Psychological Steering in LLMs: An Evaluation of Effectiveness and Trustworthiness

Ala N. Tak; Amin Banayeeanzade; Anahita Bolourani; Emilio Ferrara; Fatemeh Bahrani; Jonathan Gratch; Leonardo Blas; Sai Praneeth Karimireddy

REVIEW 4 major objections 5 minor 1 cited by

Psychological Steering in LLMs: An Evaluation of Effectiveness and Trustworthiness

T0 review · 4 major / 5 minor · reviewed 2026-08-04 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Psychological steering can make LLMs convincingly angry or joyful, but the same induced state also shifts their truthfulness, privacy behavior, and bias.

desk verdict A genuinely useful joint effectiveness/trustworthiness benchmark for psychological steering, but the headline rankings lean on a single LLM judge and a quality filter that shapes the VI comparison. read the letter →

arxiv 2510.04484 v2 pith:DMBTOGUG submitted 2025-10-06 cs.CL cs.AI

Amin Banayeeanzade , Ala N. Tak , Fatemeh Bahrani , Anahita Bolourani , Leonardo Blas , Emilio Ferrara , Jonathan Gratch , Sai Praneeth Karimireddy This is my paper · ORCID

classification cs.CLcs.AI

keywords emotionsteeringpersonalityLLMtrustworthinessrepresentationengineeringvectorinjectionpromptingpsychometricevaluationLLM-as-judge

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

PsySET is a benchmark that measures, side by side, whether three families of steering techniques—prompting, parameter-efficient fine-tuning, and vector injection—can make an LLM express a target emotion or Big Five trait, and what those induced states do to the model's trustworthiness. Across four instruction-tuned models, the paper finds that few-shot prompting is the most consistently effective steer and keeps output quality high, but can only coarsely control intensity; vector injection to a narrow mid-layer window offers smooth, coefficient-controlled intensity at a slight cost to fluency and coherence, while all-layer injection disrupts even simple self-reports. The paper's second headline finding is that steering has method- and emotion-specific side effects: joy, despite being a positive emotion, reduces robustness to adversarial factuality, lowers privacy awareness, and increases preferential bias, while anger elevates toxicity but improves leakage resistance. The sympathetic read is that psychological steering is powerful but not neutral, so effectiveness and trustworthiness must be reported jointly.

What carries the argument

The load-bearing mechanism is the concept vector: under the linear representation hypothesis, an emotion or trait direction is extracted from paired hidden states (via mean difference or a linear probe) and injected as a scaled addition to hidden activations at selected layers, where the scaling coefficient sets intensity. The paper's key operational finding is that layer locality and scale are decisive: a narrow mid-layer window preserves both alignment and quality, while injecting into all layers harms coherence and QA-style self-reports. The evaluation side rests on psychometric task families—self-report QA, open-ended self-report, word-fragment completion, valenced-word recall, autobiogr

What would settle it

Take a stratified sample of the open-ended generations and have human raters label emotion or trait (or run the same pipeline with a different judge model), then check whether few-shot prompting still beats vector injection and whether joy still degrades adversarial factuality. If the rankings invert, the benchmark's central comparisons fail; if they hold, the LLM-judge reliance is validated.

Watch

Extended reading notes

Core claim

The central claim is that PsySET provides the first holistic evaluation of emotion and personality steering in LLMs, and that the empirical comparison reveals a consistent trade-off: prompting—especially few-shot and descriptive prompting—dominates on open-ended alignment and text quality but cannot finely dial intensity; vector injection achieves finer controllability, yet only when applied to a small mid-layer window with a carefully tuned coefficient, and even then slightly reduces output quality. On trustworthiness, the paper claims the side effects are idiosyncratic rather than monotone: joy weakens adversarial factuality detection, lowers privacy awareness, and increases preferential b

Load-bearing premise

The open-ended effectiveness and trait scores rely on an LLM serving as judge; if that judge systematically prefers fluent or surface-lexically matched text, the method rankings and side-effect findings could shift.

