Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

PersonaFuse claims a post-trained 8B LLM can switch its expressed Big Five traits query by query, beating fine-tuning baselines by 38–69% on emotion benchmarks without losing reasoning or safety.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-04 22:18 UTC pith:7HNWHK4I

load-bearing objection Solid new post-training recipe for personality-adaptive LLMs, but the no-degradation claim rests on a missing control group and the absolute GSM8k numbers contradict it. the 3 major comments →

arxiv 2509.07370 v2 pith:7HNWHK4I submitted 2025-09-09 cs.CL

PersonaFuse: A Personality Activation-Driven Framework for Enhancing Human-LLM Interactions

Yixuan Tang , Yi Yang , Ahmed Abbasi This is my paper
classification cs.CL
keywords personality adaptationBig Five modelTrait Activation Theorymixture of expertsLLM post-trainingsocial-emotional intelligencepersona routingchain-of-thought data generation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

PersonaFuse is a post-training recipe that lets a large language model change its expressed personality from one query to the next, instead of using the same generic tone for every user. Drawing on the Big Five model, it builds ten expert adapters — one for the high and low end of each trait — and a learned router that reads each query and mixes the experts whose traits fit the situation. The paper reports that an 8B Llama-3.1 model trained this way beats its fine-tuned baselines by roughly 38–69% on emotional-intelligence benchmarks, improves or holds its scores on reasoning and safety tests where the baselines visibly degrade, and wins human-preference comparisons on emotion-based dialogues against several larger commercial models. The authors offer this as evidence that emotional adaptability in AI does not have to be purchased with general ability or safety.

Core claim

The paper argues an LLM's social-emotional intelligence — the ability to read and respond to another person's emotional state — can be substantially improved by making personality a context-dependent output rather than a fixed prompt or a single fine-tuning goal. PersonaFuse turns Trait Activation Theory, the psychology finding that situational cues determine which traits get expressed, into a Mixture-of-Experts module: ten LoRA adapters each embody one pole of a Big Five trait, and a small persona-encoder embeds the query and routes to trait-relevant experts. Supervision comes from Persona-CoT, synthetic data where an LLM reasons from social and task cues to a binary trait-activation vector

What carries the argument

Persona-MoE: ten LoRA adapters, each trained on queries calling for one pole of a Big Five trait, gated by a situation-aware router. The router's persona encoder, a 0.5B LLM, maps each query to a persona embedding; cosine similarity to ten learnable expert embeddings, passed through a temperature-scaled softmax, yields the mixture weights. Persona-CoT is the data engine: Falcon3-10B-Instruct reasons from social and task cues to a binary trait-activation vector per query. That vector ties the design together — it partitions training data across the ten experts, supervises the contrastive routing objective, and anchors the joint fine-tuning stage.

Load-bearing premise

The load-bearing premise is that the ten-dimensional trait labels Falcon3-10B-Instruct produces during Persona-CoT correctly say which personality traits a good assistant response should express; the router, the expert partitioning, and all training supervision just learn to reproduce those labels, so if the labels are wrong for a query, the model's 'emotionally intelligent' response is only as good as the teacher's judgment.

What would settle it

Have human judges label which Big Five traits an ideal response should express on a held-out, culturally varied set of dialogues, and compare with Falcon3's generated activation vectors; low agreement on emotion-heavy queries would show the router is copying one teacher model's biases. A quicker probe: present the same underlying need in two surface tones — for example, a crisis described in a cheerful voice — and check whether routing follows the tone or the need.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

Share X Bluesky LinkedIn Reddit HN

If this is right

  • A post-trained 8B model can match or beat much larger proprietary models on human preference in emotion-based dialogue (57.9% win rate vs GPT-4o, 66.7% vs DeepSeek-R1-Distill), while remaining competitive on pure reasoning — emotional adaptability is achievable at small scale.
  • Because the router mixes small adapters over frozen base weights, the approach avoids the catastrophic forgetting that the paper's supervised fine-tuning baselines show on GPQA and SafetyBench; it even reports small gains on those benchmarks.
  • The framework transfers to a 1.7B base model with consistent gains across all benchmark groups, pointing to deployment in resource-constrained settings.
  • Gains carry into downstream human-centered applications: customer service (Shop MMLU) and mental-health counseling (MentalChat16K), including concrete rises in empathy and active-listening scores.
  • Router weights expose which personality traits are activated for a given query, making personality expression inspectable and controllable at inference time instead of being buried in a prompt.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • If the router genuinely learns cue-to-trait mappings, the same architecture could route on other behavioral axes — patience, humor, directness — or fold in explicit user preferences, a step the paper lists as future work.
  • The synthetic trait labels carry the full weight of the method, so a useful stress test is comparing Falcon3's activation vectors with human judgments on a culturally varied query set; large disagreements would mean part of the emotional gain is echoing one teacher model's biases.
  • In counseling or healthcare deployments, router weights could double as an audit trail: a provider could inspect which personality traits the model activated for each interaction, which prompting-based personalization cannot offer.
  • The 'no sacrifice' result is measured against the frozen base model; a natural extension is testing whether the same trained router and adapters keep working when the base model is upgraded without retraining.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. PersonaFuse proposes a post-training framework for LLM social-emotional adaptability. It instantiates Trait Activation Theory and the Big Five model via ten LoRA personality experts (one per trait pole), a Qwen2.5-0.5B-based persona encoder that routes queries by cosine similarity to expert embeddings, and a synthetic training-data pipeline (Persona-CoT) in which Falcon3-10B-Instruct annotates social/task cues and a 10-dimensional trait activation vector p. Training proceeds in three stages: per-expert LoRA warm-up, contrastive router training, and joint end-to-end training. The paper evaluates the resulting Llama-3.1-8B-based model on EmoBench, EQ-Bench, ToMBench, GPQA, GSM8k, Arena-Hard, SafetyBench, Shop MMLU, MentalChat16K, and a human preference study. The headline claim is that PersonaFuse substantially improves social-emotional intelligence 'without sacrificing general reasoning ability or model safety.'

