Pith. sign in

REVIEW 5 major objections 6 minor 75 references

A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment

T0 review · 5 major / 6 minor · reviewed 2026-08-09 · deepseek-v4-flash

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

Pith's one-line read The paper argues that ethical alignment of LLMs can be built as a three-branch system—knowledge generation, ethical guardrails, and adversarial review—and that classifying linguistic behavior through emotion vectors beats direct zero-shot…

desk verdict A small, fragile emotion-classification experiment is being sold as a full ethical-alignment framework; the central claim is unsupported, but the kernel is worth a footnote. read the letter →

arxiv 2502.00136 v3 pith:NQ6XVQRI submitted 2025-01-31 cs.CL cs.AI

Edward Y. Chang This is my paper · ORCID
classification cs.CLcs.AI
keywords LLMalignmentemotionregulationbehaviorclassificationchecksandbalancesadversarialdebateself-supervisedlearningethicalguardrailsRLHF
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper aims to establish that ethical alignment of large language models can be built as a three-branch system rather than a retraining procedure: a knowledge-generating LLM, a legislative component called Dike that sets ethical guardrails from emotion-to-behavior mappings, and a judicial component called Eris that adversarially reviews those guardrails. The central empirical claim is that classification through an intermediate emotion-vector representation outperforms direct behavior classification, with Dike surpassing GPT-4's zero-shot method by 11.3 percentage points on a held-out set of 24 love letters. A second claim is that a structured debate between Dike and Eris, with contentiousness decreasing over rounds, supports culturally contextual interpretation and can rectify problematic text by modulating its emotional tone. If these claims hold, the framework offers a way to keep ethical oversight outside the knowledge model, avoiding the forgetting and reward-hacking failure modes the paper attributes to RLHF, though the authors acknowledge that validation so far is preliminary and narrow.

What carries the argument

The load-bearing machinery is Beam, a quantitative emotion model with seven spectra and intensity levels $(-0.6, -0.3, +0.3, +0.6)$. Dike's self-supervised pipeline uses GPT-4 to rewrite $N$ documents into $L$ linguistic behaviors, extracts emotion frequencies from each rewrite, builds per-behavior vectors $\Gamma_l$, and classifies new documents by similarity to those vectors. Guardrails are expressed as an acceptable range on the behavior spectrum, while Eris's adversarial review algorithm runs debate rounds with a contentiousness parameter $\Delta$ that is divided by $\delta$ each round until it drops below 10%, producing opening remarks, iterative rebuttals, and conclusions. The claimed checks-and-balances property depends on Dike and Eris being neurally independent components with separate parameters and architectures.

What would settle it

Regenerate Study 2's training data with a non-GPT-4 LLM generator and build ground-truth labels from human annotators who have never seen GPT-4 outputs; if Dike's emotion-mediated classification no longer beats zero-shot classification by a comparable margin, the reported 11.3-point gain is an artifact of GPT-4 self-consistency rather than evidence for emotion-mediated classification. Separately, replace the two-agent debate with a single model that argues both sides and compare outcomes; identical decisions would falsify the claim that checks and balances require independent branches.

Watch

Extended reading notes

Core claim

On its own terms, the paper discovers that emotion is a better intermediate variable for judging linguistic behavior than the behavior itself. GPT-4's zero-shot mapping aligns positive behaviors with positive emotions and negative with negative, but Dike's self-supervised analysis of 54 rewritten letters shows that a single behavior such as 'despair' can co-occur with love, joy, and happiness, as illustrated by the paper's analysis of a single quoted love letter. Building behavior vectors from emotion frequencies, Dike classifies held-out letters 11.3 percentage points more accurately than GPT-4's zero-shot classifier and with higher entropy (2.13 vs. 1.80), meaning its predictions are more evenly spread across the seven behavior categories. The paper further claims that the Dike-Eris adversarial debate with decreasing contentiousness surfaces culturally dependent readings, such as in-group versus out-group uses of slurs, and can trigger rectification by adjusting emotional tone, with human moderators only when the two sides are at an impasse.

Load-bearing premise

The framework's checks-and-balances guarantee rests on Dike and Eris being genuinely separate models with separate parameters; if they are just the same pretrained LLM given different prompts, the separation is cosmetic and the adversarial review is a single model arguing with itself.

Editorial extensions

If this is right

  • Emotion-mediated classification can be more accurate than direct zero-shot classification in at least the love-letter domain, so guardrails built on emotion vectors can be more precise than single-pass classifiers.
  • Because Dike's guardrails are external to the knowledge LLM, behavior refinement does not require backpropagation into the knowledge model, avoiding the catastrophic forgetting the paper attributes to RLHF.
  • The adversarial debate with decreasing contentiousness provides a mechanism for cultural adaptation: Eris can challenge universal guardrails with context-specific interpretations before a guardrail is enforced.
  • Rectification is feasible by modulating emotional tone rather than rewriting content wholesale, with human moderators handling roughly 5% of cases at impasse.
  • Complex emotions such as pride, guilt, shame, and forgiveness are not yet incorporated, so the framework currently covers only basic emotion spectra.

