arxiv: 2605.13858 · v1 · submitted 2026-04-13 · 💻 cs.NE · cs.CL· cs.LG

Recognition: 2 theorem links

· Lean Theorem

A Hormone-inspired Emotion Layer for Transformer language models (HELT)

Eslam Reda , Sara El-Metwally

Authors on Pith no claims yet

Pith reviewed 2026-05-15 07:28 UTC · model grok-4.3

classification 💻 cs.NE cs.CLcs.LG

keywords hormone-inspired emotion modelingtransformer augmentationcontinuous emotion representationaffective computingT5 modelper-hormone attention headsemotional response generationmulti-objective training

0 comments

The pith

A hormone-emotion block added to T5 computes six continuous hormone values from specialized attention heads to modulate hidden states for more appropriate emotional responses.

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

The paper introduces HormoneT5 by inserting a Hormone Emotion Block into transformer encoders to simulate the continuous, multi-dimensional nature of human emotional states through six hormone-like signals. These signals arise from per-hormone attention heads that use orthogonal queries, temperature scaling, and deep projections, then combine into an embedding that directly adjusts the model's hidden representations before decoding. The approach replaces discrete emotion labels with a biologically motivated modulation mechanism trained via a multi-objective loss that includes sequence generation, hormone prediction with margins, and diversity regularization. If the modulation works as intended, generated text should better match the emotional tone of inputs in ways standard transformers cannot achieve.

Core claim

The central claim is that six continuous hormone-like values, each produced by its own attention head with orthogonally initialized queries and temperature-scaled softmax, can be projected into an emotional embedding that modulates encoder hidden states; when trained with the combined loss, the resulting HormoneT5 model reaches over 85 percent per-hormone accuracy within a 0.15 tolerance on a curated emotion-labeled dataset and produces responses that human raters judge significantly more emotionally appropriate than baseline T5 outputs.

What carries the argument

The Hormone Emotion Block, which runs six parallel per-hormone attention heads on the encoder outputs, derives continuous hormone values, and injects their projected embedding to scale or shift the hidden states before they reach the decoder.

If this is right

HormoneT5 reaches 85 percent or higher accuracy on each of the six hormone predictions within a 0.15 tolerance threshold.
Hormone differentiation ranges exceed 0.85 across all six hormones when inputs carry contrasting emotional tones.
Human raters express a statistically significant preference for HormoneT5 responses in emotional appropriateness and empathetic quality.
The added diversity regularization term prevents attention collapse and maintains separation among the six hormone channels.

Where Pith is reading between the lines

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

The same block could be attached to decoder layers or used across multiple dialogue turns to enforce longer-term emotional consistency.
If the hormone values can be aligned with physiological signals, the architecture might serve as a testbed for comparing model behavior against measured human endocrine responses.
Replacing the fixed six hormones with a learned number of channels could reveal whether a small fixed set is sufficient or whether more dimensions improve performance on complex affective tasks.

Load-bearing premise

That the six hormone-like values extracted by the attention heads meaningfully simulate aspects of human emotional processing and that modulating the hidden states with their embedding produces measurably better emotional responses.

What would settle it

A blind A/B test on identical emotional prompts where human raters show no statistically significant preference for HormoneT5 outputs over standard T5 outputs, or where the per-hormone predictions fail to separate contrasting emotional tones by at least 0.85.

Figures

Figures reproduced from arXiv: 2605.13858 by Eslam Reda, Sara El-Metwally.

**Figure 2.** Figure 2: HormoneT5 architecture overview showing the Hormone Emotion Block inserted between the T5 encoder and [PITH_FULL_IMAGE:figures/full_fig_p007_2.png] view at source ↗

**Figure 3.** Figure 3: Training dynamics over 50 epochs showing loss curves, per-hormone accuracy, and differentiation range [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗

**Figure 4.** Figure 4: Hormone activations by emotional tone showing clear differentiation between contrasting emotions. [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗

**Figure 5.** Figure 5: Qualitative examples showing HormoneT5 responses with corresponding hormone activations for each emotional tone. 8 Ablation Studies and Analysis To rigorously understand the contribution of each architectural component in HormoneT5, we conducted systematic ablation studies. Our baseline full model established a strong performance ceiling with an average hormone Mean Squared Error (MSE) of 0.027, a predicti… view at source ↗

read the original abstract

Large Language Models have demonstrated remarkable capabilities in generating contextually relevant and grammatically correct text. However, they fundamentally lack the ability to process and respond to emotional context in a manner analogous to human emotional cognition. Current approaches to emotion modeling in NLP systems rely primarily on discrete emotion classification or simplistic sentiment analysis, which fail to capture the continuous, multi-dimensional nature of human emotional states. In this paper, we introduce HormoneT5, a novel architecture that augments transformer language models with a biologically-inspired Hormone Emotion Block that simulates the human endocrine system's role in emotional processing. Our approach computes six continuous hormone-like values through specialized per-hormone attention heads, each with orthogonally initialized learnable queries, temperature-scaled attention mechanisms, and deep output projections. These hormone values are then transformed into an emotional embedding that modulates the encoder hidden states, enabling emotionally-appropriate response generation. We propose a multi-objective training framework combining sequence-to-sequence loss, hormone prediction loss with margin penalties, and diversity regularization to prevent attention collapse. Experimental results on our curated emotion-labeled dataset demonstrate that HormoneT5 achieves 85%+ per-hormone accuracy within a 0.15 tolerance threshold, with hormone differentiation ranges exceeding 0.85 across all six hormones between contrasting emotional tones. Human evaluation studies show significant preference (p < 0.01) for HormoneT5-generated responses in terms of emotional appropriateness and empathetic quality compared to baseline T5 outputs. Our work opens new directions for biologically-grounded affective computing and emotionally intelligent conversational agents.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