Significance. If the central trade-off claim were established, the paper would make a useful contribution: it operationalizes a psychological theory in a modular MoE architecture, provides a scalable data-generation recipe, and evaluates across a broad set of human-facing benchmarks. The authors are also to be credited for explicitly acknowledging in the conclusion that LLM annotation of traits is unreliable and that the synthetic data are limited in cultural and domain coverage. However, the main empirical claim is currently unsupported by the experiments as reported. The most serious gap is the absence of the unmodified base model in the general-intelligence and safety tables, which makes the 'without sacrificing' assertion untestable and, given the absolute GSM8k score, questionable. The trait-activation supervision also lacks any human validation, so the theoretical grounding of the routing mechanism remains an assumption rather than a demonstrated property.

major comments (3)
  1. [Section 6.3, Tables 14 and 15 (also Table 7)] The paper's central claim is that PersonaFuse improves social-emotional intelligence 'without sacrificing general reasoning ability or model safety.' The only direct evidence for the 'without sacrificing' part is Tables 14 and 15, yet these tables do not include Llama-3.1-8B-Instruct, the base model from which PersonaFuse is built. They compare only against other post-trained variants (Direct-finetuned, Human-Like-finetuned, Random Route). Without the base row, the relative gains over Direct-finetuned show only that PersonaFuse forgets less than a naive fine-tune; they cannot show preservation of the base's capabilities. The concern is concrete: PersonaFuse's GSM8k score is 0.5557, well below the published Llama-3.1-8B-Instruct accuracy of roughly 0.84 on the same benchmark. Even allowing for evaluation-settings differences, a gap of this magnitude cannot be waved away without running th
  2. [Section 4.2, Stage 2, and footnote 3; Section 9 limitations] The entire routing mechanism is supervised by trait activation vectors p generated by Falcon3-10B-Instruct with no human validation. The ten experts are partitioned, the contrastive router labels are defined, and the trait-consistency batches are constructed from these p vectors. The paper's theoretical framing asserts that TAT-driven cues identify the appropriate personality traits, but the only operational link is an unvalidated LLM's output. The limitation paragraph in Section 9 explicitly concedes that 'the annotation accuracy is not always reliable.' This matters because if p is systematically biased, the router learns the generator's opinion about which traits fit which queries, not a ground-truth trait-situation mapping. The two qualitative examples in Table 10 and the embedding classification results in Table 6 are encouraging but do not validate the labels themselves. I ask for
  3. [Section 7, Table 5] The human evaluation is the only place where PersonaFuse is compared with Llama-3.1-8B-Instruct, and it consists of 20 emotion-dialogue and 20 logical-reasoning examples, with a 7-annotator majority vote. The reported 71.9% win rate over Llama-3.1-8B-Instruct on logical reasoning is presented as evidence that the framework preserves reasoning. With only 20 items, this estimate has a very wide confidence interval, and the evaluation criteria (Perceived Usefulness and Social Presence) do not measure objective correctness. This cannot carry the load of the no-degradation claim, nor can it justify the summary statement that PersonaFuse 'maintains' reasoning. If the human evaluation is retained as evidence for the trade-off claim, the authors should report per-item agreement, significance tests, and a larger balanced sample.
minor comments (5)
  1. [Section 4.3 and Appendix A] The loss-weight hyperparameters are inconsistent. Section 4.3 says beta = 1.0 and gamma = 0.2, while Appendix A says 'alpha and beta are 0.5, and gamma is 1.0 in the loss function.' This is critical for reproducibility and should be reconciled.
  2. [Section 7.1] The text says 'Sample responses are provided in Table??' with an unresolved placeholder. Please fix the cross-reference.
  3. [Section 4.3] Equation numbering is out of order: Eq. (8) appears before Eqs. (3) and (5), and Eq. (2) appears after Eq. (3). Renumber the equations sequentially.
  4. [Throughout] The naming is inconsistent: 'Persona-CoT' and 'Persona-COT' are both used; 'Direct-finetuned' and 'Direct-Finetuned' vary across text and tables; 'Human-Like-Finetuned' appears as 'Baseline' at one point in Section 6.1. Please standardize.
  5. [Figure 6] The y-axis labels in Figure 6 are heavily rotated and partially illegible. The figure would benefit from horizontal labels or a legend-based redesign.

Circularity Check

1 steps flagged

Human-eval logical reasoning set overlaps with Persona-CoT training-data source; core architecture otherwise self-contained.

specific steps
  1. fitted input called prediction [Section 5 (Training Data Generation with Persona-CoT) and Section 7.1 (Human Evaluation, Data Source)]
    "To increase the diversity of training queries, we compile data from multiple publicly available sources, including: ... Infinity-Instruct (Zhao et al. 2024a), a synthesized instruction-following dataset. We randomly sample a total of 100,000 queries from these public datasets ... For logical reasoning, we select examples tagged as 'logical reasoning' from the Infinity Instruct dataset (Li et al. 2025a) ... Neither dataset is used in the Persona-CoT data generation process."

