LLM Assertiveness can be Mechanistically Decomposed into Emotional and Logical Components

Arush Tagade; Hikaru Tsujimura

REVIEW 3 major objections 2 minor 21 references

LLM Assertiveness can be Mechanistically Decomposed into Emotional and Logical Components

T0 review · 3 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Random voter splits can determine which aggregation rule to trust, without assuming a generative model.

desk verdict Potentially important claim, but the only text I got is the abstract and I can't verify the method; the orthogonality-by-construction worry is real and needs a full-paper look. read the letter →

arxiv 2508.17182 v2 pith:LIKSU3CQ submitted 2025-08-24 cs.LG cs.AIcs.CL

Hikaru Tsujimura , Arush Tagade This is my paper · ORCID

classification cs.LGcs.AIcs.CL MSC 91B1491B1268Q17

keywords rulepickingrulesrankaggregationsocialchoiceconsistencyrandomsplitKendall-TaudistancemaximumlikelihoodestimatorNP-completeness

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 introduces a problem social choice has not directly asked: given many defensible ways to aggregate evaluators' rankings, how should we choose the aggregation rule itself? The proposed answer, Aggregation by Consistency (AbC), is to split the evaluators randomly into two independent panels, run each candidate rule on both halves, and select the rule whose two outputs disagree least under a Kendall-tau distance that penalizes ties. The paper shows AbC satisfies natural axioms, that welfare-maximizing rule pickers fail one of these axioms, and that asking for perfect split consistency over all positional scoring rules is NP-complete. Despite that hardness, a sampling implementation picks the maximum-likelihood rule on synthetic Mallows and Plackett–Luce data, and gives rule-change guidance on peer review, elections, and Formula One.

What carries the argument

The central mechanism is a random two-panel split combined with the expected Kendall-Tau distance with ties: each voter is assigned uniformly at random to one of two groups, every candidate rule is run on both group profiles, and the distance between the two resulting weak rankings (with ties penalized at half weight) measures disagreement. AbC selects the candidate rule minimizing this expected disagreement. The distance function and the split procedure together make consistency a computable, data-driven objective.

What would settle it

Generate profiles from a known noise model such as Plackett-Luce with a fixed ground-truth ranking, include the MLE among the candidate rules, and compute each candidate's exact expected Kendall-Tau split distance. If any non-MLE rule has consistently lower expected split distance than the MLE across repeated random splits, then consistency does not track the best rule, and the paper's central criterion would fail.

Watch

Extended reading notes

Core claim

The central claim is that consistency under random splits is enough to pick among aggregation rules without assuming any generative model. AbC takes a set of candidate social welfare functions, splits the voters uniformly at random into two groups, and chooses the candidate rule with the smallest expected Kendall-Tau distance between the rankings it produces on the two halves. The paper proves AbC is reversal-symmetric and plurality-shuffling consistent, and that any welfare-maximizing rule picker fails the latter. It also proves that deciding whether any positional scoring rule can achieve perfect consistency on a split is NP-complete. Experiments then show that when data is generated from

Load-bearing premise

That a rule's split consistency—how much its output changes when voters are randomly divided—faithfully measures how good the rule is for that profile; if the two halves can disagree for reasons unrelated to rule quality, AbC's selections stop being meaningful.

Editorial extensions

If this is right

Rule choice becomes an empirical, data-driven step: run AbC on your own profile and pick from any candidate set without committing to a generative model or a fixed axiom set.
When data really is generated by Mallows or Plackett–Luce, AbC chooses the maximum likelihood estimator, so it inherits the statistical rationale without needing to assume it.
In real evaluation settings, AbC can diagnose proposed changes: it shows both Formula One scoring-rule changes improved split consistency, while trimmed Borda would hurt ALMA proposal-ranking consistency.
Computing perfect consistency over all positional scoring rules is NP-complete, but the Monte Carlo implementation remains practical in the experiments.
No anonymous rule picking rule can satisfy all three axioms AbC captures, delimiting what any such method can promise.