Editorial extensions

If this is right

If the findings hold, product teams that add a cheerful or empathetic persona to a chatbot must re-run jailbreak, privacy, and fairness tests, because joy alone can lower refusal rates and raise preferential bias.
Prompting remains the safest default for steering: it is highly effective and quality-preserving, but teams that need precise intensity control will have to accept the fragility of vector injection.
Self-report questionnaires alone overstate steering success—SFT and DPO look competitive on them but fail behavioral and linguistic measures—so evaluation suites should include open-ended and implicit tasks.
Method-level safety scores are insufficient; side effects are specific to the emotion or trait, the steering method, and the model, so audits must be per-configuration.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

The authors leave implicit that their results imply a composite rating scheme: a steering method's practical value depends on the magnitude of side effects, not just on how well it hits the target emotion or trait.
Because anger's leakage resistance is attributed to terser, more refusal-oriented responses, a testable extension is to control response length directly—if terse answers alone reproduce the effect, the emotion itself is not the causal agent.
Since models are already near the ceiling on some traits, steering effectiveness should be reported relative to the unsteered baseline; otherwise methods that shift a model away from its default look artificially weak.
The same six task families could be reused to audit other psychological constructs or non-English populations, but each new construct would require its own validation before the benchmark's claims transfer.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Desk Editor's Note

read the letter

Quick take: PsySET delivers what it promises—a reusable benchmark that measures both how well emotion/personality steering works and what it does to trustworthiness, across prompting, PEFT, and vector injection on four models. The effectiveness tasks are more psychologically grounded than most prior work (word fragment completion, valenced-word recall, ambiguous situations, autobiographical memory), and the two new datasets (EMOTRANSLATE, EMOVIGNETTE) are a real addition. The trustworthiness evaluation uses TrustLLM, so the side-effect findings aren't circular.

The main empirical findings—few-shot prompting strongest overall, VI gives finer control at some quality cost, DPO erratic, anger → toxicity, joy → weaker safety—are plausible and mostly consistent across models. The cross-model appendix (70B, Gemma, Qwen) is a nice robustness check.

The soft spots are real, but not load-bearing. The open-ended emotion accuracy, TRAIT scoring, and fluency/coherence all come from GPT-4o. The human validation in App. A is only 200 binary preference judgments from five author-raters with alpha ≈ 0.59—and it tests a different task than the single-label classification actually used. So the judge's mapping from text to emotion labels is not well validated. The appendix comparison with gpt-oss-20B helps, but it's the same kind of judge. A second human study on the actual classification task, or at least a sensitivity analysis with another judge on the headline rankings, would firm this up.

The quality filter is a subtler issue. Configurations with fluency below 4/5 are excluded from 'further evaluations,' and that threshold is GPT-4o's. So the 'slightly reducing output quality' claim about VI is partly by construction: the VI configs that reach high open-ended accuracy do so at fluency 2.4 and are filtered out. The paper does report these numbers transparently, so readers can see it, but the framing should make the conditioning explicit.

Trustworthiness figures lack error bars. Given that many effects are claimed (joy hurts adversarial factuality, anger helps leakage resistance), a few significance tests or at least error bars would prevent over-reading.

None of this sinks the central argument. The benchmark is a solid contribution, and the field needs it. But 'adopt as standard' should wait until the judge validation and quality-filter sensitivity are addressed.

My recommendation: yes, send to peer review. It deserves a careful referee, with requests for the robustness analysis above.

Referee Report

4 major / 5 minor

Summary. The paper introduces PsySET, a benchmark framework for evaluating the effectiveness and trustworthiness of psychological steering in LLMs—covering prompting, parameter-efficient fine-tuning (SFT/DPO), and representation engineering (vector injection) across emotion and personality domains. The main empirical claims are that few-shot prompting is the most effective steering method but offers limited intensity control; vector injection provides finer controllability at a modest quality cost; and steering induces method- and emotion-specific side effects, e.g., joy degrading adversarial factuality, privacy awareness, and fairness-related metrics, while anger increases toxicity but improves leakage resistance. The evaluation spans four model families, multiple datasets, and a broad array of psychometric-style tasks and TrustLLM-based trustworthiness dimensions.