    Section 5 lists Infinity-Instruct as one of the sources from which PersonaFuse's 100k Persona-CoT training queries were sampled. Section 7.1 then evaluates logical reasoning on examples tagged 'logical reasoning' from the Infinity Instruct dataset and asserts that neither eval dataset was used in Persona-CoT generation. That assertion contradicts Section 5, so the human-eval logical-reasoning win rates (e.g., 71.9% vs Llama-3.1-8B-Instruct in Table 5) are computed on the same data source used to construct training queries. The model's 'reasonable performance on logical reasoning tasks' is therefore at least partly a re-test of its training distribution, not an independent measure of preserved general reasoning. This is a train/eval overlap rather than an equation-level tautology, and it do

full rationale

The core PersonaFuse derivation is not circular: trait activation vectors p are synthetic intermediate labels produced by Falcon3-10B-Instruct, and the router and experts are trained to match those labels (Eqs. 1, 3, 6, 8, 2), but the main social-emotional and safety claims are benchmarked on external, independently-scored suites (EmoBench, EQ-Bench, ToMBench, SafetyBench) and on GPT-4-judged Arena-Hard/MentalChat. Those external results could in principle falsify the approach, so the LLM-generated label dependence is a data-quality/validity concern, not a self-justifying loop. The paper's own limitation section acknowledges that LLM trait annotation 'is not always reliable.' The strongest circularity-adjacent defect is the human-evaluation logical-reasoning set: Section 7.1 claims the evaluation datasets are not used in Persona-CoT generation, yet Section 5 names Infinity-Instruct as a training-query source. That overlap makes the human-eval logical-reasoning win rate a contaminated supporting result. Separately, the headline 'without sacrificing general reasoning ability' is weakened by the absence of the base Llama-3.1-8B-Instruct row in Tables 14-15, and the absolute GSM8k score (0.5557) is far below the base model's published ~0.84; however, this is a missing-control validity gap, not a reduction of the result to its inputs. Overall, one supporting evaluation reduces to training-distribution overlap; the central architecture and external-benchmark findings retain independent content, so a moderate score is appropriate.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 0 invented entities

The framework introduces no physical entities. It does introduce a 'persona embedding space' and ten expert adapters, but these are learned representations rather than new ontological objects. The main external imports are psychological theories and synthetic labels, and the latter are the weakest link.

free parameters (5)
  • router temperature tau = 1.0
    Set by hand in Section 4.1 with the stated goal of balanced expert mixture; no tuning analysis is shown.
  • contrastive margin m = not specified
    Appears in the contrastive loss (Eq. 3) but no value is reported, leaving an unquantified design choice.
  • loss weights alpha, beta, gamma = inconsistent: beta=1.0, gamma=0.2 in Section 4.3; alpha=0.5, beta=0.5, gamma=1.0 in Appendix A
    The paper gives two conflicting sets of values for the same objectives, so the final configuration is ambiguous.
  • LoRA rank and alpha = rank=8, alpha=16
    Chosen in Section 5 without ablation support.
  • training batch sizes, learning rates, steps = stage-dependent: batch 32/64/32, lr 1e-4/1e-4/1e-5, steps 1000/not stated/300
    Standard hyperparameters selected by hand in Appendix A; no sensitivity analysis.
axioms (4)
  • domain assumption The Big Five model and Trait Activation Theory correctly predict which personality traits are appropriate for a given conversational context (Table 1 correlations).
    The whole framework is built on psychology literature correlations (e.g., therapist needs high agreeableness and low extraversion). If these do not transfer to LLM response styles, the expert design and data generation lose their theoretical grounding.
  • ad hoc to paper LLM-generated trait activation vectors p are accurate enough to supervise both expert specialization and router training.
    Section 4.2 produces p using Falcon3-10B-Instruct with no human verification; the authors themselves concede in the limitations that 'the annotation accuracy is not always reliable.'
  • domain assumption Freezing the base LLM and training only LoRA adapters and the router preserves general knowledge.
    The paper relies on this to justify the claim of no catastrophic forgetting, but the absolute reasoning scores suggest substantial degradation relative to the unmodified base model.
  • standard math Standard deep learning machinery: softmax, cosine similarity, contrastive loss, optimizer convergence.
    These are routine and not in question.

pith-pipeline@v1.3.0-alltime-deepseek · 30026 in / 12519 out tokens · 144407 ms · 2026-08-04T22:18:33.356659+00:00 · methodology

0 comments
Cite this review

Pith. "Pith review of PersonaFuse: A Personality Activation-Driven Framework for Enhancing Human-LLM Interactions." pith.science (2026). https://pith.science/paper/7HNWHK4I