The paper introduces a concrete per-hormone attention block for continuous modulation in T5 but its accuracy numbers rest on author-defined labels that track the same emotional tones used to build the dataset.

read the letter

The main takeaway is that this work gives a specific recipe for injecting six continuous hormone-like values into a transformer decoder via dedicated attention heads, but the experimental support for those values actually doing emotional work is thin. The architecture itself is the clearest new piece: orthogonal query initialization per hormone, temperature scaling inside the attention, deep projections to produce the values, and then an embedding that modulates encoder states. The training objective mixes standard seq2seq loss with a margin-penalized hormone prediction term plus diversity regularization to avoid collapse. That combination is not just another discrete emotion classifier, so the design choice is worth noting for anyone building affective layers. The reported 85% per-hormone accuracy within 0.15 tolerance and differentiation ranges above 0.85 on contrasting tones are the headline numbers, and the human preference study reaches p<0.01. Those claims are stated plainly. The soft spot is that the ground-truth hormone targets appear to be rule-based assignments derived from the same tonal categories used to label the data. Without independent psychological validation or inter-annotator agreement reported, the high accuracy can be explained by the model simply learning the authors' mapping rather than discovering anything about human emotional processing. Dataset size, exact baseline numbers, error bars, and how the six targets were set are not visible in the abstract, which makes the numbers hard to interpret. The scope is also narrow—one base model and one curated set—so the result stays incremental even if the architecture holds up. Readers working on bio-inspired or continuous-state affective models would get the most from the block design and loss formulation. The paper shows clear thinking about how to structure the modulation and avoid trivial solutions, so it is coherent on its own terms. I would send it to peer review so referees can check the label construction and demand fuller baselines and ablations.

Referee Report

3 major / 2 minor

Summary. The paper proposes HormoneT5, a T5-based architecture augmented with a Hormone Emotion Block that uses six per-hormone attention heads (with orthogonal query initialization, temperature scaling, and deep projections) to compute continuous hormone-like values. These values are converted to an emotional embedding that modulates encoder hidden states. Training uses a multi-objective loss combining seq2seq, hormone prediction with margin penalties, and diversity regularization. On a curated emotion-labeled dataset the model reportedly achieves 85%+ per-hormone accuracy within a 0.15 tolerance and hormone differentiation ranges >0.85 between contrasting tones; a human study finds significant preference (p<0.01) for HormoneT5 outputs over baseline T5 on emotional appropriateness.

Significance. If the hormone values can be shown to be independently meaningful rather than author-defined targets, the approach would offer a novel continuous, multi-dimensional mechanism for affective modulation in transformers. The multi-objective training and per-hormone attention design are technically interesting, but the current experimental support is too thin to establish either biological grounding or practical superiority.

major comments (3)

[Abstract / §4] Abstract and §4 (experimental results): the headline claims of 85%+ per-hormone accuracy within 0.15 tolerance and differentiation ranges >0.85 are measured against ground-truth hormone labels on the authors' curated dataset. No description is given of how these six continuous targets were assigned (e.g., rule-based mapping from discrete emotion categories), inter-annotator reliability, or external psychological validation. Without such grounding the metric reduces to reproduction of the authors' own labeling scheme rather than evidence of simulated emotional processing.
[Abstract / §4] Abstract and §4: the reported numerical results supply neither dataset size, baseline quantitative numbers (standard T5 or other affective models), error bars, nor the precise definition of the 0.15 tolerance threshold. These omissions make the 85%+ accuracy and p<0.01 human-preference claims unverifiable and prevent assessment of effect size or statistical robustness.
[§3] §3 (architecture): the assertion that the six hormone-like values 'simulate the human endocrine system's role in emotional processing' is load-bearing for the paper's framing yet rests on an unvalidated mapping. The manuscript should either supply references to established psychological or neuroscientific models that justify the chosen hormone set and their continuous ranges, or reframe the contribution as an engineering heuristic rather than a biologically grounded simulation.

minor comments (2)

[§4] The human-evaluation protocol (number of raters, exact questions, blinding procedure, and response sampling) is not described; this should be added to allow replication of the p<0.01 result.
[§3] Notation for the per-hormone attention heads and the modulation operation (how the emotional embedding is added or multiplied into hidden states) is introduced without an explicit equation; a single diagram or equation would improve clarity.