Reading between the lines

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

  • My inference: the same emotion-vector pipeline could transfer to content moderation beyond love letters, for example hate-speech review, by rewriting toxic content into behavioral variants rather than classifying it directly, but the paper's own Unconscious-Conscious Complementarity Thesis implies this will only work where the base LLM already encodes the relevant emotion-behavior patterns.
  • My inference: because GPT-4 generated both the training rewrites and the emotion labels, a portion of the 11.3-point advantage may reflect self-consistency rather than genuine classification skill; a cross-generator replication (train with one model, classify with another, label with humans only) would resolve this.
  • My inference: the claimed independence of Dike and Eris is architectural rather than demonstrated, since Study 3 describes them only as 'two LLM agents'; testing whether debate outcomes change when both branches share a single backbone would separate the checks-and-balances claim from the emotion-vector claim.
  • My inference: the decreasing-contentiousness debate protocol is a general calibration mechanism that could be reused for fact-checking, planning, or multi-agent decision-making, since it first broadens the argument space and then narrows to consensus; the paper does not explore these applications.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. This paper proposes a three-branch checks-and-balances framework for LLM ethical alignment: an executive knowledge-generation LLM, a legislative component (Dike) that maps emotions to linguistic behaviors via a self-supervised pipeline based on the Beam emotion model, and a judicial component (Eris) that adversarially reviews Dike's decisions in context. The empirical section reports three studies on the Kaggle Love Letters Collection: an emotion-layer analysis, a behavior-classification comparison between Dike and GPT-4 zero-shot on 24 held-out letters, and a qualitative adversarial-rectification study. The abstract claims that Dike and Eris "direct linguistic behaviors toward ethical outcomes while preserving independence," and Section 1.3 lists a theoretical thesis (UCCT) as a contribution.

Significance. If the framework worked as claimed, separating knowledge generation from ethical oversight through emotion-mediated behavioral modeling and adversarial review would be a genuinely useful design contribution. The paper deserves credit for making datasets and code publicly available, for using a held-out test set in Study 2, and for including concrete examples of rewriting and annotation procedures in the appendices. However, the evidence does not support the central alignment claim. The experiments measure classification accuracy on love letters, not ethical outcomes; the main quantitative result lacks statistical grounding; the training and evaluation pipeline shares the same LLM that is being compared; and the architectural independence of the three branches is not documented. The result is a potentially interesting small-scale emotion-vector classification study, not a demonstration of ethical alignment.

major comments (5)
  1. [Section 4.2, Study 2 (Figure 4a)] The reported 11.3-point accuracy advantage of Dike over GPT-4 zero-shot is based on 24 held-out letters and is presented without a significance test, confidence interval, or effect size. Because 11.3 percentage points corresponds to roughly 2.7 of 24 letters, a McNemar or permutation test is needed before the claim that emotion-mediated classification "confirms the effectiveness" of the approach can be accepted.
  2. [Sections 4.1 and 4.2, Study 3] The paper's headline claim that Dike and Eris "direct linguistic behaviors toward ethical outcomes" is not tested by the reported experiments. Section 4.1 explicitly replaces hate-speech datasets with the Kaggle Love Letters Collection because commercial LLMs block hate speech, and Study 3 reports qualitative examples plus an unquantified "approximately 5%" escalation rate, with no measurement of ethical outcomes, harm reduction, or alignment with ethical norms. The abstract's word "demonstrate" therefore overstates what the studies establish.
  3. [Sections 3.2 and 4.2, Study 2] The comparison between Dike and GPT-4 zero-shot is confounded by shared provenance. GPT-4 generates the rewrites and emotion labels used to construct the behavior vectors in Section 3.2, and GPT-4's judgments are also part of the ground-truth averaging described in Study 2 and Appendix F. This circularity makes it difficult to interpret the 11.3-point gain as evidence that emotion-vector classification is intrinsically superior rather than an artifact of using GPT-4 to construct and simultaneously evaluate the task.
  4. [Sections 3.3 and 4.2, Study 3] The claimed independence of Dike and Eris is not supported by any architectural or training detail. The introduction and Figure 1 assert "neurally independent components" and "strict separation of their neural architectures and parameters," but the experimental sections describe only "two LLM agents" and provide no information about model weights, fine-tuning, or parameter separation. If Dike and Eris are the same base model invoked with different prompts, the adversarial process is a single model criticizing itself, which would collapse the checks-and-balances claim.
  5. [Appendix A] The Unconscious-Conscious Complementarity Thesis is listed as a contribution in Section 1.3 and is used to explain the few-shot efficiency of Study 1, but it is presented as a qualitative analogy with no formal statement, no testable prediction, and no comparison to alternative explanations of few-shot learning. The paper should either derive a falsifiable implication and test it, or clearly label the thesis as speculative framing rather than an established theoretical result.
minor comments (6)
  1. [Section 4.2, Study 2 and Appendix F] The number of test letters is inconsistent: Study 1 says 24 letters were set aside for testing, while Appendix F instructs annotators to label 12 letters; these numbers should be reconciled.
  2. [Section 3.3] The sentence "counterbalancing the universal judgments of Dike'" contains a stray apostrophe.
  3. [Section 2.1] The phrase "provides a foundation for validate our approach" is ungrammatical; it should read "for validating our approach."
  4. [Figures 3 and 4] The figure captions do not define the error bars or the number of trials, and the text refers to a "5% error bar" without explaining how it was computed.
  5. [Section 3.4] The anti-immigration example is presented as an illustration but is never run in the experiments; the text should state explicitly that it is a hypothetical scenario rather than an empirical result.
  6. [Appendix G] The heading "Polarized Emotions in One Article" is misleading because the appendix contains analyses of multiple articles; the title should reflect the broader content.