Reading between the lines

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

Inference: the split-disagreement measure could be adapted beyond rankings—committees, scores, or approval sets—using other distances such as Jaccard dissimilarity, extending AbC to multi-winner selection.
Inference: a testable extension is to run AbC on real domains with independent gold standards (e.g., historical elections with known winners) to check whether 'most consistent' also means 'most accurate'; the synthetic MLE results support this, but real-world evidence is circumstantial.
Inference: if consistency-as-quality holds broadly, AbC gives organizations a principled way to audit aggregation procedures—peer review, contests, or AI feedback aggregation—by re-running splits on past data.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

The headline result – that Llama 3.2 assertiveness decomposes into orthogonal emotional and logical components with distinct steering effects – is exactly the kind of testable mechanistic claim worth taking seriously. If it holds, it would give applied interpretability a concrete handle on overconfidence. Credit where it's due: the abstract is clear, the experimental plan is sensible, and the authors are open about using open models and human-annotated datasets. That is the right raw material.

But I cannot actually review this paper. The full text I was sent is a different manuscript – a social-choice paper on rule-picking by Berker, Armstrong, Conitzer, and Shah (arXiv:2508.17177). That is almost certainly a metadata/pipeline error, not an author problem, but it means I have only the abstract to judge. And from the abstract alone, the load-bearing weakness is exactly what the stress-test note flagged: the abstract never says how the emotional and logical components were extracted. If the decomposition uses PCA, SVD, or any orthogonal projection, then the two components are orthogonal by construction. Calling one cluster 'emotional' and the other 'logical' after the fact is then post hoc labeling, not discovery. The parallel to the Elaboration Likelihood Model is an interpretation, not a mechanistic result, unless the components are validated against independent probes.

The causal steering experiment also looks potentially circular. The abstract says the steering vectors were 'derived from these sub-components' and then tested for 'distinct causal effects.' If those vectors come from the same fine-tuned models and the same activation contrasts that defined the components, the test is in-sample unless there is a held-out split or counterfactual validation. The abstract gives no indication either way.

I want to be careful about proportion. These are concerns about the abstract, not proven flaws in the paper. It is entirely possible that the full manuscript reports an out-of-sample PCA or a non-orthogonal sparse decomposition with independent probes, and then the central claim is plausible. But the abstract as written does not support that.

Who gets value from this? People working on mechanistic interpretability of LLM confidence and on targeted debiasing. It deserves a serious referee – the question is important enough and the claim is sharp enough that a competent reviewer should look under the hood. I would not cite it in its current abstract-only form.

Send it to peer review, but require the authors to disclose the decomposition method, show out-of-sample validation, and release code/data. Then the orthogonality question can be settled rather than guessed at.

Referee Report

3 major / 2 minor

Summary. The abstract describes a mechanistic interpretability study claiming that LLM assertiveness in Llama 3.2 models decomposes into two orthogonal sub-components, labeled emotional and logical, with distinct causal effects when steering vectors derived from these components are applied. The study is positioned as evidence for a multi-component structure of assertiveness, paralleling the Elaboration Likelihood Model. However, the full text supplied for review is not this paper: it is a social-choice manuscript titled 'Designing Rules to Pick a Rule: Aggregation by Consistency' (arXiv:2508.17177), with different authors and content. No methods, experiments, or results from the claimed LLM study are present, making it impossible to evaluate the central claims from the body.

Significance. If the claimed decomposition into separate, steerable emotional and logical assertiveness components is correct and mechanistically validated, it would be a notable contribution to mechanistic interpretability and to understanding overconfidence in LLMs. The use of an open-source model (Llama 3.2), human-annotated assertiveness data, and steering-vector interventions are potentially reproducible and falsifiable, which are strengths in the stated design. However, none of the actual evidence, algorithmic details, or validation checks are visible in the supplied manuscript, so the significance is conditional on content that is currently absent.

major comments (3)