@misc{pith2026250907370,
  author       = {Pith},
  title        = {Pith review of: PersonaFuse: A Personality Activation-Driven Framework for Enhancing Human-LLM Interactions},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/7HNWHK4I}},
  note         = {Machine review of arXiv:2509.07370}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Recent advancements in Large Language Models (LLMs) demonstrate remarkable capabilities across various fields. These developments have led to more direct communication between humans and LLMs in various situations, such as social companionship and psychological support. However, LLMs often exhibit limitations in emotional perception and social competence during real-world conversations. These limitations partly originate from their inability to adapt their communication style and emotional expression to different social and task contexts. In this work, we introduce PersonaFuse, a novel LLM post-training framework that enables LLMs to adapt and express different personalities for varying situations. Inspired by Trait Activation Theory and the Big Five personality model, PersonaFuse employs a Mixture-of-Expert architecture that combines persona adapters with a dynamic routing network, enabling contextual trait expression. Experimental results show that PersonaFuse substantially outperforms baseline models across multiple dimensions of social-emotional intelligence. Importantly, these gains are achieved without sacrificing general reasoning ability or model safety, which remain common limitations of direct prompting and supervised fine-tuning approaches. PersonaFuse also delivers consistent improvements in downstream human-centered applications, such as mental health counseling and review-based customer service. Finally, human preference evaluations against leading LLMs, including GPT-4o and DeepSeek, demonstrate that PersonaFuse achieves competitive response quality despite its comparatively smaller model size. These findings demonstrate that PersonaFuse offers a theoretically grounded and practical approach for developing social-emotional enhanced LLMs, marking a significant advancement toward more human-centric AI systems.

Figures

Figures reproduced from arXiv: 2509.07370 by Ahmed Abbasi, Yixuan Tang, Yi Yang.

Figure 1
Figure 1. Figure 1: Response comparison between GPT-4o and our model. "{Truncate}" indicates truncated [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Our proposed theory driven PersonaFuse framework (a) Persona-MoE, the LLM architec [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: System architecture of PersonaFuse. The framework consists of three main components: [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: The naive response (top) is directly generated by LLM. Our proposed approach Persona [PITH_FULL_IMAGE:figures/full_fig_p012_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Performance improvements across social-emotional intelligence benchmarks over the [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Performance improvements on general intelligence and safety benchmarks over the Direct [PITH_FULL_IMAGE:figures/full_fig_p017_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Performance improvements on practical application tasks in customer service and mental [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Screenshot of the annotation interface used in Human Evaluation. [PITH_FULL_IMAGE:figures/full_fig_p030_8.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

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

  1. Can Large Language Models Revolutionize Survey Research? Experiments with Disaster Preparedness Responses

    cs.AI 2026-05 unverdicted novelty 5.0

    A PMT-constrained LLM framework with A-TLM configuration outperforms classical imputation methods on RMSE and bias for block-wise missing disaster survey data.

Reference graph

Works this paper leans on

90 extracted references · 66 canonical work pages · cited by 1 Pith paper · 1 internal anchor

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution isbn issn journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in "" FUNCTION format.date year ...

  3. [3]

    Information Systems Research 35(2):441--459

    Abbasi A, Parsons J, Pant G, Sheng ORL, Sarker S (2024) Pathways for design research on artificial intelligence. Information Systems Research 35(2):441--459

    2024

  4. [4]

    Information Systems Research 29(3):612--640

    Adamopoulos P, Ghose A, Todri V (2018) The impact of user personality traits on word of mouth: Text-mining social media platforms. Information Systems Research 29(3):612--640

    2018

  5. [5]

    Humanities and Social Sciences Communications 11(1):1--30

    Afroogh S, Akbari A, Malone E, Kargar M, Alambeigi H (2024) Trust in ai: progress, challenges, and future directions. Humanities and Social Sciences Communications 11(1):1--30

    2024

  6. [6]

    Allal LB, Lozhkov A, Bakouch E, Blázquez GM, Penedo G, Tunstall L, Marafioti A, Kydlíček H, Lajarín AP, Srivastav V, Lochner J, Fahlgren C, Nguyen XS, Fourrier C, Burtenshaw B, Larcher H, Zhao H, Zakka C, Morlon M, Raffel C, von Werra L, Wolf T (2025) Smollm2: When smol goes big -- data-centric training of a small language model

    2025

  7. [7]

    (2022) Constitutional ai: Harmlessness from ai feedback

    Bai Y, Kadavath S, Kundu S, Askell A, Kernion J, Jones A, Chen A, Goldie A, Mirhoseini A, McKinnon C, et al. (2022) Constitutional ai: Harmlessness from ai feedback. arXiv preprint arXiv:2212.08073

    Pith/arXiv arXiv 2022

  8. [8]

    Barrick MR, Mount MK, Judge TA (2001) Personality and performance at the beginning of the new millennium: What do we know and where do we go next? International Journal of Selection and assessment 9(1-2):9--30

    2001

  9. [9]

    APL Machine Learning 2(2):026119, ISSN 2770-9019

    Buehler EL, Buehler MJ (2024) X-LoRA: Mixture of low-rank adapter experts, a flexible framework for large language models with applications in protein mechanics and molecular design . APL Machine Learning 2(2):026119, ISSN 2770-9019

    2024

  10. [10]

    Findings of the Association for Computational Linguistics: ACL 2025, 24761--24778 (Association for Computational Linguistics), ISBN 979-8-89176-256-5

    Bui N, Nguyen HT, Kumar S, Theodore J, Qiu W, Nguyen VA, Ying R (2025) Mixture-of-personas language models for population simulation. Findings of the Association for Computational Linguistics: ACL 2025, 24761--24778 (Association for Computational Linguistics), ISBN 979-8-89176-256-5