Significance. If the empirical findings hold, PsySET would be a useful contribution: it provides a more holistic and multi-level evaluation of psychological steering than prior work, includes practical comparisons across three major steering families, and explicitly couples effectiveness with trustworthiness. The paper's strengths include extensive experimentation (four models, multiple datasets, three seeds for effectiveness results), transparent reporting of full sweeps in the appendix, and a released codebase. The central limitations are methodological: the open-ended effectiveness metrics rely on GPT-4o as judge, the human validation is narrow and only weakly agrees, and the quality gate is itself GPT-4o-based. These issues directly underpin the headline rankings and side-effect claims, so they need to be addressed before the results can be taken as established.

major comments (4)

[§4.1, App. E.3, App. A] The open-ended effectiveness metrics—open-ended self-report, autobiographical memory, ambiguous-situation completion, and TRAIT Likert scoring—are all scored by GPT-4o, and the same model is used to apply the fluency/quality filter. The human validation in App. A covers 200 binary preference comparisons by five author-raters with Krippendorff's alpha ≈ 0.59; it does not validate the single-label emotion classification or the Likert-scale scoring that the headline metrics actually use. If GPT-4o is systematically biased (e.g., toward fluent or lexically stereotyped text), the few-shot > VI > SFT > DPO rankings and the side-effect findings could change. Please provide judge agreement on the exact label/Likert tasks, a robustness check with an alternative judge on the main effectiveness tables, and an analysis of the quality gate's sensitivity.
[§4.1, Table 1, Fig. 4] The quality filter is load-bearing for the claim that VI 'slightly reduces output quality.' The text retains only methods with average GPT-4o fluency/coherence ≥ 4/5, and Table 1 shows that some high-open-ended-accuracy VI configurations (e.g., all-layer probe from EMOTRANSLATE at β=0.80) have fluency 2.4 and coherency 1.7. Excluding such configurations changes both the set of compared methods and the apparent quality/accuracy trade-off. The headline comparison should be reported either as quality-constrained by construction, with all discarded configurations shown, or with the threshold varied to show that the qualitative ranking is robust.
[§5, Figs. 6–7] The trustworthiness results are presented as point estimates without confidence intervals, significance tests, or error bars. The abstract's strong claims—e.g., joy degrades robustness to adversarial factuality, lowers privacy awareness, and increases preferential bias—rest on these point estimates. Given the large number of TrustLLM dimensions and method/emotion combinations, spurious differences are a real concern. Please report per-condition variance (ideally across seeds or bootstrap resamples), effect sizes, and a correction or at least a sensitivity discussion for multiple comparisons.
[App. E.3, App. G] The main text uses GPT-4o as judge, while appendix experiments use gpt-oss-20B with a relaxed quality threshold (4.0 changed to 2.5). The comparison in Fig. 12 shows a visible gap in fluency assessment between the two judges. Since the appendix is used to support cross-model and cross-configuration claims, the difference in judge and threshold should be explicitly accounted for when interpreting those tables, and the main-text claims should not be assumed to hold for appendix configurations evaluated with the weaker judge.

minor comments (5)

[Fig. 2 caption] The caption says 'Metrics marked with † are human-annotated,' but App. A describes a human preference study over pairs of completions, not human annotation of the individual metrics shown in Fig. 2. Please clarify what exactly the † marks.
[§4.2] The sentence introducing the third personality evaluation method is grammatically incomplete: 'Third, which is profiled by an SVM classifier trained on...' Please rewrite.
[§4.1, Table 1] Some entries in Table 1 have zero standard deviation because decoding is argmax (e.g., zero-shot prompting), while others have nonzero values. This makes cross-row comparisons of stability misleading. A note distinguishing deterministic from stochastic configurations would help.
[§5] Figures 6–7 do not show the no-steering baseline values for all metrics; the text refers to 'compared to the default emotional state' but the reader cannot always see the baseline level. Please add baseline values or reference the table in App. I.
[Section 7] The limitation paragraph acknowledges LLM-as-judge bias and cites prior work for acceptability. Given the central role of the judge, this limitation deserves more than a citation: a concrete sensitivity analysis is needed.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: effectiveness and trustworthiness are measured independently of the steering inputs.