[Full Text (title page)] The manuscript body supplied for review is not the paper described in the abstract. The title, authors, and content correspond to arXiv:2508.17177 ('Designing Rules to Pick a Rule: Aggregation by Consistency'), a social-choice paper. No methods, experimental setup, layer counts, datasets, or steering-vector details for the LLM assertiveness study appear anywhere. Under the instruction to treat all supplied text as the manuscript, this is a load-bearing integrity issue: the abstract's claims are entirely unsupported by the body. The correct full text must be provided before any technical evaluation is possible.
[Abstract] The claim that high-assertive representations 'decompose into two orthogonal sub-components' is unverifiable because the decomposition method is not stated. If the method is PCA, SVD, or any orthogonal projection, orthogonality is true by construction rather than an empirical finding. The 'emotional' and 'logical' labels would then be assigned post hoc to whatever directions the math returns. The paper needs to specify the decomposition algorithm and validate that the components correspond to independent emotional/logical axes, for example via non-orthogonal decompositions (e.g., ICA) or via independent behavioral probes on held-out data not used to define the components.
[Abstract] The causal steering results are reported only qualitatively ('broadly influence,' 'more localized effects') with no effect sizes, confidence intervals, baselines, or out-of-sample evaluation. Since the steering vectors are 'derived from these sub-components' and the components are extracted from the same fine-tuned models, there is a risk of circularity. The authors should clarify whether components and steering vectors are validated on data not used to define them, e.g., across fine-tuning seeds, datasets, or held-out layers.

minor comments (2)

[Abstract] The terms 'emotional' and 'logical' are not operationally defined. The paper should specify how human annotations map to these categories and how cluster labels are assigned.
[Abstract] The phrase 'layers most sensitive to assertiveness contrasts' is vague; the threshold or selection criterion for 'most sensitive' is not given.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation is exhibited; the abstract under-specifies the decomposition method and the supplied full text is a different paper, so no equation-level reduction can be identified.

full rationale

The claimed result in the abstract—that high-assertive representations decompose into two orthogonal emotional/logical components and that steering vectors from them have distinct causal effects—cannot be checked for circularity from the abstract alone because the decomposition method is not stated. The concern that orthogonality might be imposed by PCA/SVD is a plausible method-level worry, but it is speculation about an undisclosed procedure rather than a quoteable reduction. Per the hard rules, circularity must be exhibited by the paper's own equations or by a load-bearing self-citation. No such exhibit exists in the supplied text. The full text provided (arXiv:2508.17177, 'Designing Rules to Pick a Rule') is a different paper and contains no argument about LLM assertiveness, emotional/logical components, or steering vectors. Within that full text, the self-reference to Berker et al. 2025 appears only as a suggested future distance function, and Caiata et al. 2025 is used as an experimental methodology reference, not to establish the main theorem. The central theoretical results are proved with self-contained arguments from stated definitions. Therefore there is no identifiable step where a prediction equals its input by construction, and the appropriate finding is no significant circularity (score 0).

Assumptions & free parameters 4 free parameters · 4 assumptions · 2 invented entities

This ledger is abstract-only because the supplied full text is a different manuscript. The entries below are the standard free choices and assumptions of this method family, inferred from the abstract, not confirmed from the paper. Every entry should be re-audited when the actual full text is available.

free parameters (4)

Layer selection threshold for 'most sensitive' layers
The abstract states layers are selected by similarity metrics as most sensitive to assertiveness contrasts but reports no threshold, layer count, or selection rule; a typical free choice in this method family.
Number of components in the decomposition (2) = 2 (emotional, logical)
The abstract asserts two orthogonal sub-components; whether the number of components was fixed a priori or chosen from data is not stated. Orthogonality may be imposed by the method (e.g., PCA/SVD).
Fine-tuning hyperparameters for Llama 3.2 on assertiveness datasets
The abstract says the models are fine-tuned on human-annotated assertiveness datasets but reports no hyperparameters, data sizes, or train/test splits.
Steering vector coefficients
Steering vectors are 'derived from these sub-components'; coefficients and scales are not reported in the abstract.