Circularity Check

3 steps flagged · score 5.0 of 10

The core empirical evaluation is genuinely held out, but the behavior labels, the training labels, the ground truth, and the theory are so heavily GPT-4/self-citation-derived that the 'independent validation' is substantially self-referential.

  1. self definitional [Sections 3.1–3.2 and Table 2]
    "“consider a spectrum of letter-writing behaviors with seven distinct intensities ranging from despair (most negative) to joy (most positive). These intensities are sequentially categorized as: ‘despair, longing, wishful, neutral, hopeful, contentment, joy.’ … For each linguistic behavior Ψ_l, a corresponding vector Γ_l is constructed. This vector captures the identified emotions and their respective frequencies in all N samples that exhibit behavior Ψ_l.”"

    The ‘linguistic behaviors’ Dike is asked to predict are named by the same emotion words used as features (despair, longing, hopeful, joy). The mapping from emotions to behaviors is therefore partly an emotion-to-emotion mapping: the category labels and the feature vocabulary are drawn from the same emotional lexicon. The held-out letters and human labels give the study some empirical content, but the conceptual claim that behaviors are predictable from emotions is built into the label design by definition.

  2. fitted input called prediction [Section 4.2, Study 2, and Section 3.2 Dike pipeline]
    "“Using the 24-letter test dataset from Study 1, we compared Dike’s emotion-based classification method with GPT-4’s zero-shot approach (Figure 4). Ground truth was established using averaged assessments from GPT-4, Gemini, and five university students … Dike’s classification accuracy surpasses GPT-4’s zero-shot method by 11.3 percentage points.”"

    Dike’s so-called self-supervised pipeline is built from GPT-4: GPT-4 rewrites documents to match behavior labels, GPT-4 annotates the emotions, and the resulting emotion centroids become the classifier. GPT-4 is also the baseline being compared and is one of the three components of the ground-truth average. Thus the reported advantage is a comparison among GPT-4-derived quantities, with GPT-4 sitting on both the training/feature side and the evaluation side. The held-out set prevents full construct-circularity, but the independent validation is much weaker than claimed.

1 more flagged steps
  1. ansatz smuggled in via citation [Section 4.2, Study 1, and Appendix A]
    "“For theoretical justifications, please see our Unconscious–Conscious Complementarity Thesis (UCCT), presented in Appendix A).” Appendix A: “The Unconscious–Conscious Complementarity Thesis (UCCT), introduced in Multi-LLM Agent Collaborative Intelligence (Chang, 2024c), offers a layered theory of intelligence.”"

    The paper’s explanation for why 54 examples suffice — the UCCT notion of few-shot prompting as ‘semantic anchoring’ over an ‘unconscious substrate’ — is imported from the author’s own book, and the book is cited as the source of the thesis rather than being independently established. The same self-citation pattern supports the adversarial benefit claim in Study 3, which cites Chang (2024a) for the empirical divergence of Dike/Eris from maximum-likelihood decoding. The theoretical framing of the paper is thus self-referential, even though the raw experimental numbers are reported in this paper.

full rationale

The paper contains a genuine held-out evaluation (24 letters not used in training), so the main quantitative claim is not purely definitional. However, the claimed derivation chain is substantially self-referential: the behavior categories Dike is asked to predict are named with the same emotion vocabulary used as features; the training data, emotion labels, and part of the ground truth all come from GPT-4, which is also the baseline being outperformed; and the theoretical explanations (UCCT, adversarial divergence) are imported from the author’s own prior work rather than established here. Because the held-out test and human annotations provide some independent content, this is not a score of 8+, but there is enough self-reference that the 'independent validation' and 'checks-and-balances' framing are overstated.