Simulated Author's Rebuttal

3 responses · 1 unresolved

We thank the referee for the detailed and constructive report. The comments highlight important gaps in clarity and framing that we will address in revision. Below we respond point by point to the three major comments.

read point-by-point responses

Referee: [Abstract / §4] Abstract and §4 (experimental results): the headline claims of 85%+ per-hormone accuracy within 0.15 tolerance and differentiation ranges >0.85 are measured against ground-truth hormone labels on the authors' curated dataset. No description is given of how these six continuous targets were assigned (e.g., rule-based mapping from discrete emotion categories), inter-annotator reliability, or external psychological validation. Without such grounding the metric reduces to reproduction of the authors' own labeling scheme rather than evidence of simulated emotional processing.

Authors: We agree that the current manuscript does not adequately describe the label-generation process. In the revised version we will add a dedicated subsection in §4 that specifies the exact rule-based mapping from the source discrete emotion categories to the six continuous [0,1] hormone targets, including the functional forms and chosen ranges. We will also report inter-annotator agreement statistics obtained during dataset curation. A full external psychological or neuroscientific validation of the mapping was not performed in this study; we will explicitly state this limitation and list it as future work. revision: partial
Referee: [Abstract / §4] Abstract and §4: the reported numerical results supply neither dataset size, baseline quantitative numbers (standard T5 or other affective models), error bars, nor the precise definition of the 0.15 tolerance threshold. These omissions make the 85%+ accuracy and p<0.01 human-preference claims unverifiable and prevent assessment of effect size or statistical robustness.

Authors: We accept that these details were omitted. The revised manuscript will report the exact training-set size, quantitative results for the unmodified T5 baseline and at least one additional affective model, standard-error bars computed over three independent runs, and the precise definition of the 0.15 tolerance (fraction of predictions lying inside the interval [target−0.15, target+0.15]). The statistical test underlying the p<0.01 human-preference result will also be fully documented. revision: yes
Referee: [§3] §3 (architecture): the assertion that the six hormone-like values 'simulate the human endocrine system's role in emotional processing' is load-bearing for the paper's framing yet rests on an unvalidated mapping. The manuscript should either supply references to established psychological or neuroscientific models that justify the chosen hormone set and their continuous ranges, or reframe the contribution as an engineering heuristic rather than a biologically grounded simulation.

Authors: We agree that the current wording overstates biological fidelity. In the revision we will replace all instances of “simulate” with “inspired by” and will reframe the Hormone Emotion Block explicitly as an engineering heuristic that draws on the continuous, multi-dimensional character of endocrine signaling. We will add citations to established dimensional affect models (e.g., valence-arousal and related continuous representations) and will include a short discussion clarifying that the six-hormone choice is a pragmatic design decision rather than a claim of neuroscientific equivalence. revision: yes

standing simulated objections not resolved

External psychological or neuroscientific validation of the specific six-hormone mapping and chosen continuous ranges

Circularity Check

1 steps flagged

Hormone accuracy and differentiation claims reduce to reproducing author-defined labels on the curated dataset

specific steps

fitted input called prediction [Abstract (experimental results paragraph)]
"Experimental results on our curated emotion-labeled dataset demonstrate that HormoneT5 achieves 85%+ per-hormone accuracy within a 0.15 tolerance threshold, with hormone differentiation ranges exceeding 0.85 across all six hormones between contrasting emotional tones."

The per-hormone accuracy and differentiation ranges are computed against continuous hormone targets that the authors assign to the dataset via rule-based mapping from the same emotional-tone categories used to curate the data. The model is trained with a hormone prediction loss against these targets; therefore the reported 85%+ accuracy and >0.85 ranges are the direct result of fitting the authors' own label definitions rather than an independent test of emotional simulation.

full rationale

The central experimental result (85%+ per-hormone accuracy within 0.15 tolerance and differentiation ranges >0.85) is measured against ground-truth hormone values assigned by the authors to the emotion-labeled dataset. Because those targets are constructed from the same tonal/emotional categories used to build the data, the reported metrics demonstrate reproduction of the authors' label mapping rather than independent evidence that the per-hormone attention heads simulate human emotional processing. No external validation or inter-annotator agreement for the continuous hormone targets is provided, so the success metric is statistically forced by the training objective and label construction. The human preference study inherits the same dependency. This matches the fitted-input-called-prediction pattern but does not collapse the entire architecture to a tautology; the modulation mechanism itself remains a non-circular design choice.

Axiom & Free-Parameter Ledger

3 free parameters · 2 axioms · 1 invented entities

The central claim depends on the untested premise that a biological hormone analogy produces useful continuous emotion signals and that internal modulation improves output quality; several design choices are introduced without external grounding.

free parameters (3)

number of hormones
Fixed at six to represent multi-dimensional emotion; chosen by the authors.
tolerance threshold
Set at 0.15 for accuracy calculation; post-hoc choice affecting reported performance.
temperature scaling factor
Used in attention; value not stated and therefore fitted or tuned.

axioms (2)

domain assumption Human emotional states can be usefully approximated by six continuous hormone-like scalars
Invoked to justify the per-hormone attention design.
domain assumption Adding an emotional embedding derived from these scalars to encoder states produces more appropriate responses
Core mechanistic assumption of the modulation step.

invented entities (1)

Hormone Emotion Block no independent evidence
purpose: Compute and apply six hormone values to modulate transformer states
New module introduced by the paper with no prior independent validation

pith-pipeline@v0.9.0 · 5577 in / 1615 out tokens · 48453 ms · 2026-05-15T07:28:11.699961+00:00 · methodology

discussion (0)