    2025

  11. [11]

    arXiv preprint arXiv:2501.05032

    C al k EY, Akku s TR (2025) Enhancing human-like responses in large language models. arXiv preprint arXiv:2501.05032

    arXiv 2025

  12. [12]

    Chapman BP, Talbot N, Tatman AW, Britton PC (2009) Personality traits and the working alliance in psychotherapy trainees: An organizing role for the five factor model? Journal of social and clinical psychology 28(5):577--596

    2009

  13. [13]

    Transactions on Machine Learning Research ISSN 2835-8856, survey Certification

    Chen J, Wang X, Xu R, Yuan S, Zhang Y, Shi W, Xie J, Li S, Yang R, Zhu T, Chen A, Li N, Chen L, Hu C, Wu S, Ren S, Fu Z, Xiao Y (2024 a ) From persona to personalization: A survey on role-playing language agents. Transactions on Machine Learning Research ISSN 2835-8856, survey Certification

    work page 2024

  14. [14]

    Findings of the Association for Computational Linguistics: EMNLP 2023, 1170--1183 (Association for Computational Linguistics)

    Chen Y, Xing X, Lin J, Zheng H, Wang Z, Liu Q, Xu X (2023) S oul C hat: Improving LLM s' empathy, listening, and comfort abilities through fine-tuning with multi-turn empathy conversations. Findings of the Association for Computational Linguistics: EMNLP 2023, 1170--1183 (Association for Computational Linguistics)

    work page 2023

  15. [15]

    Management Science 70(12):9101--9117

    Chen Z, Chan J (2024) Large language model in creative work: The role of collaboration modality and user expertise. Management Science 70(12):9101--9117

    work page 2024

  16. [16]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 15959--15983 (Association for Computational Linguistics)

    Chen Z, Wu J, Zhou J, Wen B, Bi G, Jiang G, Cao Y, Hu M, Lai Y, Xiong Z, Huang M (2024 b ) T o MB ench: Benchmarking theory of mind in large language models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 15959--15983 (Association for Computational Linguistics)

    work page 2024

  17. [17]

    arXiv preprint arXiv:2110.14168

    Cobbe K, Kosaraju V, Bavarian M, Chen M, Jun H, Kaiser L, Plappert M, Tworek J, Hilton J, Nakano R, Hesse C, Schulman J (2021) Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

    Pith/arXiv arXiv 2021

  18. [18]

    Findings of the Association for Computational Linguistics: ACL 2025, 6342--6362

    Dan Y, Zhou J, Chen Q, Tian J, He L (2025) P-react: Synthesizing topic-adaptive reactions of personality traits via mixture of specialized lora experts. Findings of the Association for Computational Linguistics: ACL 2025, 6342--6362

    work page 2025

  19. [19]

    MIS quarterly 319--340

    Davis FD (1989) Perceived usefulness, perceived ease of use, and user acceptance of information technology. MIS quarterly 319--340

    work page 1989

  20. [20]

    Information systems research 19(1):93--105

    Devaraj S, Easley RF, Crant JM (2008) Research note—how does personality matter? relating the five-factor model to technology acceptance and use. Information systems research 19(1):93--105

    work page 2008

  21. [21]

    Devlin J, Chang MW, Lee K, Toutanova K (2019) BERT : Pre-training of deep bidirectional transformers for language understanding. Proceedings of the 2019 Conference of the North A merican Chapter of the Association for Computational Linguistics: Human Language Technologies , 4171--4186 (Association for Computational Linguistics)

    work page 2019

  22. [22]

    (2024) The llama 3 herd of models

    Dubey A, Jauhri A, Pandey A, Kadian A, Al-Dahle A, Letman A, Mathur A, Schelten A, Yang A, Fan A, et al. (2024) The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    Pith/arXiv arXiv 2024

  23. [23]

    big five

    Engvik H (1999) Therapeutic popularity and personality: Association between peer therapist nominations and the “big five” personality factors. Scandinavian Journal of Psychology 40(4):261--267

    work page 1999

  24. [24]

    Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 11371--11380 (ELRA and ICCL)

    Feng W, Hao C, Zhang Y, Han Y, Wang H (2024) Mixture-of- L o RA s: An efficient multitask tuning method for large language models. Proceedings of the 2024 Joint International Conference on Computational Linguistics, Language Resources and Evaluation (LREC-COLING 2024), 11371--11380 (ELRA and ICCL)

    work page 2024

  25. [25]

    Journal of research in personality 56:82--92

    Fleeson W, Jayawickreme E (2015) Whole trait theory. Journal of research in personality 56:82--92

    work page 2015

  26. [26]

    The Thirteenth International Conference on Learning Representations

    Gan Z, Liu Y (2025) Towards a theoretical understanding of synthetic data in LLM post-training: A reverse-bottleneck perspective. The Thirteenth International Conference on Learning Representations

    work page 2025

  27. [27]

    Proceedings of the National Academy of Sciences 122(24):e2501660122

    Gao Y, Lee D, Burtch G, Fazelpour S (2025) Take caution in using llms as human surrogates. Proceedings of the National Academy of Sciences 122(24):e2501660122

    work page 2025

  28. [28]

    arXiv preprint arXiv:2406.20094

    Ge T, Chan X, Wang X, Yu D, Mi H, Yu D (2024) Scaling synthetic data creation with 1,000,000,000 personas. arXiv preprint arXiv:2406.20094