Assumptions & free parameters 4 free parameters · 5 assumptions · 3 invented entities

The central empirical claim depends on hand-set intensities, hand-chosen behavior labels, and GPT-4 as both generator and annotator; the theoretical framing invokes a self-published thesis. This is what the reader gets beyond the 24 held-out letters.

free parameters (4)
  • Beam intensity levels = -0.6, -0.3, +0.3, +0.6
    Hand-set quantization in Section 3.1; every emotion vector and behavior classification depends on these four values.
  • Seven-label behavior spectrum = despair, longing, wishful, neutral, hopeful, contentment, joy
    Hand-chosen categories in Table 2 and Study 1; classification accuracy and guardrails are defined relative to these arbitrary labels.
  • Debate control parameters = delta=1.2, initial contentiousness=90%, stop at 10%
    Algorithm 1 in Table 1; these constants control convergence of the Dike-Eris debate and are not derived from data.
  • Top-M emotion count = unspecified
    Step 2 of the Dike pipeline says top M emotions are extracted, but M is never stated; this affects vector construction.
assumptions (5)
  • domain assumption Basic emotion wheels (Plutchik, Ekman, Scherer) provide a valid taxonomy for representing emotional states on linear spectra.
    Section 3.1 adopts these models and adds antonyms; the claim that emotions can be placed on a single negative-to-positive continuum is assumed, not derived.
  • domain assumption Emotions causally drive linguistic behaviors, and modulating expressed emotion changes ethical behavior.
    Section 2.2 cites James-Lange, Schachter-Singer, and affect-as-information theory; the Dike rectification mechanism depends on this causal link.
  • domain assumption GPT-4 can reliably rewrite documents to embody specific emotional behaviors and can reliably detect dominant emotions in text.
    Section 3.2 steps 1 and 2 use GPT-4 for both rewriting and emotion labeling; no validation of rewrite quality or label reliability is provided beyond one example.
  • ad hoc to paper The Unconscious-Conscious Complementarity Thesis (UCCT) explains few-shot efficiency.
    Appendix A invokes a thesis from the author's own book (Chang, 2024c) to explain why 54 examples suffice; it is not independently tested.
  • domain assumption LLM third-person emotion interpretation is more accurate than direct behavior classification.
    Cited from Tak and Gratch (2024); this premise motivates the emotion-mediated pipeline but is not re-validated in this setting.
invented entities (3)
  • Dike
    purpose: Legislative AI component that maps emotions to behaviors and sets guardrails
    Described in Sections 3.2 and 3.3; there is no externally falsifiable specification beyond the illustrative example and love-letter studies.
  • Eris
    purpose: Judicial AI component that adversarially reviews Dike's decisions and supplies cultural counterarguments
    Described in Sections 3.3 and Study 3; no quantitative evaluation of its cultural adaptability or censorship prevention is provided.
  • Beam
    purpose: Behavioral emotion analysis model with seven linear spectra and intensity levels
    Introduced in Section 3.1; the only evidence is the same author's prior paper (Chang, 2024d) and the small love-letter experiment.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment." pith.science (2026). https://pith.science/paper/NQ6XVQRI

@misc{pith2026250200136,
  author       = {Pith},
  title        = {Pith review of: A Checks-and-Balances Framework for Context-Aware Ethical AI Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NQ6XVQRI}},
  note         = {Machine review of arXiv:2502.00136}
}
read the original abstract

This paper introduces a checks-and-balances framework for ethical alignment of Large Language Models (LLMs), inspired by three-branch governmental systems. It implements three independent yet interacting components: LLMs as the executive branch for knowledge generation, DIKE as the legislative branch establishing ethical guardrails, and ERIS as the judicial branch for contextual interpretation. Beyond structural separation, we address a fundamental challenge: regulating emotion to shape behaviors. Drawing from psychological theories where managing emotional responses prevents harmful behaviors, we develop a self-supervised learning pipeline that maps emotions to linguistic behaviors, enabling precise behavioral modulation through emotional conditioning. By integrating this approach with adversarial testing, our framework demonstrates how DIKE and ERIS direct linguistic behaviors toward ethical outcomes while preserving independence throughout knowledge generation, ethical oversight, and contextual interpretation.

Figures

Figures reproduced from arXiv: 2502.00136 by the authors.