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

computes six continuous hormone-like values through specialized per-hormone attention heads, each with orthogonally initialized learnable queries, temperature-scaled attention mechanisms, and deep output projections... ˜H=H⊙(1+α·e expanded)
IndisputableMonolith/Foundation/AlphaCoordinateFixation.lean J_uniquely_calibrated_via_higher_derivative unclear

?

unclear
Relation between the paper passage and the cited Recognition theorem.

multi-objective training framework combining sequence-to-sequence loss, hormone prediction loss with margin penalties, and diversity regularization

What do these tags mean?

matches: The paper's claim is directly supported by a theorem in the formal canon.
supports: The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends: The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses: The paper appears to rely on the theorem as machinery.
contradicts: The paper's claim conflicts with a theorem or certificate in the canon.
unclear: Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

82 extracted references · 82 canonical work pages · 7 internal anchors

[1]

New approach in quantification of emotional intensity from the speech signal: emotional temperature.Expert Systems with Applications, 42(24):9554– 9564, 2015

Jesús B Alonso, Josué Cabrera, Manuel Medina, and Carlos M Travieso. New approach in quantification of emotional intensity from the speech signal: emotional temperature.Expert Systems with Applications, 42(24):9554– 9564, 2015

work page 2015
[2]

A wide evaluation of chatgpt on affective computing tasks.IEEE Transactions on Affective Computing, 15(4):2204–2212, 2024

Mostafa M Amin, Rui Mao, Erik Cambria, and Björn W Schuller. A wide evaluation of chatgpt on affective computing tasks.IEEE Transactions on Affective Computing, 15(4):2204–2212, 2024

work page 2024
[3]

The molecular basis of love.International Journal of Molecular Sciences, 26(4):1533, 2025

Jaroslava Babková and Gabriela Repiská. The molecular basis of love.International Journal of Molecular Sciences, 26(4):1533, 2025

work page 2025
[4]

Physiology of emotion

Rituparna Barooah. Physiology of emotion. InApplication of Biomedical Engineering in Neuroscience, pages 415–435. Springer, 2019

work page 2019
[5]

Isabel Barradas, Zartasha Naeem Khan, and Angelika Peer. Emotion recognition from peripheral physiological signals: A systematic review of trends, challenges and opportunities.ACM Transactions on Interactive Intelligent Systems, 16(1):1–41, 2026

work page 2026
[6]

PhD thesis, Rensselaer Polytechnic Institute, 2025

Ankita Bhaumik.Towards Emotional Reasoning by Dialogue Agents. PhD thesis, Rensselaer Polytechnic Institute, 2025

work page 2025
[7]

Emotional memory: A dimensional analysis

Margaret M Bradley. Emotional memory: A dimensional analysis. InEmotions, pages 97–134. Psychology Press, 2014

work page 2014
[8]

Simon and Schuster, 2015

Loretta Graziano Breuning.Habits of a happy brain: retrain your brain to boost your serotonin, dopamine, oxytocin, & endorphin levels. Simon and Schuster, 2015

work page 2015
[9]

Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Nee- lakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners.Advances in neural information processing systems, 33:1877–1901, 2020

work page 1901
[10]

Emobank: Studying the impact of annotation perspective and representation format on dimensional emotion analysis

Sven Buechel and Udo Hahn. Emobank: Studying the impact of annotation perspective and representation format on dimensional emotion analysis. InProceedings of the 15th Conference of the European Chapter of the Association for Computational Linguistics: Volume 2, Short Papers, pages 578–585, 2017

work page 2017
[11]

Biochemistry of hormones that influences feelings.Annals of Pharmacovigilance & Drug, 1, 2019

Monica Butnariu and Ioan Sarac. Biochemistry of hormones that influences feelings.Annals of Pharmacovigilance & Drug, 1, 2019

work page 2019
[12]

Generating Long Sequences with Sparse Transformers

Rewon Child. Generating long sequences with sparse transformers.arXiv preprint arXiv:1904.10509, 2019

work page internal anchor Pith review Pith/arXiv arXiv 1904
[13]

Emosphere++: Emotion-controllable zero-shot text-to-speech via emotion-adaptive spherical vector.IEEE Transactions on Affective Computing, 2025

Deok-Hyeon Cho, Hyung-Seok Oh, Seung-Bin Kim, and Seong-Whan Lee. Emosphere++: Emotion-controllable zero-shot text-to-speech via emotion-adaptive spherical vector.IEEE Transactions on Affective Computing, 2025

work page 2025
[14]

Plug and play language models: A simple approach to controlled text generation.arXiv preprint arXiv:1912.02164, 2019

Sumanth Dathathri, Andrea Madotto, Janice Lan, Jane Hung, Eric Frank, Piero Molino, Jason Yosinski, and Rosanne Liu. Plug and play language models: A simple approach to controlled text generation.arXiv preprint arXiv:1912.02164, 2019

work page arXiv 1912
[15]

Bert: Pre-training of deep bidirectional transformers for language understanding

Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. Bert: Pre-training of deep bidirectional transformers for language understanding. InProceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers), pages 4171–4186, 2019. 16

work page 2019
[16]

Happiness & health: the biological factors- systematic review article.Iranian journal of public health, 43(11):1468, 2014