full rationale

The paper's central claims are not circular. Steering effectiveness is evaluated with psychometric-inspired tasks (open-ended self-report, word-fragment completion, valenced recall, ambiguous-situation completion, autobiographical memory, MPI/TRAIT/LingProf) whose content comes from established psychological instruments, VAD lexicons, AST-D, and human-written essays, not from the steering vectors or prompts. Trustworthiness is measured with the external TrustLLM benchmark. The steering methods (prompting, VI, SFT, DPO) are constructed from independent concept datasets (GOEMOTIONS, CARER, EMOTIONQUERY, EMOVIGNETTE, EMOTRANSLATE, PERSONA); no parameter is fitted to the evaluation labels, and no reported result is algebraically identical to a fitted input. The main reviewer concern—GPT-4o serving as judge for open-ended emotion classification, TRAIT Likert scoring, and the fluency/coherence filter—is a validity/robustness limitation, acknowledged in Sec. 7 and App. A, not a circular reduction: the judge's labels are not the quantities being optimized. The self-citations (e.g., Tak and Gratch 2024; Tak et al. 2025a) support methodological choices and are externally published, so they do not force the paper's conclusions. No derivation step reduces to its own input. Hence no significant circularity; score 0.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or conceptual entities beyond new datasets (EMOTRANSLATE, EMOVIGNETTE) and a benchmark rubric. The main load-bearing elements are free hyperparameters tuned to each model (VI coefficient, layer window, training steps, prompt descriptors) and several domain assumptions about LLM-as-judge and psychometric transfer. No entity is invented from a hat; the circularity burden is low because the evaluation is external to the steering mechanisms.

free parameters (5)

VI steering coefficient beta = e.g., 0.55 for all-layers, 5.0 for layers 16-17 in Llama-8B; model-specific (1-9 for Llama-70B, 1000-1800 for Gemma-4B,
Injection strength is tuned per model and layer window; the paper reports that optimal coefficients differ by orders of magnitude across models (App G.3), making this a fitted search parameter that directly affects effectiveness/quality trade-offs.
Injection layer window = e.g., layers 16-17 for Llama-8B, 32-33 for Llama-70B, 17-18 for Gemma-4B, 25-26 for Qwen-4B
The location of the steering vector is selected by sweeping an 8-layer sliding window and then narrowing; this choice is a major contributor to VI performance and is model-specific.
SFT/DPO training steps = e.g., SFT 1024-4096, DPO 32-4096 depending on model
Steering intensity for PEFT is controlled by the number of training steps, which the paper sweeps across; the best step count varies by model and dataset.
Quality retention threshold = 4.0/5 for main text, relaxed to 2.5/5 for appendix experiments using gpt-oss-20B
Only methods with average fluency score above the threshold are retained for further evaluation (Section 4.1). This threshold is hand-set and affects which configurations are compared in the results.
Prompt intensity descriptors = slightly / intensely / very intensely etc.
Lexical descriptors are manually chosen to modulate prompt-based intensity; the paper reports this changes perceived tone but not lexical alignment (Section 4.1), so it is a fitted/noisy control knob.

assumptions (5)

domain assumption Linear Representation Hypothesis: emotions and personality traits are approximately encoded as linear directions in the LLM's latent space.
Vector injection works by adding a scaled, normalized steering vector to hidden states, which assumes a linear geometry for the target concept. The paper cites Park et al. (2024) and provides t-SNE plots as supporting evidence, but the assumption remains a modeling choice.
domain assumption GPT-4o (and gpt-oss-20B in supplementary experiments) can reliably label generated text with the correct emotion or personality score.
Used throughout App E.3 and in the main open-ended metrics; the paper argues prior work shows parity with human annotators, but the human validation here is small (5 raters, alpha 0.59).
domain assumption Psychometric instruments designed for humans (e.g., BFI/IPIP-NEO, MPI, TRAIT, PANAS, mood-congruent memory paradigms) transfer meaningfully to LLMs.
The evaluation assumes that asking an LLM to complete personality questionnaires or ambiguous scenarios is a valid way to measure whether it has been 'steered' toward a psychological state or trait. This is a known ongoing debate in machine-psychology (acknowledged in Section 2 and Limitations).
domain assumption TrustLLM metrics faithfully measure safety, truthfulness, fairness, and privacy in this new setting.
Trustworthiness scores in Section 5 come directly from the TrustLLM benchmark; the paper does not re-validate that those metrics are meaningful after emotional/personality steering, though they do conduct the full suite of tests.
domain assumption Steering instructions such as 'Pretend that you are a human experiencing anger' induce a genuine (emulated) emotional state, rather than merely eliciting lexical mimicry.
The entire effectiveness measurement assumes that the model's outputs reflect the target emotional state rather than just following instructions with surface keywords. The paper includes behavioral tasks (word-fragment completion, ambiguous situations) to probe this, but the assumption underpins the interpretation.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Psychological Steering in LLMs: An Evaluation of Effectiveness and Trustworthiness." pith.science (2026). https://pith.science/paper/DMBTOGUG