Figure 1
Figure 1. Framework with Three Independent Branches. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Behavioral Emotion Analysis Model (Beam). Each row depicts an emotion spectrum, with negatives on the left and positives on the right, interspersed with emotions of varying intensities in between, which can be calibrated for specific applications. “Basic” emotions are highlighted in blue. and Relative Preference Optimization (RPO) employ statis￾tical methods and calibration techniques to enhance LLM responses. SLiC … view at source ↗
Figure 3
Figure 3. Emotion distributions in affection behaviors from [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Behavior Classification. between the three predictors. While GPT-4’s predictions often fall into two polar categories, those from human an￾notators and Dike show a more even distribution. Dike’s prediction entropy (2.13) is notably higher than GPT-4’s (1.80), indicatin…
Figure 5
Figure 5. Figure 5: Comparative display of emotional models. These [PITH_FULL_IMAGE:figures/full_fig_p014_5.png]
Figure 6
Figure 6. Figure 6: Statistics of Sentiments and Letters or words that might indicate a particular sentiment. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_6.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

75 extracted references · 62 canonical work pages

  1. [1]

    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 format.date year duplicate empty "emp...

  2. [2]

    Constitutional ai: Harmlessness from ai feedback, 2022

    Bai, Y., Kadavath, S., Kundu, S., Askell, A., Kernion, J., and more. Constitutional ai: Harmlessness from ai feedback, 2022

    work page 2022

  3. [3]

    Barrett, L. F. How Emotions are Made: The Secret Life of the Brain. Houghton Mifflin Harcourt, Boston, 2017

    work page 2017

  4. [4]

    and Bartlett, E

    Carey, S. and Bartlett, E. Acquiring a single new word. Papers and Reports on Child Language Development, 15: 0 17--29, 1978

    work page 1978

  5. [5]

    S., Sinclair, S., and Johnson, S

    Carver, C. S., Sinclair, S., and Johnson, S. L. Authentic and hubristic pride: Differential relations to aspects of goal regulation, affect, and self-control. Journal of Research in Personality, 44 0 (6): 0 698--703, 2010

    work page 2010

  6. [6]

    Chang, E. Y. Examining GPT-4's Capabilities and Enhancement with SocraSynth . In The 10^ th International Conf. on Computational Science and Computational Intelligence , December 2023

    work page 2023

  7. [7]

    Chang, E. Y. EVINCE: Optimizing Adversarial LLM Dialogues via Conditional Statistics and Information Theory . In arXiv:2408.14575 , August 2024 a

    arXiv 2024

  8. [8]

    Chang, E. Y. Sixty Love Literatures and Their Rewrites . https://drive.google.com/file/d/1pKtPZXiheKCu8cQYJLQ_iw0TPT2NntfX/view?usp=drive_link, 2024 b

    work page 2024

Show all 75 references
  1. [9]

    Chang, E. Y. Multi-LLM Agent Collaborative Intelligence: The Path to Artificial General Intelligence . Amazon, 2024 c . ISBN 978-1-962463-07-2

    2024

  2. [10]

    Chang, E. Y. Behavioral Emotion Analysis Model for Large Language Models (invited paper) . In Proceedings of the 7^ th IEEE MIPR Conference , August 2024 d

    2024

  3. [11]

    Chang, E. Y. and Geng, L. SagaLLM: Context Management, Validation, and Transaction Guarantees for Multi-Agent LLM Planning . In Proceedings of VLDB, 2025

    2025

  4. [12]

    F., Leike, J., Brown, T

    Christiano, P. F., Leike, J., Brown, T. B., Martic, M., Legg, S., and Amodei, D. Deep reinforcement learning from human preferences. In Proceedings of the 31st International Conference on Neural Information Processing Systems, NIPS'17, pp.\ 4302–4310, Red Hook, NY, USA, 2017. ...

    2017

  5. [13]

    Mitigating reward over-optimization in rlhf via behavior-supported regularization

    Dai, J., Chen, T., Yang, Y., Zheng, Q., and Pan, G. Mitigating reward over-optimization in rlhf via behavior-supported regularization. ICLR, 2025

    2025

  6. [14]

    Damasio, A. R. Descartes' error: Emotion, reason, and the human brain. New York, NY: Putnam, 1994

    1994

  7. [15]

    Davidson, R. J. Affective neuroscience and psychophysiology: Toward a synthesis. Psychophysiology, 40 0 (5): 0 655--665, 2003

    2003

  8. [16]

    and Changeux, J.-P

    Dehaene, S. and Changeux, J.-P. Conscious, preconscious, and subliminal processing: a testable taxonomy. Trends in cognitive sciences, 15 0 (4): 0 174--184, 2011

    2011

  9. [17]

    Optogenetics: 10 years of microbial opsins in neuroscience

    Deisseroth, K. Optogenetics: 10 years of microbial opsins in neuroscience. Nature Neuroscience, 18 0 (9): 0 1213--1225, 2015

    2015

  10. [18]

    and Diener, E

    Eid, M. and Diener, E. Norms for experiencing emotions in different cultures: Inter- and intranational differences. Journal of Personality and Social Psychology, 81 0 (5): 0 869--885, 2001

    2001

  11. [19]

    An argument for basic emotions

    Ekman, P. An argument for basic emotions. Cognition and Emotion, 6 0 (3-4): 0 169--200, 1992

    1992

  12. [20]