Dariush Dfarhud, Maryam Malmir, and Mohammad Khanahmadi. Happiness & health: the biological factors- systematic review article.Iranian journal of public health, 43(11):1468, 2014

work page 2014
[17]

Hierarchical token prepending: Enhancing information flow in decoder-based llm embeddings.arXiv preprint arXiv:2511.14868, 2025

Xueying Ding, Xingyue Huang, Mingxuan Ju, Liam Collins, Yozen Liu, Leman Akoglu, Neil Shah, and Tong Zhao. Hierarchical token prepending: Enhancing information flow in decoder-based llm embeddings.arXiv preprint arXiv:2511.14868, 2025

work page arXiv 2025
[18]

Biological connection to the feeling of happiness.Journal of Clinical & Diagnostic Research, 14(10), 2020

JOSMITHA DSOUZA, Anirban Chakraborty, and Jacintha Veigas. Biological connection to the feeling of happiness.Journal of Clinical & Diagnostic Research, 14(10), 2020

work page 2020
[19]

The fear primacy hypothesis in the structure of emotional states: a systematic literature review

Andrei Efremov. The fear primacy hypothesis in the structure of emotional states: a systematic literature review. Psychological Reports, page 00332941241313106, 2025

work page 2025
[20]

An argument for basic emotions.Cognition & emotion, 6(3-4):169–200, 1992

Paul Ekman. An argument for basic emotions.Cognition & emotion, 6(3-4):169–200, 1992

work page 1992
[21]

Don’t get too excited-eliciting emotions in llms

Gino Franco Fazzi, Julie Skoven Hinge, Stefan Heinrich, and Paolo Burelli. Don’t get too excited-eliciting emotions in llms. InProceedings of the Extended Abstracts of the CHI Conference on Human Factors in Computing Systems, pages 1–9, 2025

work page 2025
[22]

Using millions of emoji occurrences to learn any-domain representations for detecting sentiment, emotion and sarcasm.arXiv preprint arXiv:1708.00524, 2017

Bjarke Felbo, Alan Mislove, Anders Søgaard, Iyad Rahwan, and Sune Lehmann. Using millions of emoji occurrences to learn any-domain representations for detecting sentiment, emotion and sarcasm.arXiv preprint arXiv:1708.00524, 2017

work page arXiv 2017
[23]

Token prepending: A training-free approach for eliciting better sentence embeddings from llms

Yuchen Fu, Zifeng Cheng, Zhiwei Jiang, Zhonghui Wang, Yafeng Yin, Zhengliang Li, and Qing Gu. Token prepending: A training-free approach for eliciting better sentence embeddings from llms. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 3168–3181, 2025

work page 2025
[24]

Biochemical foundations of emotion regulation: Implications for pharmacological and psychological interventions-a narrative review.OBM Neurobiology, 9(4):1–17, 2025

Marina M Gerges and Vladimir Kosonogov. Biochemical foundations of emotion regulation: Implications for pharmacological and psychological interventions-a narrative review.OBM Neurobiology, 9(4):1–17, 2025

work page 2025
[25]

Interoceptive mechanisms and emotional processing.Annual Review of Psychology, 76, 2025

Benedict M Greenwood and Sarah N Garfinkel. Interoceptive mechanisms and emotional processing.Annual Review of Psychology, 76, 2025

work page 2025
[26]

Sentiment analysis with llms and slms

Venkata Gunnu, Shubham Shah, Anvesh Minukuri, and Jayanth Gopu. Sentiment analysis with llms and slms. In Practical Solutions for Modern NLP Challenges: Mastering LLMs and SLMs for Real-World NLP in Cloud and Open-Source, pages 175–224. Springer, 2026

work page 2026
[27]

Endorphins: the basis of pleasure?Journal of neurology, neurosurgery, and psychiatry, 55(4):247, 1992

CH Hawkes. Endorphins: the basis of pleasure?Journal of neurology, neurosurgery, and psychiatry, 55(4):247, 1992

work page 1992
[28]

The endocrine system: an overview.Alcohol health and research world, 22(3):153, 1998

Susanne Hiller-Sturmhöfel and Andrzej Bartke. The endocrine system: an overview.Alcohol health and research world, 22(3):153, 1998

work page 1998
[29]

Parameter-efficient transfer learning for nlp

Neil Houlsby, Andrei Giurgiu, Stanislaw Jastrzebski, Bruna Morrone, Quentin De Laroussilhe, Andrea Gesmundo, Mona Attariyan, and Sylvain Gelly. Parameter-efficient transfer learning for nlp. InInternational conference on machine learning, pages 2790–2799. PMLR, 2019

work page 2019
[30]

Universal Language Model Fine-tuning for Text Classification

Jeremy Howard and Sebastian Ruder. Universal language model fine-tuning for text classification.arXiv preprint arXiv:1801.06146, 2018

work page internal anchor Pith review Pith/arXiv arXiv 2018
[31]

Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models.ICLR, 1(2):3, 2022

work page 2022
[32]