assumptions (4)

domain assumption Residual activation geometry in Llama 3.2 encodes assertiveness in directions that similarity metrics can recover
The method of extracting residual activations and computing similarity metrics to localize assertiveness presupposes that assertiveness is represented in the activation space in a recoverable way. Invoked throughout the abstract.
domain assumption Fine-tuned Llama 3.2 models faithfully reflect assertiveness as annotated by humans
The analysis rests on fine-tuning with human-annotated assertiveness datasets; if fine-tuning distorted or did not capture assertiveness, the decomposition would be of an artifact.
domain assumption The Elaboration Likelihood Model dual-route theory is a valid interpretive frame for LLM internals
The abstract claims the two components 'parallel the dual-route Elaboration Likelihood Model in Psychology'; the mapping from psychological theory to network sub-components is assumed, not derived.
ad hoc to paper Orthogonal sub-components are a real property of the model rather than an artifact of the decomposition
The abstract presents orthogonality as a discovered property. If the decomposition method (e.g., PCA/SVD/orthogonalization) guarantees orthogonality by construction, this axiom is violated; the abstract does not state the method.

invented entities (2)

'Emotional' assertiveness sub-component
purpose: One of two orthogonal latent directions claimed to underpin high-assertive behavior, with broad influence on prediction accuracy
From the abstract alone, the entity is defined by the same activation data used to construct it; its causal role is tested only inside the paper. No externally falsifiable handle is visible.
'Logical' assertiveness sub-component
purpose: The second orthogonal latent direction, claimed to exert more localized causal effects than the emotional component
Same as above: the entity is internal to the paper's own fine-tuned models and decomposition; independent evidence is not visible from the abstract.

how reviews work

0 comments

Cite this review

Pith. "Pith review of LLM Assertiveness can be Mechanistically Decomposed into Emotional and Logical Components." pith.science (2026). https://pith.science/paper/LIKSU3CQ

@misc{pith2026250817182,
  author       = {Pith},
  title        = {Pith review of: LLM Assertiveness can be Mechanistically Decomposed into Emotional and Logical Components},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LIKSU3CQ}},
  note         = {Machine review of arXiv:2508.17182}
}

read the original abstract

Large Language Models (LLMs) often display overconfidence, presenting information with unwarranted certainty in high-stakes contexts. We investigate the internal basis of this behavior via mechanistic interpretability. Using open-sourced Llama 3.2 models fine-tuned on human annotated assertiveness datasets, we extract residual activations across all layers, and compute similarity metrics to localize assertive representations. Our analysis identifies layers most sensitive to assertiveness contrasts and reveals that high-assertive representations decompose into two orthogonal sub-components of emotional and logical clusters-paralleling the dual-route Elaboration Likelihood Model in Psychology. Steering vectors derived from these sub-components show distinct causal effects: emotional vectors broadly influence prediction accuracy, while logical vectors exert more localized effects. These findings provide mechanistic evidence for the multi-component structure of LLM assertiveness and highlight avenues for mitigating overconfident behavior.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

21 extracted references · 9 canonical work pages

[1]

Refusal in language models is mediated by a single direction, 2024

Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction, 2024. URL http://arxiv.org/abs/2406.11717

arXiv 2024
[2]

Voelkel, Shane Muldowney, Johannes C

Hui Bai, Jan G. Voelkel, Shane Muldowney, Johannes C. Eichstaedt, and Robb Willer. LLM -generated messages can persuade humans on policy issues. 16 0 (1): 0 6037, 2025. ISSN 2041-1723. doi:10.1038/s41467-025-61345-5. URL https://www.nature.com/articles/s41467-025-61345-5

doi:10.1038/s41467-025-61345-5 2025
[3]

Measuring the persuasiveness of language models, 2024

Esin Durmus, Liane Lovitt, Alex Tamkin, Stuart Ritchie, Jack Clark, and Deep Ganguli. Measuring the persuasiveness of language models, 2024. URL https://www.anthropic.com/news/measuring-model-persuasiveness

work page 2024
[4]

Toy models of superposition

Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition. Transformer Circuits Thread, 2022. https://transformer-circuits.pub/2022/t...