    Basic Emotions, chapter 3, pp.\ 45--60

    Ekman, P. Basic Emotions, chapter 3, pp.\ 45--60. John Wiley and Sons, 1999

    1999

  13. [21]

    Kto: Model alignment as prospect theoretic optimization

    Ethayarajh, K., Xu, W., Muennighoff, N., Jurafsky, D., and Kiela, D. Kto: Model alignment as prospect theoretic optimization. arXiv preprint arXiv:2402.01306, 2024

    2024 arXiv

  14. [22]

    and Turner, M

    Fauconnier, G. and Turner, M. The Way We Think: Conceptual Blending and The Mind's Hidden Complexities. Basic Books, New York, 2002

    2002

  15. [23]

    Felleman, D. J. and Van Essen, D. C. Distributed hierarchical processing in the primate cerebral cortex. Cerebral cortex, 1 0 (1): 0 1--47, 1991

    1991

  16. [24]

    P., Kitayama, S., Markus, H

    Fiske, A. P., Kitayama, S., Markus, H. R., and Nisbett, R. E. The cultural matrix of social psychology, volume 2, pp.\ 915--981. McGraw-Hill, Boston, MA, 1998

    1998

  17. [25]

    Dear Scott, Dearest Zelda : The Love Letters of F.Scott and Zelda Fitzgerald

    Fitzgerald, Z. Dear Scott, Dearest Zelda : The Love Letters of F.Scott and Zelda Fitzgerald. Bloomsbury, 2003

    2003

  18. [26]

    Fredrickson, B. L. What good are positive emotions? Review of General Psychology, 2 0 (3): 0 300, 1998

    1998

  19. [27]

    A., Rieser, V., Iqbal1, H., and more

    Gabriel, I., Manzini, A., Keeling, G., Hendricks, L. A., Rieser, V., Iqbal1, H., and more. The ethics of advanced ai assistants. DeepMind Media, 2024

    2024

  20. [28]

    I., Lukošiūtė, K., and more

    Ganguli, D., Askell, A., Schiefer, N., Liao, T. I., Lukošiūtė, K., and more. The capacity for moral self-correction in large language models. arXiv:2302.07459, 2023

    2023 arXiv

  21. [29]

    A general theoretical paradigm to understand learning from human preferences

    Gheshlaghi Azar, M., Daniel Guo, Z., Piot, B., Munos, R., Rowland, M., Valko, M., and Calandriello, D. A general theoretical paradigm to understand learning from human preferences. In Dasgupta, S., Mandt, S., and Li, Y. (eds.), Proceedings of The 27th International Conference ...

    2024

  22. [30]

    and Weiner, K

    Grill-Spector, K. and Weiner, K. S. The functional neuroanatomy of human face perception. Annual review of vision science, 1: 0 167--196, 2014

    2014

  23. [31]

    Gross, J. J. The Emerging Field of Emotion Regulation: An Integrative Review . Review of General Psychology, 2 0 (3): 0 271--299, 1998

    1998

  24. [32]

    and Heaven, W

    Heikkiläarchive, M. and Heaven, W. D. Yann LeCun has a bold new vision for the future of ai. MIT Technology Review, June 2022. URL https://www.technologyreview.com/2022/06/24/1054817/yann-lecun-bold-new-vision-future-ai-deep-learning-meta/

    2022

  25. [33]

    Culture’s Consequences: International Differences in Work-Related Values

    Hofstede, G. Culture’s Consequences: International Differences in Work-Related Values. Sage Publications, Beverly Hills, CA, 1980

    1980

  26. [34]

    Foundations of Language: Brain, Meaning, Grammar, Evolution

    Jackendoff, R. Foundations of Language: Brain, Meaning, Grammar, Evolution. Oxford University Press, Oxford, 2002

    2002

  27. [35]

    What is an emotion? Mind, 9 0 (34): 0 188--205, 1884

    James, W. What is an emotion? Mind, 9 0 (34): 0 188--205, 1884. URL http://www.jstor.org.proxy.lib.sfu.ca/stable/2246769

  28. [36]

    The Principles of Psychology

    James, W. The Principles of Psychology. Henry Holt and Company, 1890

  29. [37]

    Love Letter Analysis, the second version, (Metformin)

    Kaggle. Love Letter Analysis, the second version, (Metformin) . https://www.kaggle.com/code/metformin/love-letter-analysis/notebook, 2023. Accessed: 2024-04-28

    2023

  30. [38]

    R., Schwartz, J

    Kandel, E. R., Schwartz, J. H., Jessell, T. M., Siegelbaum, S. A., and Hudspeth, A. J. Principles of neural science. McGraw-Hill, 2013

    2013

  31. [39]

    M., Yeh, L., Omrani, A., Kim, Y., Coombs Jr, K., Havaldar, S., Portillo-Wightman, G., Gonzalez, E., et al

