.1 · p. 17
  • Figure 7 evaluates single-turn Craigslist Bargain outcomes using Agreement Rate, Price Improvement, Affiliative Politeness, and Softening Markers across prompt, global-steering, unsteered, and STAR conditions.The setting contains n=5,155 examples and marks adjusted significance against the Prompt Baseline.Source: Figure 7 · p. 18
  • Figure 8 evaluates the same four outcome categories in multi-turn Craigslist Bargain dialogues, comparing Prompt Baseline, UU, and STAR (SS).The multi-turn setting contains n=515 examples and uses adjusted significance against the Prompt Baseline.Source: Figure 8 · p. 18
  • The emotional dialogue prompt asks Person A to express feelings, seek connection rather than quick advice, and adapt disclosure to Person B’s supportive or dismissive tone.This fixed prompt operationalizes the interactional setting used for the emotional evaluation.Source: §E.1 · p. 19
  • The four multi-turn configurations vary whether steering is absent, introduced at the second turn, removed after the first, or maintained across both turns.UU, US, SU, and SS enable analysis of persistence, decay, reversal, and cumulative behavioral change.Source: §E.1 · pp. 17, 19

E.2 Negotiation Experiments

The negotiation experiments evaluate STAR through single- and multi-turn bargaining setups, comparing localized steering with global, unsteered, and prompt-based controls. Behavioral outcomes are measured with politeness, agreement, price-improvement, and human-quality metrics.

E.2 Negotiation Experiments

  • STAR targets negotiation behavior at attribution-identified layers and token positions, with simulations initialized after a seller concession to activate bargaining dynamics.The study evaluates immediate counter-offer effects and multi-turn persistence using CraigslistBargain dialogues.Source: §E.2 · p. 19
  • Single-turn experiments use 5,155 valid dialogues, with Mistral-7B-Instruct as seller and steered Llama-3.1-8B as buyer.The steering vector targets counter-offer behavior at layers ℓ* ∈ {2, 4}.Source: §E.2 · p. 19

F Evaluation Metrics

  • Negotiation outcomes include agreement rate and price improvement, distinguishing accepted offers and economic gains from surface-level verbosity.Agreement is inferred from the seller’s final turn, while price improvement compares simulated and dataset final prices.Source: §E.2 · p. 19, Negotiation Outcomes · p. 21

E.3 Baseline Methodology

  • Baseline conditions include global activation steering, unsteered generation, and zero-shot instruction priming, enabling comparisons between localized internal intervention and language-level control.Global steering applies the same vector across layers and positions, removing STAR’s attribution-guided localization.Source: §E.3 · p. 20

F.1 Automatic Evaluation Metrics

  • Automatic evaluation combines categorical politeness markers with continuous empathy-related lexical measures and controls multiple comparisons using Benjamini–Hochberg FDR correction.Features include gratitude, apologizing, hedges, directness, indirect requests, dismissiveness, and seven Empath categories.Source: Affective Outcomes · p. 21, Negotiation Outcomes · pp. 21–22

F.2 Human Evaluation Protocol

The human evaluation protocol measures whether steering preserves usable, coherent, emotionally appropriate, and natural dialogue quality. It uses blinded multi-annotator ratings, composite dimensions, and corrected pairwise tests, while reporting underpowered comparisons and weaker reliability for contradiction judgments.

F.2 Human Evaluation Protocol

  • Human evaluation rates Content Quality, Contextual Fit, Emotional Quality, and Perceived Naturalness to detect failures that automatic distributional metrics cannot establish.The overall score is Overall = mean(CQ, CF, EQ, PN).Source: §F.2 · pp. 22–23
  • The study samples 200 items per domain from larger evaluation sets, with per-condition sample sizes ranging from N=10 to N=46.Items are stratified across steering conditions and selected with a fixed random seed.Source: §F.2 · p. 22
  • Annotator agreement improves after calibration, with mean ICC2k rising from 0.774 in training to 0.883 after calibration, while single-turn agreement exceeds multi-turn agreement.Table 9 reports mean κ = 0.764 for single-turn versus 0.644 for multi-turn items.Source: Table 8 · p. 22, §F.2 · p. 22, Table 9 · p. 23
  • Annotations use 1–5 Likert scales, reverse-code selected items, exclude categorical contradiction and disputed epistemic-accuracy items from composites, and aggregate by majority vote.When all three annotators disagree, the median rounded to the nearest integer is used.Source: §F.2 · pp. 22–23
  • STAR is the only single-turn condition with corrected significant differences, outperforming Prompt Baseline on CQ, CF, PN, and Overall and Global Steering on EQ and Overall.No multi-turn pairwise comparison survives correction, and the observed maximum overall difference is Δ = +1.33 versus Prompt Baseline.Source: §F.2 · pp. 23–24
  • The evaluation is underpowered for moderate effects at its sample sizes, and CQ5 requires extra caution because it has weaker and more volatile agreement than other dimensions.Results with .01 < padj < .05 are described as suggestive rather than confirmatory.Source: §F.2 · p. 24

F.3 Extended Negotiation Results

Extended negotiation analyses show that localized steering changes accommodation and politeness patterns, with effects driven primarily by each agent’s own steering state. STAR generally retains an advantage over global steering and differs mechanistically from prompt-based control, while economic outcomes remain unchanged.

F.3 Extended Negotiation Results

  • STAR produces equal or larger effects than global steering on every directly compared feature, with hedging and apologizing most sensitive to attribution-guided layer selection (V = 0.11 and 0.12).The result is consistent with behaviors being concentrated in a narrow band of mid-network layers.Source: §F.3 · p. 24, Table 12 · p. 28
  • Global steering is statistically indistinguishable from the prompt baseline on four of six categorical features, whereas STAR separates from the baseline on apologizing and indirect requests.No baseline comparison significantly changes agreement rate or price improvement.Source: §F.3 · p. 24, Table 13 · p. 29
  • In the single-turn human-quality comparison, STAR significantly exceeds Global Steering by +0.77 and Prompt Baseline by +1.33, while multi-turn pairwise comparisons remain non-significant.Figure 10 reports padj < 0.05 for these single-turn comparisons and padj ≥ .97 throughout the multi-turn tier.Source: Figure 10 · p. 25, Table 10 · p. 26, Table 11 · p. 27
  • STAR shifts buyer behavior toward accommodation across all six politeness markers, with medium effects on gratitude (V = 0.31), indirect requests (V = 0.30), and directness (V = 0.30).Only price improvement reaches significance under both localized and global steering.Source: Table 12 · p. 28
  • Asymmetric results indicate that steering effects are agent-local: UU differs from both US and SU, while SS versus US is null with all V ≤ 0.08.The authors note that this null could reflect a behavioral ceiling rather than strict agent independence.Source: §F.3 · p. 24, Table 14 · p. 30
  • Bilateral steering sustains accommodation shifts across all six categorical features, including medium effects for gratitude (V = 0.34), apologizing (V = 0.31), and indirect requests (V = 0.36).Continuous outcomes are uniformly non-significant, so steering changes negotiation style without changing agreements or price outcomes.Source: Table 15 · p. 31

Related papers

Loading 2511.12832v3…