2022
[5]

EPISTEMIC INTEGRITY IN LARGE LANGUAGE MODELS

Bijean Ghafouri, Shahrad Mohammadzadeh, James Zhou, Pratheeksha Nair, Jacob-Junqi Tian, Hikaru Tsujimura, Mayank Goel, Sukanya Krishna, Reihaneh Rabbany, Jean-François Godbout, and Kellin Pelrine. EPISTEMIC INTEGRITY IN LARGE LANGUAGE MODELS . 2024

work page 2024
[6]

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava S...

arXiv 2024
[7]

Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA : Low-rank adaptation of large language models, 2021. URL http://arxiv.org/abs/2106.09685

arXiv 2021
[8]

probable

Tzur M. Karelitz and David V. Budescu. You say "probable" and i say "likely": Improving interpersonal communication with verbal probability phrases. 10 0 (1): 0 25--41, 2004. ISSN 1939-2192, 1076-898X. doi:10.1037/1076-898X.10.1.25. URL https://doi.apa.org/doi/10.1037/1076-898X.10.1.25

doi:10.1037/1076-898x.10.1.25 2004

Show all 21 references

[9]

The SFU opinion and comments corpus: A corpus for the analysis of online news comments

Varada Kolhatkar, Hanhan Wu, Luca Cavasso, Emilie Francis, Kavan Shukla, and Maite Taboada. The SFU opinion and comments corpus: A corpus for the analysis of online news comments. 4 0 (2): 0 155--190, 2020. ISSN 2509-9507, 2509-9515. doi:10.1007/s41701-019-00065-w. URL http://...

2020 doi
[10]

Jinqi Lai, Wensheng Gan, Jiayang Wu, Zhenlian Qi, and Philip S. Yu. Large language models in law: A survey. 5: 0 181--196, 2024. ISSN 26666510. doi:10.1016/j.aiopen.2024.09.002. URL https://linkinghub.elsevier.com/retrieve/pii/S2666651024000172

2024 doi
[11]

Taming overconfidence in LLMs : Reward calibration in RLHF , 2025

Jixuan Leng, Chengsong Huang, Banghua Zhu, and Jiaxin Huang. Taming overconfidence in LLMs : Reward calibration in RLHF , 2025. URL http://arxiv.org/abs/2410.09724

2025 arXiv
[12]

The role of assertiveness in a storytelling game with persuasive robotic non-player characters

Raul Paradeda, Maria José Ferreira, Raquel Oliveira, Carlos Martinho, and Ana Paiva. The role of assertiveness in a storytelling game with persuasive robotic non-player characters. In Proceedings of the Annual Symposium on Computer-Human Interaction in Play, pages 453--465. AC...

2019
[13]

Measuring sentence-level and aspect-level (un)certainty in science communications, 2021

Jiaxin Pei and David Jurgens. Measuring sentence-level and aspect-level (un)certainty in science communications, 2021. URL http://arxiv.org/abs/2109.14776

2021 arXiv
[14]

Personal involvement as a determinant of argument-based persuasion

Richard E Petty, John T Cacioppo, and Rachel Goldman. Personal involvement as a determinant of argument-based persuasion. 1981

1981
[15]

Cognitive reframing of negative thoughts through human-language model interaction

Ashish Sharma, Kevin Rushton, Inna Lin, David Wadden, Khendra Lucas, Adam Miner, Theresa Nguyen, and Tim Althoff. Cognitive reframing of negative thoughts through human-language model interaction. In Proceedings of the 61st Annual Meeting of the Association for Computational L...