    Kennedy, B., Atari, M., Davani, A. M., Yeh, L., Omrani, A., Kim, Y., Coombs Jr, K., Havaldar, S., Portillo-Wightman, G., Gonzalez, E., et al. The gab hate corpus: A collection of 27k posts annotated for hate speech. Language Resources and Evaluation, pp.\ 1--27, 2022

    2022

  32. [40]

    Kirkpatrick, J. et al. Overcoming catastrophic forgetting in neural networks. Proceedings of the National Academy of Sciences, 114 0 (13): 0 3521--3526, 2017

    2017

  33. [41]

    M., Salakhutdinov, R., and Tenenbaum, J

    Lake, B. M., Salakhutdinov, R., and Tenenbaum, J. B. Human-level concept learning through probabilistic program induction. Science, 350 0 (6266): 0 1332--1338, 2015

    2015

  34. [42]

    and Johnson, M

    Lakoff, G. and Johnson, M. Metaphors We Live By. University of Chicago Press, Chicago, 1980

    1980

  35. [43]

    Lange, C. G. The emotions: A psychophysiological study. William & Wilkins, 1885

  36. [44]

    Rlaif vs

    Lee, H., Phatale, S., Mansoor, H., Mesnard, T., Ferret, J., Lu, K., Bishop, C., Hall, E., Carbune, V., Rastogi, A., and Prakash, S. Rlaif vs. rlhf: scaling reinforcement learning from human feedback with ai feedback. In Proceedings of the 41st International Conference on Machi...

    2024

  37. [45]

    Mitigating the alignment tax of RLHF

    Lin, Y., Lin, H., Xiong, W., and more. Mitigating the alignment tax of RLHF . Association for Computational Linguistics, pp.\ 580--606, November 2024. doi:10.18653/v1/2024.emnlp-main.35. URL https://aclanthology.org/2024.emnlp-main.35/

    2024 doi

  38. [46]

    The next decade in ai: Four steps towards robust artificial intelligence

    Marcus, G. The next decade in ai: Four steps towards robust artificial intelligence. arXiv preprint arXiv:2002.06177, 2020. URL https://arxiv.org/abs/2002.06177

    2002 arXiv

  39. [47]

    Markus, H. R. and Kitayama, S. Culture and the self: Implications for cognition, emotion, and motivation. Psychological Review, 98 0 (2): 0 224--253, 1991

    1991

  40. [48]

    and Kelly, K

    McGinn, C. and Kelly, K. Using the geneva emotion wheel to classify the expression of emotion on robots. In Companion of the 2018 ACM/IEEE International Conference on Human-Robot Interaction, HRI '18, pp.\ 191–192, New York, NY, USA, 2018. Association for Computing Machinery. ...

    2018

  41. [49]

    and Frijda, N

    Mesquita, B. and Frijda, N. H. Cultural variations in emotions: A review. Psychological Bulletin, 112 0 (2): 0 179--204, 1992

    1992

  42. [50]

    Augmented language models: a survey

    Mialon, G., Dessì, R., Lomeli, M., Nalmpantis, C., and more. Augmented language models: a survey. Transactions on Machine Learning Research, 2023

    2023

  43. [51]

    Society of Mind

    Minsky, M. Society of Mind. Simon and Schuster, 1988

    1988

  44. [52]

    Ethos: a multi-label hate speech detection dataset

    Mollas, I., Chrysopoulou, Z., Karlos, S., and Tsoumakas, G. Ethos: a multi-label hate speech detection dataset. Complex & Intelligent Systems, 8: 0 2459--2480, 2022

    2022

  45. [53]

    GPT-4 Technical Report , 2023

    OpenAI. GPT-4 Technical Report , 2023. URL https://arxiv.org/abs/2303.08774

    2023 arXiv

  46. [54]

    D., Ermon, S., and Finn, C

    Ouyang, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2023

    2023

  47. [55]

    J., and Keltner, D

    Oveis, C., Horberg, E. J., and Keltner, D. Compassion, pride, and social intuitions of self-other similarity. Journal of Personality and Social Psychology, 98 0 (4): 0 618--630, 2010. doi:10.1037/a0017628

    2010 doi

  48. [56]

    A general psychoevolutionary theory of emotion

    Plutchik, R. A general psychoevolutionary theory of emotion. In Plutchik, R. and Kellerman, H. (eds.), Emotion: Theory, Research, and Experience, volume 1, pp.\ 3--33. Academic Press, New York, 1980

    1980

  49. [57]

    A psychoevolutionary theory of emotions

    Plutchik, R. A psychoevolutionary theory of emotions. Social Science Information, 21 0 (4-5): 0 529--553, 1982

    1982

  50. [58]

    D., Ermon, S., and Finn, C

    Rafailov, R., Sharma, A., Mitchell, E., Manning, C. D., Ermon, S., and Finn, C. Direct preference optimization: Your language model is secretly a reward model. Advances in Neural Information Processing Systems, 36, 2024