Recent Advances in Multimodal Affective Computing: An NLP Perspective

Guimin Hu, Yi Xin, Weimin Lyu, Haojian Huang, Chang Sun, Zhihong Zhu, Lin Gui, Ruichu Cai, Erik Cambria, and Hasti Seifi. Recent trends of multimodal affective computing: A survey from nlp perspective.arXiv preprint arXiv:2409.07388, 2024

work page internal anchor Pith review Pith/arXiv arXiv 2024
[33]

Explainable sentiment analysis with deepseek-r1: Performance, efficiency, and few-shot learning.IEEE Intelligent Systems, 2025

Donghao Huang and Zhaoxia Wang. Explainable sentiment analysis with deepseek-r1: Performance, efficiency, and few-shot learning.IEEE Intelligent Systems, 2025

work page 2025
[34]

Evaluating and inducing personality in pre-trained language models.Advances in Neural Information Processing Systems, 36:10622–10643, 2023

Guangyuan Jiang, Manjie Xu, Song-Chun Zhu, Wenjuan Han, Chi Zhang, and Yixin Zhu. Evaluating and inducing personality in pre-trained language models.Advances in Neural Information Processing Systems, 36:10622–10643, 2023

work page 2023
[35]

Personallm: Investigating the ability of large language models to express personality traits

Hang Jiang, Xiajie Zhang, Xubo Cao, Cynthia Breazeal, Deb Roy, and Jad Kabbara. Personallm: Investigating the ability of large language models to express personality traits. InFindings of the association for computational linguistics: NAACL 2024, pages 3605–3627, 2024

work page 2024
[36]

Siyu Jin, Fang Xu, Zihui Yuan, Gengfeng Niu, and Zongkui Zhou. Falling in love with ai virtual agents: the role of physical attractiveness and perceived interactivity in parasocial romantic relationships.Humanities and Social Sciences Communications, 2026. 17

work page 2026
[37]

Transformers are rnns: Fast autoregressive transformers with linear attention

Angelos Katharopoulos, Apoorv Vyas, Nikolaos Pappas, and François Fleuret. Transformers are rnns: Fast autoregressive transformers with linear attention. InInternational conference on machine learning, pages 5156–5165. PMLR, 2020

work page 2020
[38]

Exploring the frontiers of llms in psychological applications: A comprehensive review.Artificial Intelligence Review, 58(10):305, 2025

Luoma Ke, Song Tong, Peng Cheng, and Kaiping Peng. Exploring the frontiers of llms in psychological applications: A comprehensive review.Artificial Intelligence Review, 58(10):305, 2025

work page 2025
[39]

Varshney, Caiming Xiong, and Richard Socher

Nitish Shirish Keskar, Bryan McCann, Lav R Varshney, Caiming Xiong, and Richard Socher. Ctrl: A conditional transformer language model for controllable generation.arXiv preprint arXiv:1909.05858, 2019

work page arXiv 1909
[40]

Hormones and the endocrine system.Cham: Springer International Publishing, 2016

Bernhard Kleine and Winfried G Rossmanith. Hormones and the endocrine system.Cham: Springer International Publishing, 2016

work page 2016
[41]

Understanding emotions through nlp.arXiv preprint, 2025

AA Armanulla Khan GS Uday Kumar. Understanding emotions through nlp.arXiv preprint, 2025

work page 2025
[42]

Prefix-Tuning: Optimizing Continuous Prompts for Generation

Xiang Lisa Li and Percy Liang. Prefix-tuning: Optimizing continuous prompts for generation.arXiv preprint arXiv:2101.00190, 2021

work page internal anchor Pith review Pith/arXiv arXiv 2021
[43]

Towards emotional support dialog systems

Siyang Liu, Chujie Zheng, Orianna Demasi, Sahand Sabour, Yu Li, Zhou Yu, Yong Jiang, and Minlie Huang. Towards emotional support dialog systems. InProceedings of the 59th annual meeting of the association for computational linguistics and the 11th international joint conference on natural language processing (volume 1: Long papers), pages 3469–3483, 2021

work page 2021
[44]

RoBERTa: A Robustly Optimized BERT Pretraining Approach

Yinhan Liu, Myle Ott, Naman Goyal, Jingfei Du, Mandar Joshi, Danqi Chen, Omer Levy, Mike Lewis, Luke Zettlemoyer, and Veselin Stoyanov. Roberta: A robustly optimized bert pretraining approach.arXiv preprint arXiv:1907.11692, 2019

work page internal anchor Pith review Pith/arXiv arXiv 1907
[45]

Emollms: A series of emotional large language models and annotation tools for comprehensive affective analysis

Zhiwei Liu, Kailai Yang, Qianqian Xie, Tianlin Zhang, and Sophia Ananiadou. Emollms: A series of emotional large language models and annotation tools for comprehensive affective analysis. InProceedings of the 30th ACM SIGKDD Conference on Knowledge Discovery and Data Mining, pages 5487–5496, 2024

work page 2024
[46]

Developing conversational virtual humans for social emotion elicitation based on large language models.Expert Systems with Applications, 246:123261, 2024

Jose Llanes-Jurado, Lucía Gómez-Zaragozá, Maria Eleonora Minissi, Mariano Alcañiz, and Javier Marín-Morales. Developing conversational virtual humans for social emotion elicitation based on large language models.Expert Systems with Applications, 246:123261, 2024

work page 2024
[47]