2023 doi
[16]

Editorial: ChatGPT : Challenges, opportunities, and implications for teacher education

Torrey Trust, Jeromie Whalen, and Chrystalla Mouza. Editorial: ChatGPT : Challenges, opportunities, and implications for teacher education. 2023

2023
[17]

Vazquez, Ulisse Mini, and Monte MacDiarmid

Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini, and Monte MacDiarmid . Steering language models with activation engineering, 2024. URL http://arxiv.org/abs/2308.10248

2024 arXiv
[18]

Mitigating overconfidence in large language models: A behavioral lens on confidence estimation and calibration

Bingbing Wen, Chenjun Xu, Bin Han, Robert Wolfe, Lucy Lu Wang, and Bill Howe. Mitigating overconfidence in large language models: A behavioral lens on confidence estimation and calibration. 2024

2024
[19]

Webis-persuasive-debaters-on-reddit- CMV -2022, 2022

Matti Wiegmann, Khalid Al-Khatib, Vishal Khanna , and Benno Stein. Webis-persuasive-debaters-on-reddit- CMV -2022, 2022. URL https://zenodo.org/record/7034173

2022
[20]

Windschitl and Gary L

Paul D. Windschitl and Gary L. Wells. Measuring psychological uncertainty: Verbal versus numeric methods. 2 0 (4): 0 343--364, 1996. ISSN 1939-2192, 1076-898X. doi:10.1037/1076-898X.2.4.343. URL https://doi.apa.org/doi/10.1037/1076-898X.2.4.343

1996 doi
[21]

SteerConf : Steering LLMs for confidence elicitation, 2025

Ziang Zhou, Tianyuan Jin, Jieming Shi, and Qing Li. SteerConf : Steering LLMs for confidence elicitation, 2025. URL http://arxiv.org/abs/2503.02863

2025 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Refusal in language models is mediated by a single direction, 2024

Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction, 2024. URL http://arxiv.org/abs/2406.11717

arXiv 2024

[2] [2]

Voelkel, Shane Muldowney, Johannes C

Hui Bai, Jan G. Voelkel, Shane Muldowney, Johannes C. Eichstaedt, and Robb Willer. LLM -generated messages can persuade humans on policy issues. 16 0 (1): 0 6037, 2025. ISSN 2041-1723. doi:10.1038/s41467-025-61345-5. URL https://www.nature.com/articles/s41467-025-61345-5

doi:10.1038/s41467-025-61345-5 2025

[3] [3]

Measuring the persuasiveness of language models, 2024

Esin Durmus, Liane Lovitt, Alex Tamkin, Stuart Ritchie, Jack Clark, and Deep Ganguli. Measuring the persuasiveness of language models, 2024. URL https://www.anthropic.com/news/measuring-model-persuasiveness

work page 2024

[4] [4]

Toy models of superposition

Nelson Elhage, Tristan Hume, Catherine Olsson, Nicholas Schiefer, Tom Henighan, Shauna Kravec, Zac Hatfield-Dodds, Robert Lasenby, Dawn Drain, Carol Chen, Roger Grosse, Sam McCandlish, Jared Kaplan, Dario Amodei, Martin Wattenberg, and Christopher Olah. Toy models of superposition. Transformer Circuits Thread, 2022. https://transformer-circuits.pub/2022/t...

2022

[5] [5]

EPISTEMIC INTEGRITY IN LARGE LANGUAGE MODELS

Bijean Ghafouri, Shahrad Mohammadzadeh, James Zhou, Pratheeksha Nair, Jacob-Junqi Tian, Hikaru Tsujimura, Mayank Goel, Sukanya Krishna, Reihaneh Rabbany, Jean-François Godbout, and Kellin Pelrine. EPISTEMIC INTEGRITY IN LARGE LANGUAGE MODELS . 2024

work page 2024

[6] [6]

Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zhang, Aurelien Rodriguez, Austen Gregerson, Ava S...

arXiv 2024

[7] [7]

Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen

Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, and Weizhu Chen. LoRA : Low-rank adaptation of large language models, 2021. URL http://arxiv.org/abs/2106.09685

arXiv 2021

[8] [8]

probable

Tzur M. Karelitz and David V. Budescu. You say "probable" and i say "likely": Improving interpersonal communication with verbal probability phrases. 10 0 (1): 0 25--41, 2004. ISSN 1939-2192, 1076-898X. doi:10.1037/1076-898X.10.1.25. URL https://doi.apa.org/doi/10.1037/1076-898X.10.1.25

doi:10.1037/1076-898x.10.1.25 2004