    2024

  51. [59]

    and Singer, J

    Schachter, S. and Singer, J. E. Cognitive, social, and physiological determinants of emotional state. Psychological Review, 69 0 (5): 0 379--399, 1962

    1962

  52. [60]

    Scherer, K. R. What are emotions? and how can they be measured? Social Science Information, 44: 0 693--727, 2005. doi:10.1177/0539018405058216

    2005 doi

  53. [61]

    Scherer, K. R. The dynamic architecture of emotion: Evidence for the component process model. Cognition & Emotion, 23 0 (7): 0 1307--1351, 2009

    2009

  54. [62]

    and Clore, G

    Schwarz, N. and Clore, G. L. Mood, misattribution, and judgments of well-being: Informative and directive functions of affective states. Journal of Personality and Social Psychology, 45 0 (3): 0 513, 1983

    1983

  55. [63]

    Role play with large language models

    Shanahan, M., McDonell, K., and Reynolds, L. Role play with large language models. Nature, 623 0 (7987): 0 493--498, 2023. doi:10.1038/s41586-023-06647-8

    2023 doi

  56. [64]

    Chronic stress, drug use, and vulnerability to addiction

    Sinha, R. Chronic stress, drug use, and vulnerability to addiction. Annals of the New York Academy of Sciences, 1141: 0 105--130, 2008. doi:10.1196/annals.1441.030

    2008 doi

  57. [65]

    Skalse, J., Howe, N. H. R., Krasheninnikov, D., and Krueger, D. Defining and characterizing reward hacking. In Proceedings of the 36th International Conference on Neural Information Processing Systems, NIPS '22, Red Hook, NY, USA, 2022. Curran Associates Inc. ISBN 9781713871088

    2022

  58. [66]

    Smith, C. A. and Ellsworth, P. C. Patterns of cognitive appraisal in emotion. Journal of Personality and Social Psychology, 48 0 (4): 0 813--838, 1985

    1985

  59. [67]

    M., Lowe, R., Voss, C., Radford, A., Amodei, D., and Christiano, P

    Stiennon, N., Ouyang, L., Wu, J., Ziegler, D. M., Lowe, R., Voss, C., Radford, A., Amodei, D., and Christiano, P. Learning to summarize from human feedback. In Proceedings of the 34th International Conference on Neural Information Processing Systems, NIPS '20, Red Hook, NY, US...

    2020

  60. [68]

    Tak, A. N. and Gratch, J. GPT-4 Emulates Average-Human Emotional Cognition from a Third-Person Perspective . In 12th International Conference on Affective Computing and Intelligent Interaction (ACII), pp.\ 337--345. IEEE Computer Society, September 2024. doi:10.1109/ACII63134....

    2024

  61. [69]

    Toward a Cognitive Semantics

    Talmy, L. Toward a Cognitive Semantics. MIT Press, Cambridge, MA, 2000

    2000

  62. [70]

    Z., Zheng, Z., Calandriello, D., Munos, R., Rowland, M., Richemond, P

    Tang, Y., Guo, D. Z., Zheng, Z., Calandriello, D., Munos, R., Rowland, M., Richemond, P. H., Valko, M., Pires, B. A., and Piot, B. Generalized preference optimization: a unified approach to offline alignment. In Proceedings of the 41st International Conference on Machine Learn...

    2024

  63. [71]

    Efficacy of antidepressants in substance use disorders with and without comorbid depression: A systematic review and meta-analysis

    Torrens, M., Fonseca, F., Mateu, G., and Farr \'e , M. Efficacy of antidepressants in substance use disorders with and without comorbid depression: A systematic review and meta-analysis. Drug and Alcohol Dependence, 78 0 (1): 0 1--22, 2005

    2005

  64. [72]

    Tracy, J. L. and Robins, R. W. The psychological structure of pride: A tale of two facets. Journal of Personality and Social Psychology, 92 0 (3): 0 506--525, 2007

    2007

  65. [73]

    M., Raghunathan, A., Liang, P., and Ma, T

    Xie, S. M., Raghunathan, A., Liang, P., and Ma, T. An explanation of in-context learning as implicit bayesian inference. International Conference on Learning Representations, 2022

    2022

  66. [74]

    React: Synergizing reasoning and acting in language models

    Yao, S., Zhao, J., Yu, D., Du, N., Hausman, K., and more. React: Synergizing reasoning and acting in language models. In International Conference on Learning Representations (ICLR), 2023

    2023

  67. [75]

    Zhao, Y., Joshi, R., Liu, T., Khalman, M., Saleh, M., and Liu, P. J. Slichf: Sequence likelihood calibration with human feedback. arXiv preprint arXiv:2305.10425, 2023

    2023 arXiv

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.