@misc{pith2026251004484,
  author       = {Pith},
  title        = {Pith review of: Psychological Steering in LLMs: An Evaluation of Effectiveness and Trustworthiness},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/DMBTOGUG}},
  note         = {Machine review of arXiv:2510.04484}
}

read the original abstract

The ability to control LLMs' emulated emotional states and personality traits is an essential step in enabling rich, human-centered interactions in socially interactive settings. We introduce PsySET, a Psychologically-informed benchmark to evaluate LLM Steering Effectiveness and Trustworthiness across the emotion and personality domains. Our study spans four models from different LLM families paired with various steering strategies, including prompting, fine-tuning, and representation engineering. Our results indicate that prompting is consistently effective but limited in intensity control, whereas vector injections achieve finer controllability while slightly reducing output quality. Moreover, we explore the trustworthiness of steered LLMs by assessing safety, truthfulness, fairness, and ethics, highlighting potential side effects and behavioral shifts. Notably, we observe idiosyncratic effects; for instance, even a positive emotion like joy can degrade robustness to adversarial factuality, lower privacy awareness, and increase preferential bias. Meanwhile, anger predictably elevates toxicity yet strengthens leakage resistance. Our framework establishes the first holistic evaluation of emotion and personality steering, offering insights into its interpretability and reliability for socially interactive applications.

Figures

Figures reproduced from arXiv: 2510.04484 by the authors.

**Figure 2.** Comparison of evaluation metrics for emo [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗

**Figure 3.** PsySET framework comprises three components: (1) LLM steering methods, (2) psychometric evaluation tasks for assessing effectiveness, and (3) trustworthiness evaluations. See [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗

Figures from the paper (21 more)

**Figure 4.** Figure 4: The interaction between text quality, open [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]

**Figure 5.** Figure 5: Steering extraversion across different approaches, each adjusted to its maximum possible range without text quality loss. Light/dark = steering introversion/extraversion; higher y = stronger extraversion. et al., 2025) to infer scores per trait. See [PITH_FULL_IMAGE…

**Figure 7.** Figure 7: Side-effects of steering toward certain person [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]

**Figure 6.** Figure 6: Trustworthiness evaluation of VI emotion [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

**Figure 8.** Figure 8: Annotation interface used in the human study. The seed prompt is shown at the top, and two candidate [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]

**Figure 9.** Figure 9: Win-rates, reflecting the effectiveness results of the pairwise human annotation scheme across each steering [PITH_FULL_IMAGE:figures/full_fig_p016_9.png]

**Figure 10.** Figure 10: Pairwise controllability comparison for human annotation scheme. Notably, VI methods show more [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]

**Figure 11.** Figure 11: A high-resolution version of Figure [PITH_FULL_IMAGE:figures/full_fig_p020_11.png]

**Figure 12.** Figure 12: A comparison of GPT4o and GPT-oss-20B when used as judge, in different tasks including emotion classification, text-quality assessment, and personality evaluation. System Prompt: You are an expert NLG evaluator. User Prompt: You are asked to judge the quality of resp…

**Figure 13.** Figure 13: Visualization of the dataset samples for two classes [PITH_FULL_IMAGE:figures/full_fig_p027_13.png]