    Pith/arXiv arXiv 2024

  29. [29]

    Quality of Life Research 21(3):475--486

    Gershon RC, Lai JS, Bode R, Choi S, Moy C, Bleck T, Miller D, Peterman A, Cella D (2012) Neuro-qol: quality of life item banks for adults with neurological disorders: item development and calibrations based upon clinical and general population testing. Quality of Life Research 21(3):475--486

    work page 2012

  30. [30]

    Proceedings of the National Academy of Sciences 120(30):e2305016120

    Gilardi F, Alizadeh M, Kubli M (2023) Chatgpt outperforms crowd workers for text-annotation tasks. Proceedings of the National Academy of Sciences 120(30):e2305016120

    work page 2023

  31. [31]

    (2025) Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Guo D, Yang D, Zhang H, Song J, Zhang R, Xu R, Zhu Q, Ma S, Wang P, Bi X, et al. (2025) Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948

    Pith/arXiv arXiv 2025

  32. [32]

    Han E, Yin D, Zhang H (2023) Bots with feelings: should ai agents express positive emotion in customer service? Information Systems Research 34(3):1296--1311

    work page 2023

  33. [33]

    (2025) Which economic tasks are performed with ai? evidence from millions of claude conversations

    Handa K, Tamkin A, McCain M, Huang S, Durmus E, Heck S, Mueller J, Hong J, Ritchie S, Belonax T, et al. (2025) Which economic tasks are performed with ai? evidence from millions of claude conversations

    work page 2025

  34. [34]

    Procedia - Social and Behavioral Sciences 30:839--843, ISSN 1877-0428, 2nd World Conference on Psychology, Counselling and Guidance - 2011

    Homayouni A (2011) Personality traits and emotional intelligence as predictors of learning english and math. Procedia - Social and Behavioral Sciences 30:839--843, ISSN 1877-0428, 2nd World Conference on Psychology, Counselling and Guidance - 2011

    work page 2011

  35. [35]

    International Conference on Learning Representations

    Hu EJ, yelong shen, Wallis P, Allen-Zhu Z, Li Y, Wang S, Wang L, Chen W (2022) Lo RA : Low-rank adaptation of large language models. International Conference on Learning Representations

    work page 2022

  36. [36]

    Operations Research

    Huang C, Tang Z, Hu S, Jiang R, Zheng X, Ge D, Wang B, Wang Z (2025) Orlm: A customizable framework in training large models for automated optimization modeling. Operations Research

    work page 2025

  37. [37]

    arXiv preprint arXiv:2507.21919

    Ibrahim L, Hafner FS, Rocher L (2025) Training language models to be warm and empathetic makes them less reliable and more sycophantic. arXiv preprint arXiv:2507.21919

    Pith/arXiv arXiv 2025

  38. [38]

    (2024 a ) Mixtral of experts

    Jiang AQ, Sablayrolles A, Roux A, Mensch A, Savary B, Bamford C, Chaplot DS, Casas Ddl, Hanna EB, Bressand F, et al. (2024 a ) Mixtral of experts. arXiv preprint arXiv:2401.04088

    Pith/arXiv arXiv 2024

  39. [39]

    Findings of the Association for Computational Linguistics: NAACL 2024, 3605--3627 (Association for Computational Linguistics)

    Jiang H, Zhang X, Cao X, Breazeal C, Roy D, Kabbara J (2024 b ) P ersona LLM : Investigating the ability of large language models to express personality traits. Findings of the Association for Computational Linguistics: NAACL 2024, 3605--3627 (Association for Computational Linguistics)

    work page 2024

  40. [40]

    Nature Communications 15(1):9074

    Jin Q, Wang Z, Floudas CS, Chen F, Gong C, Bracken-Clarke D, Xue E, Yang Y, Sun J, Lu Z (2024 a ) Matching patients to clinical trials with large language models. Nature Communications 15(1):9074

    work page 2024

  41. [41]

    The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchmarks Track

    Jin Y, Li Z, Zhang C, Cao T, Gao Y, Jayarao PS, Li M, Liu X, Sarkhel R, Tang X, Wang H, Wang Z, Xu W, Yang J, Yin Q, Li X, Nigam P, Xu Y, Chen K, Yang Q, Jiang M, Yin B (2024 b ) Shopping MMLU : A massive multi-task online shopping benchmark for large language models. The Thirty-eight Conference on Neural Information Processing Systems Datasets and Benchm...

    work page 2024

  42. [42]

    Quality Innovation Prosperity 24(3):90--105

    Jir \'a sek M, Sudzina F (2020) Big five personality traits and creativity. Quality Innovation Prosperity 24(3):90--105

    work page 2020

  43. [43]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 15232--15261 (Association for Computational Linguistics)

    Kang D, Kim S, Kwon T, Moon S, Cho H, Yu Y, Lee D, Yeo J (2024) Can large language models be good emotional supporter? mitigating preference bias on emotional support conversation. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 15232--15261 (Association for Computational Linguistics)

    work page 2024

  44. [44]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, 14581--14602 (Association for Computational Linguistics), ISBN 979-8-89176-251-0

    Kim D, Kang D, Moon T (2025) D o MIX : An efficient framework for exploiting domain knowledge in fine-tuning. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, 14581--14602 (Association for Computational Linguistics), ISBN 979-8-89176-251-0

    work page 2025

  45. [45]

    Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 14397--14413 (Association for Computational Linguistics)

    Kim H, Sclar M, Zhou X, Bras R, Kim G, Choi Y, Sap M (2023) FANT o M : A benchmark for stress-testing machine theory of mind in interactions. Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, 14397--14413 (Association for Computational Linguistics)

    work page 2023

  46. [46]

    (2017) Overcoming catastrophic forgetting in neural networks

    Kirkpatrick J, Pascanu R, Rabinowitz N, Veness J, Desjardins G, Rusu AA, Milan K, Quan J, Ramalho T, Grabska-Barwinska A, et al. (2017) Overcoming catastrophic forgetting in neural networks. Proceedings of the national academy of sciences 114(13):3521--3526

    work page 2017

  47. [47]

    Advances in neural information processing systems 35:22199--22213

    Kojima T, Gu SS, Reid M, Matsuo Y, Iwasawa Y (2022) Large language models are zero-shot reasoners. Advances in neural information processing systems 35:22199--22213

    work page 2022

  48. [48]

    The Twelfth International Conference on Learning Representations

    Kotha S, Springer JM, Raghunathan A (2024) Understanding catastrophic forgetting in language models via implicit inference. The Twelfth International Conference on Learning Representations

    work page 2024

  49. [49]

    Proceedings of the AAAI Conference on Artificial Intelligence 38(16):18417--18425

    Kwon T, Ong KTi, Kang D, Moon S, Lee JR, Hwang D, Sohn B, Sim Y, Lee D, Yeo J (2024) Large language models are clinical reasoners: Reasoning-aware diagnosis framework with prompt-generated rationales. Proceedings of the AAAI Conference on Artificial Intelligence 38(16):18417--18425

    work page 2024

  50. [50]

    Larson S, Mahendran A, Peper JJ, Clarke C, Lee A, Hill P, Kummerfeld JK, Leach K, Laurenzano MA, Tang L, Mars J (2019) An evaluation dataset for intent classification and out-of-scope prediction. Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing ...

    work page 2019

  51. [51]

    (2022) Evaluating human-language model interaction

    Lee M, Srivastava M, Hardy A, Thickstun J, Durmus E, Paranjape A, Gerard-Ursin I, Li XL, Ladhak F, Rong F, et al. (2022) Evaluating human-language model interaction. arXiv preprint arXiv:2212.09746

    Pith/arXiv arXiv 2022

  52. [52]

    Li J, Du L, Zhao H, wen Zhang B, Wang L, Gao B, Liu G, Lin Y (2025 a ) Infinity instruct: Scaling instruction selection and synthesis to enhance language models

    work page 2025

  53. [53]

    Forty-second International Conference on Machine Learning

    Li T, Chiang WL, Frick E, Dunlap L, Wu T, Zhu B, Gonzalez JE, Stoica I (2025 b ) From crowdsourced data to high-quality benchmarks: Arena-hard and benchbuilder pipeline. Forty-second International Conference on Machine Learning

    work page 2025

  54. [54]

    Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, 9373--9385 (Association for Computational Linguistics), ISBN 979-8-89176-251-0

    Liu J, Zhu Y, Wang S, Wei X, Min E, Lu Y, Wang S, Yin D, Dou Z (2025 a ) LLM s + persona-plug = personalized LLM s. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics, 9373--9385 (Association for Computational Linguistics), ISBN 979-8-89176-251-0

    work page 2025

  55. [55]

    Proceedings of the 42nd International Conference on Machine Learning

    Liu R, Geng J, Wu AJ, Sucholutsky I, Lombrozo T, Griffiths TL (2025 b ) Mind your step (by step): Chain-of-thought can reduce performance on tasks where thinking makes humans worse. Proceedings of the 42nd International Conference on Machine Learning

    work page 2025

  56. [56]

    (2025) Alignment and safety in large language models: Safety mechanisms, training paradigms, and emerging challenges

    Lu H, Fang L, Zhang R, Li X, Cai J, Cheng H, Tang L, Liu Z, Sun Z, Wang T, et al. (2025) Alignment and safety in large language models: Safety mechanisms, training paradigms, and emerging challenges. arXiv preprint arXiv:2507.19672

    Pith/arXiv arXiv 2025

  57. [57]

    Journal of personality 60(2):175--215

    McCrae RR, John OP (1992) An introduction to the five-factor model and its applications. Journal of personality 60(2):175--215

    work page 1992

  58. [58]

    Human performance 11(2-3):145--165

    Mount MK, Barrick MR, Stewart GL (1998) Five-factor model of personality and performance in jobs involving interpersonal interactions. Human performance 11(2-3):145--165

    work page 1998

  59. [59]

    European Journal of Personality: Published for the European Association of Personality Psychology 22(4):323--335

    Nettle D, Liddle B (2008) Agreeableness is related to social-cognitive, but not social-perceptual, theory of mind. European Journal of Personality: Published for the European Association of Personality Psychology 22(4):323--335

    work page 2008

  60. [60]

    https://api.openai.com/v1/chat

    OpenAI (2025) Openai (jan 27 version). https://api.openai.com/v1/chat

    work page 2025

  61. [61]

    Ozer DJ, Benet-Martinez V (2006) Personality and the prediction of consequential outcomes. Annu. Rev. Psychol. 57(1):401--421

    work page 2006

  62. [62]

    MIS Quarterly 46(1)