A new three-dimensional model for emotions and monoamine neurotransmitters.Medical hypotheses, 78(2):341–348, 2012

Hugo Lövheim. A new three-dimensional model for emotions and monoamine neurotransmitters.Medical hypotheses, 78(2):341–348, 2012

work page 2012
[48]

Networks of spiking neurons: the third generation of neural network models.Neural networks, 10(9):1659–1671, 1997

Wolfgang Maass. Networks of spiking neurons: the third generation of neural network models.Neural networks, 10(9):1659–1671, 1997

work page 1997
[49]

Emotion words in lexical decision and reading aloud

Catherine Mason, Solène Hameau, and Lyndsey Nickels. Emotion words in lexical decision and reading aloud. Language, Cognition and Neuroscience, 40(4):527–546, 2025

work page 2025
[50]

Quantifying valence and arousal in text with multilingual pre-trained transformers

Gonçalo Azevedo Mendes and Bruno Martins. Quantifying valence and arousal in text with multilingual pre-trained transformers. InEuropean Conference on Information Retrieval, pages 84–100. Springer, 2023

work page 2023
[51]

Crowdsourcing a word–emotion association lexicon.Computational intelligence, 29(3):436–465, 2013

Saif M Mohammad and Peter D Turney. Crowdsourcing a word–emotion association lexicon.Computational intelligence, 29(3):436–465, 2013

work page 2013
[52]

Dementia and mci detection based on comprehensive facial expression analysis from videos during conversation.IEEE Journal of Biomedical and Health Informatics, 2025

Taichi Okunishi, Chuheng Zheng, Mondher Bouazizi, Tomoaki Ohtsuki, Momoko Kitazawa, Toshiro Horigome, and Taishiro Kishimoto. Dementia and mci detection based on comprehensive facial expression analysis from videos during conversation.IEEE Journal of Biomedical and Health Informatics, 2025

work page 2025
[53]

Generative agents: Interactive simulacra of human behavior

Joon Sung Park, Joseph O’Brien, Carrie Jun Cai, Meredith Ringel Morris, Percy Liang, and Michael S Bernstein. Generative agents: Interactive simulacra of human behavior. InProceedings of the 36th annual acm symposium on user interface software and technology, pages 1–22, 2023

work page 2023
[54]

Affective computing: Recent advances, challenges, and future trends.Intelligent Computing, 3:0076, 2024

Guanxiong Pei, Haiying Li, Yandi Lu, Yanlei Wang, Shizhen Hua, and Taihao Li. Affective computing: Recent advances, challenges, and future trends.Intelligent Computing, 3:0076, 2024

work page 2024
[55]

Deep contextualized word representations

Matthew E Peters, Mark Neumann, Mohit Iyyer, Matt Gardner, Christopher Clark, Kenton Lee, and Luke Zettlemoyer. Deep contextualized word representations. arxiv 2018.arXiv preprint arXiv:1802.05365, 12, 2018

work page internal anchor Pith review Pith/arXiv arXiv 2018
[56]

Adapterfusion: Non-destructive task composition for transfer learning

Jonas Pfeiffer, Aishwarya Kamath, Andreas Rücklé, Kyunghyun Cho, and Iryna Gurevych. Adapterfusion: Non-destructive task composition for transfer learning. InProceedings of the 16th conference of the European chapter of the association for computational linguistics: main volume, pages 487–503, 2021

work page 2021
[57]

MIT press, 2000

Rosalind W Picard.Affective computing. MIT press, 2000

work page 2000
[58]

Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research, 21(140):1–67, 2020

Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Sharan Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of transfer learning with a unified text-to-text transformer.Journal of machine learning research, 21(140):1–67, 2020. 18

work page 2020
[59]

Towards empathetic open-domain conversation models: A new benchmark and dataset

Hannah Rashkin, Eric Michael Smith, Margaret Li, and Y-Lan Boureau. Towards empathetic open-domain conversation models: A new benchmark and dataset. InProceedings of the 57th annual meeting of the association for computational linguistics, pages 5370–5381, 2019

work page 2019
[60]

A systematic review of the role of oxytocin, cortisol, and testosterone in facial emotional processing.Biology, 10(12):1334, 2021

Ángel Romero-Martínez, Carolina Sarrate-Costa, and Luis Moya-Albiol. A systematic review of the role of oxytocin, cortisol, and testosterone in facial emotional processing.Biology, 10(12):1334, 2021

work page 2021
[61]

A circumplex model of affect.Journal of personality and social psychology, 39(6):1161, 1980

James A Russell. A circumplex model of affect.Journal of personality and social psychology, 39(6):1161, 1980

work page 1980
[62]

PhD thesis, Dublin, National College of Ireland, 2025

Sreelakshmi Sajikumar.Emotion Detection in Text: A Comprehensive Analysis Using Classical, Deep Learning, and Transformer-Based Models. PhD thesis, Dublin, National College of Ireland, 2025

work page 2025
[63]

Fast Transformer Decoding: One Write-Head is All You Need

Noam Shazeer. Fast transformer decoding: One write-head is all you need.arXiv preprint arXiv:1911.02150, 2019

work page internal anchor Pith review Pith/arXiv arXiv 1911
[64]