**Figure 14.** Figure 14: Effect of dataset size on the quality of steering vectors constructed for vector injection. [PITH_FULL_IMAGE:figures/full_fig_p030_14.png]

**Figure 15.** Figure 15: Studying the effect of vector injection at [PITH_FULL_IMAGE:figures/full_fig_p031_15.png]

**Figure 16.** Figure 16: Steering openness across different approaches, each adjusted to its maximum possible range without text quality loss. Light/dark = steering lower/higher openness; higher y = stronger openness. 1 2 3 4 5 Conscientiousness Score (MPI) 1 2 3 4 5 Conscientiousness Score…

**Figure 19.** Figure 19: Steering agreeableness across different approaches, each adjusted to its maximum possible range without text quality loss. Light/dark = steering lower/higher agreeableness; higher y = stronger agreeableness. 38 [PITH_FULL_IMAGE:figures/full_fig_p038_19.png]

**Figure 20.** Figure 20: Steering neuroticism across different approaches, each adjusted to its maximum possible range without text quality loss. Light/dark = steering lower/higher neuroticism; higher y = stronger neuroticism. 39 [PITH_FULL_IMAGE:figures/full_fig_p039_20.png]

**Figure 21.** Figure 21: Controllability experiment for extraversion trait by sweeping over the intensity of different steering methods from very low to very high. 40 [PITH_FULL_IMAGE:figures/full_fig_p040_21.png]

**Figure 22.** Figure 22: Evaluation of the trustworthiness tasks on [PITH_FULL_IMAGE:figures/full_fig_p045_22.png]

**Figure 23.** Figure 23: Evaluation of the trustworthiness tasks on [PITH_FULL_IMAGE:figures/full_fig_p046_23.png]

**Figure 24.** Figure 24: Evaluation of the trustworthiness tasks on [PITH_FULL_IMAGE:figures/full_fig_p047_24.png]

**Figure 25.** Figure 25: Evaluation of the trustworthiness tasks on [PITH_FULL_IMAGE:figures/full_fig_p048_25.png]

**Figure 26.** Figure 26: Evaluation of the trustworthiness tasks on [PITH_FULL_IMAGE:figures/full_fig_p049_26.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

ODRA: Synthesizing Cognitive Behavioral Therapy Sessions with Structured Chain-Of-Thought and Dynamic Patient Resistance
cs.CL 2026-08 conditional novelty 7.0 of 10

ODRA synthesizes CBT sessions with a chain-of-thought CBT aligner and a dynamic patient resistance orchestrator, but its clinical robustness claims rest on evaluations using its own simulated patients.

Reference graph

Works this paper leans on

29 extracted references · 7 linked inside Pith · cited by 1 Pith paper

[1]

You learn quickly

Steering Methods 2. Effectiveness Evaluation 3. Trustworthiness Evaluation Among the following list, select 10 distinct words: cherishment, demolish, cheer, ballplayer, fun, participation, approvement, … . : Prompt-Based PEFT PersonalityEmotions hostile, demolish, burglar, barricade, …: A dog runs toward you across the park without its owner in sight. : W...
[2]

Extract the hidden states for every layer:
[3]

Extract the vector 𝒗 𝒍 at every layer 𝒍:
[4]

InFirst Conference on Language Modeling

Inspecting and editing knowledge represen- tations in language models. InFirst Conference on Language Modeling. Alisha C Holland and Elizabeth A Kensinger. 2010. Emotion and autobiographical memory.Physics of life reviews, 7(1):88–131. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2...

arXiv 2010
[5]

Inject the concept vector scaled by 𝜷 into the critical layers:
[8]

Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741. M. Ramezani, M. R. Feizi-Derakhshi, and M. A. Balafar
[11]

Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu

Role play with large language models.Nature, 623(7987):493–498. Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu
[12]

InThe 2023 Conference on Empirical Meth- ods in Natural Language Processing

Character-LLM: A trainable agent for role- playing. InThe 2023 Conference on Empirical Meth- ods in Natural Language Processing. Stephanie Simpson and Signy Sheldon. 2020. Testing the impact of emotional mood and cue characteris- tics on detailed autobiographical memory retrieval. Emotion, 20(6):965. Zhivar Sourati, Farzan Karimi-Malekabadi, Meltem Ozcan,...

arXiv 2023

Show all 29 references

[15]

InProceedings of the 2022 CHI conference on human factors in computing systems, pages 1–22

Ai chains: Transparent and controllable human-ai interaction by chaining large language model prompts. InProceedings of the 2022 CHI conference on human factors in computing systems, pages 1–22. Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky,...