    Padmanabhan B, Fang X, Sahoo N, Burton-Jones A (2022) Machine learning in information systems research. MIS Quarterly 46(1)

    work page 2022

  63. [63]

    arXiv preprint arXiv:2312.06281

    Paech SJ (2023) Eq-bench: An emotional intelligence benchmark for large language models. arXiv preprint arXiv:2312.06281

    Pith/arXiv arXiv 2023

  64. [64]

    The Thirty-eighth Annual Conference on Neural Information Processing Systems

    Poddar S, Wan Y, Ivison H, Gupta A, Jaques N (2024) Personalizing reinforcement learning from human feedback with variational preference learning. The Thirty-eighth Annual Conference on Neural Information Processing Systems

    work page 2024

  65. [65]

    Findings of the Association for Computational Linguistics: EMNLP 2023, 6516--6528 (Association for Computational Linguistics)

    Qian Y, Zhang W, Liu T (2023) Harnessing the power of large language models for empathetic response generation: Empirical investigations and improvements. Findings of the Association for Computational Linguistics: EMNLP 2023, 6516--6528 (Association for Computational Linguistics)

    work page 2023

  66. [66]

    Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 5370--5381 (Association for Computational Linguistics)

    Rashkin H, Smith EM, Li M, Boureau YL (2019) Towards empathetic open-domain conversation models: A new benchmark and dataset. Proceedings of the 57th Annual Meeting of the Association for Computational Linguistics, 5370--5381 (Association for Computational Linguistics)

    work page 2019

  67. [67]

    First Conference on Language Modeling

    Rein D, Hou BL, Stickland AC, Petty J, Pang RY, Dirani J, Michael J, Bowman SR (2024) GPQA : A graduate-level google-proof q&a benchmark. First Conference on Language Modeling

    work page 2024

  68. [68]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 5986--6004 (Association for Computational Linguistics)

    Sabour S, Liu S, Zhang Z, Liu J, Zhou J, Sunaryo A, Lee T, Mihalcea R, Huang M (2024) E mo B ench: Evaluating the emotional intelligence of large language models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 5986--6004 (Association for Computational Linguistics)

    work page 2024

  69. [69]

    Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, 3687--3697 (Association for Computational Linguistics)

    Saravia E, Liu HCT, Huang YH, Wu J, Chen YS (2018) CARER : Contextualized affect representations for emotion recognition. Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing, 3687--3697 (Association for Computational Linguistics)

    work page 2018

  70. [70]

    humanizing

    Schanke S, Burtch G, Ray G (2021) Estimating the impact of “humanizing” customer service chatbots. Information Systems Research 32(3):736--751

    work page 2021

  71. [71]

    The Twelfth International Conference on Learning Representations

    Sclar M, Choi Y, Tsvetkov Y, Suhr A (2024) Quantifying language models' sensitivity to spurious features in prompt design or: How i learned to start worrying about prompt formatting. The Twelfth International Conference on Learning Representations

    work page 2024

  72. [72]

    Proceedings of the 1st Workshop on Personalization of Generative AI Systems (PERSONALIZE 2024), 83--87 (Association for Computational Linguistics)

    Sorokovikova A, Rezagholi S, Fedorova N, Yamshchikov IP (2024) LLM s simulate big5 personality traits: Further evidence. Proceedings of the 1st Workshop on Personalization of Generative AI Systems (PERSONALIZE 2024), 83--87 (Association for Computational Linguistics)

    work page 2024

  73. [73]

    Team FL (2024) The falcon 3 family of open models

    work page 2024

  74. [74]

    Journal of Applied psychology 88(3):500

    Tett RP, Burnett DD (2003) A personality trait-based interactionist model of job performance. Journal of Applied psychology 88(3):500

    work page 2003

  75. [75]

    Advances in Neural Information Processing Systems 37:34737--34774

    Toshniwal S, Moshkov I, Narenthiran S, Gitman D, Jia F, Gitman I (2024) Openmathinstruct-1: A 1.8 million math instruction tuning dataset. Advances in Neural Information Processing Systems 37:34737--34774

    work page 2024

  76. [76]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 3588--3612 (Association for Computational Linguistics)

    Wang J, Mo F, Ma W, Sun P, Zhang M, Nie JY (2024 a ) A user-centric multi-intent benchmark for evaluating large language models. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 3588--3612 (Association for Computational Linguistics)

    work page 2024

  77. [77]

    Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 11897--11916 (Association for Computational Linguistics)

    Wang L, Yang N, Huang X, Yang L, Majumder R, Wei F (2024 b ) Improving text embeddings with large language models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics, 11897--11916 (Association for Computational Linguistics)

    work page 2024

  78. [78]

    Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 19519--19529 (Association for Computational Linguistics)

    Wang Y, Wang M, Manzoor MA, Liu F, Georgiev GN, Das RJ, Nakov P (2024 c ) Factuality of large language models: A survey. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 19519--19529 (Association for Computational Linguistics)

    work page 2024

  79. [79]

    (2022) Chain-of-thought prompting elicits reasoning in large language models

    Wei J, Wang X, Schuurmans D, Bosma M, Xia F, Chi E, Le QV, Zhou D, et al. (2022) Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems 35:24824--24837

    2022

  80. [80]

    The leadership quarterly

    Wong CS, Law KS (2002) Wong and law emotional intelligence scale. The leadership quarterly

    work page 2002

Showing first 80 references.