Recursive deep models for semantic compositionality over a sentiment treebank

Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D Manning, Andrew Y Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. InProceedings of the 2013 conference on empirical methods in natural language processing, pages 1631–1642, 2013

work page 2013
[65]

Large language models for subjective language understanding: A survey.arXiv preprint arXiv:2508.07959, 2025

Changhao Song, Yazhou Zhang, Hui Gao, Ben Yao, and Peng Zhang. Large language models for subjective language understanding: A survey.arXiv preprint arXiv:2508.07959, 2025

work page arXiv 2025
[66]

Evolving neural networks through augmenting topologies.Evolution- ary computation, 10(2):99–127, 2002

Kenneth O Stanley and Risto Miikkulainen. Evolving neural networks through augmenting topologies.Evolution- ary computation, 10(2):99–127, 2002

work page 2002
[67]

Learning to identify emotions in text

Carlo Strapparava and Rada Mihalcea. Learning to identify emotions in text. InProceedings of the 2008 ACM symposium on Applied computing, pages 1556–1560, 2008

work page 2008
[68]

Empathetic large language models, the social capacities and human flourishing.Inquiry, pages 1–23, 2025

Leora Urim Sung and Avigail Ferdman. Empathetic large language models, the social capacities and human flourishing.Inquiry, pages 1–23, 2025

work page 2025
[69]

Emotionality vs

Irina N Trofimova and Anastasia A Gaykalova. Emotionality vs. other biobehavioural traits: a look at neurochemi- cal biomarkers for their differentiation.Frontiers in Psychology, 12:781631, 2021

work page 2021
[70]

Implicit aspect extraction in sentiment analysis: Review, taxonomy, oppportunities, and open challenges.Information Processing & Management, 54(4):545–563, 2018

Mohammad Tubishat, Norisma Idris, and Mohammad AM Abushariah. Implicit aspect extraction in sentiment analysis: Review, taxonomy, oppportunities, and open challenges.Information Processing & Management, 54(4):545–563, 2018

work page 2018
[71]

The advantages of lexicon-based sentiment analysis in an age of machine learning.PloS one, 20(1):e0313092, 2025

A Maurits van der Veen and Erik Bleich. The advantages of lexicon-based sentiment analysis in an age of machine learning.PloS one, 20(1):e0313092, 2025

work page 2025
[72]

Attention is all you need.Advances in neural information processing systems, 30, 2017

Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.Advances in neural information processing systems, 30, 2017

work page 2017
[73]

Learning emotion category representation to detect emotion relations across languages.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

Xiangyu Wang and Chengqing Zong. Learning emotion category representation to detect emotion relations across languages.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2025

work page 2025
[74]

Hormones and emotion.Handbook of cognition and emotion, 69, 2013

Michelle M Wirth, Allison E Gaffey, and HDT Work. Hormones and emotion.Handbook of cognition and emotion, 69, 2013

work page 2013
[75]

A comprehensive review of multimodal emotion recognition: Techniques, challenges, and future directions.Biomimetics, 10(7):418, 2025

You Wu, Qingwei Mi, and Tianhan Gao. A comprehensive review of multimodal emotion recognition: Techniques, challenges, and future directions.Biomimetics, 10(7):418, 2025

work page 2025
[76]

Exploring the application boundaries of llms in mental health: A systematic scoping review.Frontiers in Psychology, 16:1715306, 2026

Jinhua Yang, Ting Liu, Yiming Taclis Luo, Tianyue Niu, Patrick Pang, Ao Xiang, and Qin Yang. Exploring the application boundaries of llms in mental health: A systematic scoping review.Frontiers in Psychology, 16:1715306, 2026

work page 2026
[77]

Emotion recognition based on multimodal physiological signals using spiking feed-forward neural networks.Biomedical Signal Processing and Control, 91:105921, 2024

Xudong Yang, Hongli Yan, Anguo Zhang, Pan Xu, Sio Hang Pan, Mang I Vai, and Yueming Gao. Emotion recognition based on multimodal physiological signals using spiking feed-forward neural networks.Biomedical Signal Processing and Control, 91:105921, 2024

work page 2024
[78]

Hormonal underpinnings of emotional regulation: Bridging endocrinology and psychology.The Journal of Neurobehavioral Sciences, 11(2):60–75, 2024

Eda Yılmazer. Hormonal underpinnings of emotional regulation: Bridging endocrinology and psychology.The Journal of Neurobehavioral Sciences, 11(2):60–75, 2024

work page 2024
[79]

A survey of controllable text generation using transformer-based pre-trained language models.ACM Computing Surveys, 56(3):1–37, 2023

Hanqing Zhang, Haolin Song, Shaoyu Li, Ming Zhou, and Dawei Song. A survey of controllable text generation using transformer-based pre-trained language models.ACM Computing Surveys, 56(3):1–37, 2023

work page 2023
[80]

Instruction tuning for large language models: A survey.ACM Computing Surveys, 58(7):1–36, 2026

Shengyu Zhang, Linfeng Dong, Xiaoya Li, Sen Zhang, Xiaofei Sun, Shuhe Wang, Jiwei Li, Runyi Hu, Tianwei Zhang, Guoyin Wang, et al. Instruction tuning for large language models: A survey.ACM Computing Surveys, 58(7):1–36, 2026

work page 2026

Showing first 80 references.