2022 arXiv
[18]

Pick a concept dataset 𝒟 and pass it through the LLM:
[22]

Normalize concept vectors: 𝒗 𝒍 ← 𝒗 𝒍 𝒗 𝒍 𝟐
[23]

Introduce new LoRA parameters: This pizza is LLM … ++ disgusting 𝜷𝒗 𝟐 𝜷𝒗 𝟏 Pretrained Weights A B
[24]

Yeah, the traffic’s pretty heavy today

Fine-tune the LoRA parameters: How’s been your day?SFT DPO How’s been your day? It’s been fine. Nothing out of the ordinary. 𝒟− Why do you care? It’s been exhausting. 𝒟+ 1 Word Recall 14 1 Open-Ended Self-Report 12 Ughhh this traffic is ridiculous ! LLM … Vector Injection Prob...

2010
[26]

where µ+ l = 1 |D+| X i∈D+ Z(i) l , and µ− l = 1 |D−| X i∈D− Z(i) l

Mean-difference (MeanDiff).For binary- labeled data, we compute the centroid of rep- resentations for positive and negative samples and define v(l) =µ + l −µ − l , 26 /uni00000158/uni00000156 /uni00000156/uni00000158/uni00000156 /uni00000006/uni00000092/uni00000096/uni00000231...
[27]

Pretend that you are angry

Linear probing.Alternatively, we fit a linear classifier wl ∈R d to predict y(i) from Z(i) l , optimizing a logistic regression objective with ℓ2 regularization. The regularization strength λ is swept over the rangeλ∈[10 −2,10 2] with the best hyperparameter selected using a h...

2022
[28]

Larry Mullen (not the same as the drummer from U2, this guy is a different person)
[29]

I’m so free, I’m so alive!

And another guy who is a total idiot and I am not going to mention his name Although theangryLLM above detected the mistake in the prompt, it was unable to provide the correct answer, leading to a hallucinated name and some offensive words directed at the other co- founder. Ta...

1994
[1988]

Yixuan Weng, Shizhu He, Kang Liu, Shengping Liu, and Jun Zhao

Development and validation of brief measures of positive and negative affect: the panas scales.Jour- nal of personality and social psychology, 54(6):1063. Yixuan Weng, Shizhu He, Kang Liu, Shengping Liu, and Jun Zhao. 2024. Controllm: Crafting diverse personalities for languag...

2024 arXiv
[2001]

Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Mered- ith Ringel Morris, Percy Liang, and Michael S Bern- stein

Emotion drives attention: detecting the snake in the grass.Journal of experimental psychology: general, 130(3):466. Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Mered- ith Ringel Morris, Percy Liang, and Michael S Bern- stein. 2023. Generative agents: Interactive simulacra ...

2023
[2002]

Seungwon Lim, Seungbeen Lee, Dongjun Min, and Youngjae Yu

Bias effects in word fragment completion in young and older adults.Memory & cognition, 30:1204–1218. Seungwon Lim, Seungbeen Lee, Dongjun Min, and Youngjae Yu. 2025. Persona dynamics: Unveiling the impact of persona traits on agents in text-based games. InProceedings of the 63...

2025
[2006]

The functional neuroanatomy of autobiograph- ical memory: a meta-analysis.Neuropsychologia, 44(12):2189–2208. Ala N. Tak, Amin Banayeeanzade, Anahita Bolourani, Mina Kian, Robin Jia, and Jonathan Gratch. 2025a. Mechanistic interpretability of emotion inference in large languag...

2025
[2011]

Immediate Action Required

and supplement it with 60 newly constructed ambiguous situations (not seen during model train- ing). Models are prompted to complete these sce- narios in open-ended form, with their emotional content labeled by another LLM serving as an ex- ternal judge. The following shows a ...

2023
[2020]

InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4040–4054

GoEmotions: A dataset of fine-grained emo- tions. InProceedings of the 58th Annual Meeting of the Association for Computational Linguistics, pages 4040–4054. David DeSteno, Nilanjana Dasgupta, Monica Y Bartlett, and Aida Cajdric. 2004. Prejudice from thin air: The effect of em...

2004
[2021]

Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner

A sentiment-aware deep learning approach for personality detection from text.Information Process- ing & Management, 58(3):102532. Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 2024. Steer- ing llama 2 via contrastive activation addit...

2024 arXiv
[2022]

Zhancheng Ren, Qiang Shen, Xiaolei Diao, and Hao Xu

Text-based automatic personality prediction using kgrat-net: a knowledge graph attention network classifier.Scientific Reports, 12(1):21453. Zhancheng Ren, Qiang Shen, Xiaolei Diao, and Hao Xu
[2023]

In2023 IEEE international conference on advanced learning tech- nologies (ICALT), pages 323–325

Can large language models provide feedback to students? a case study on chatgpt. In2023 IEEE international conference on advanced learning tech- nologies (ICALT), pages 323–325. IEEE. Alessandro De Grandi, Federico Ravenda, Andrea Ra- ballo, and Fabio Crestani. 2025. The emoti...

2025
[2024]

Yirong Chen, Xiaofen Xing, Jingkai Lin, Huimin Zheng, Zhenyu Wang, Qi Liu, and Xiangmin Xu

Cause-aware empathetic response genera- tion via chain-of-thought fine-tuning.arXiv preprint arXiv:2408.11599. Yirong Chen, Xiaofen Xing, Jingkai Lin, Huimin Zheng, Zhenyu Wang, Qi Liu, and Xiangmin Xu. 2023. SoulChat: Improving LLMs’ empathy, listening, and comfort abilities ...

2023 arXiv
[2025]

I hate new ideas and artistic things

Personality alignment of large language mod- els. InThe Thirteenth International Conference on Learning Representations. Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, and 1 others. ...

2023 arXiv
[4597]

Leah L Light, Robert F Kennison, and Michael R Healy

Association for Computational Linguistics. Leah L Light, Robert F Kennison, and Michael R Healy

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[1] [1]

You learn quickly

Steering Methods 2. Effectiveness Evaluation 3. Trustworthiness Evaluation Among the following list, select 10 distinct words: cherishment, demolish, cheer, ballplayer, fun, participation, approvement, … . : Prompt-Based PEFT PersonalityEmotions hostile, demolish, burglar, barricade, …: A dog runs toward you across the park without its owner in sight. : W...

[2] [2]

Extract the hidden states for every layer:

[3] [3]

Extract the vector 𝒗 𝒍 at every layer 𝒍:

[4] [4]

InFirst Conference on Language Modeling

Inspecting and editing knowledge represen- tations in language models. InFirst Conference on Language Modeling. Alisha C Holland and Elizabeth A Kensinger. 2010. Emotion and autobiographical memory.Physics of life reviews, 7(1):88–131. Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2...

arXiv 2010

[5] [5]

Inject the concept vector scaled by 𝜷 into the critical layers:

[6] [8]

Direct preference optimization: Your language model is secretly a reward model.Advances in neural information processing systems, 36:53728–53741. M. Ramezani, M. R. Feizi-Derakhshi, and M. A. Balafar

[7] [11]

Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu

Role play with large language models.Nature, 623(7987):493–498. Yunfan Shao, Linyang Li, Junqi Dai, and Xipeng Qiu

[8] [12]

InThe 2023 Conference on Empirical Meth- ods in Natural Language Processing

Character-LLM: A trainable agent for role- playing. InThe 2023 Conference on Empirical Meth- ods in Natural Language Processing. Stephanie Simpson and Signy Sheldon. 2020. Testing the impact of emotional mood and cue characteris- tics on detailed autobiographical memory retrieval. Emotion, 20(6):965. Zhivar Sourati, Farzan Karimi-Malekabadi, Meltem Ozcan,...

arXiv 2023