Mechanistic Personality Analysis of LLMs Steering Personality via Latent Feature Interventions

David Courtis; Ting Hu

arxiv: 2606.28770 · v1 · pith:G6OD2WKUnew · submitted 2026-06-27 · 💻 cs.AI

Mechanistic Personality Analysis of LLMs Steering Personality via Latent Feature Interventions

David Courtis , Ting Hu This is my paper

Pith reviewed 2026-06-30 09:42 UTC · model grok-4.3

classification 💻 cs.AI

keywords mechanistic interpretabilitypersonality steeringsparse autoencodersOCEAN traitsactivation interventionsLLM controlsteering vectors

0 comments

The pith

LLMs' personality traits can be steered by adding small shifts to specific latent directions in their activations.

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

This paper explores a way to change how large language models express personality traits by working directly with their internal representations. Researchers use sparse autoencoders to find directions in the model's hidden states that relate to the Big Five personality traits. They then add a steering vector along those directions to increase or decrease the desired trait. The key finding is that this can be done with minimal impact on the model's ability to perform standard tasks. Readers might care because it suggests a more precise method for controlling model outputs compared to changing prompts or retraining the whole model.

Core claim

The authors show that latent directions identified via sparse autoencoders and contrastive activation analysis in the residual stream can be used to construct additive steering vectors. Applying these vectors enhances the expression of a target OCEAN trait in generated text. The method uses a linear weighting heuristic optimized via grid search to balance trait expression against task performance.

What carries the argument

The additive steering vector constructed from SAE latent directions via contrastive activation analysis, applied as a small shift in the residual stream.

If this is right

Targeted personality traits can be enhanced while overall language modeling performance remains high on standard benchmarks.
Grid search optimization finds combinations of feature shifts that balance personality expression with task performance.
This approach offers a mechanistic alternative to prompt engineering for controlling LLM personality.

Where Pith is reading between the lines

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

If the directions prove causal, similar interventions could allow editing of other model behaviors like safety or consistency after deployment.
The technique might generalize to steering non-personality attributes by identifying their corresponding latent directions.
It suggests personality expression in LLMs may depend on a limited set of identifiable features rather than being fully distributed.

Load-bearing premise

The latent directions identified by the sparse autoencoders and contrastive analysis causally correspond to the OCEAN personality traits rather than just appearing in the training data.

What would settle it

Measuring personality in model outputs before and after applying the steering vector and finding no statistically significant change in trait scores on validated tests.

Figures

Figures reproduced from arXiv: 2606.28770 by David Courtis, Ting Hu.

**Figure 1.** Figure 1: Overview of our personality steering pipeline, showing the flow from data generation through feature extraction, optimization, and evaluation. Processes transform the data into results, while evaluations transform results into data for the algorithm and results analysis. The gridsearch algorithm involves two benchmarks with the third being performed post-gridsearch. Each iteration of the gridsearch algo… view at source ↗

**Figure 2.** Figure 2: LLM Binary Classification per Trait, showing the percentage of correct (high confidence), low confidence, and incorrect classifications for each OCEAN trait. Conscientiousness and Neuroticism show the highest detection rates > 90%, while Openness presents significant challenges [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗

**Figure 3.** Figure 3: Human Binary Classification per Trait, showing the percentage of correct (high confidence), low confidence, and incorrect classifications for each OCEAN trait. Note the similar pattern to LLM classification, with higher uncertainty for Openness and Agreeableness. vectors for each trait as described in Section 3.4. The intervention strength parameters were calibrated through our grid search optimization t… view at source ↗

**Figure 5.** Figure 5: Optimal parameters identified through grid search for each OCEAN trait, showing positive and negative feature magnitudes and counts. Note how Openness and Conscientiousness require the most features, while Extraversion requires the fewest. describing what they have been up to recently. This should give a broad and varied view of their experience, either positive or negative. Talk about a wide variety of th… view at source ↗

**Figure 6.** Figure 6: Objective score as a function of shift magnitude for each OCEAN trait, highlighting the optimal magnitude (marked with red) before performance degradation. Neuroticism and Openness show the highest peak scores, while all traits exhibit diminishing returns with excessive intervention. specific application, facilitating rapid and cost-effective prototyping of personality-aware AI systems. Despite these adva… view at source ↗

read the original abstract

Large Language Models (LLMs) have demonstrated the ability to simulate human-like OCEAN personality traits in generated text. Previous efforts have focused on prompt engineering or fine-tuning to shape LLM personality. In this work, we propose a mechanistic interpretability approach that directly intervenes on the model's latent features. Our method identifies latent directions in the residual stream corresponding to a target OCEAN trait using sparse autoencoders (SAEs) and contrastive activation analysis. We formalize an additive steering vector in activation space and demonstrate how applying a small additive shift to the hidden states enhances the target trait while preserving overall language modeling performance. To determine the optimal combination of feature shifts, we explore a linear weighting heuristic with grid search optimization that balances personality expression with task performance. Our approach shows promise in controllably steering personality traits at the mechanistic level while maintaining high performance on standard benchmarks.

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

This applies known SAE steering methods to personality traits but supplies no numbers or checks to show the interventions actually work.

read the letter

The paper takes established sparse autoencoder and contrastive activation techniques, locates directions in the residual stream tied to OCEAN traits, and adds small steering vectors with a grid-searched linear weighting to try to boost the target trait without hurting language modeling.

It lays out the activation-space formalization and the balancing heuristic in a straightforward way. That part is clear and follows directly from prior activation engineering work.

The main gap is the lack of any reported results. The abstract says the method shows promise and demonstrates effects, yet there are no personality score deltas, no benchmark numbers with or without steering, no ablation on which SAE features matter, and no statement on whether the grid search used held-out data. Without those, it is not possible to judge whether the steering succeeds or whether the weighting step overfits.

The stress-test concern is on target. Finding directions that differ across trait-conditioned prompts does not establish that intervening on them produces causal control. The directions could be downstream correlates or prompt artifacts, in which case the additive shifts would not deliver the claimed controllability.

This is a routine extension of existing tools to a new domain rather than a new method. Readers already working on mechanistic steering for alignment might find the setup useful once the missing empirical checks are added.

I would not bring it to a reading group yet. I would not cite it. It does not look ready for peer review because the central claim rests on unshown evidence.

Referee Report

3 major / 1 minor

Summary. The paper claims to introduce a mechanistic interpretability approach for steering OCEAN personality traits in LLMs by identifying latent directions in the residual stream using sparse autoencoders (SAEs) and contrastive activation analysis. It formalizes additive steering vectors applied to hidden states and uses a linear weighting heuristic with grid search to optimize the balance between personality expression and task performance, asserting that this method allows controllable steering at the mechanistic level while maintaining high performance on standard benchmarks.

Significance. If the results hold and the interventions prove causally effective, this would represent a significant advance in controllable generation and mechanistic understanding of LLMs, offering a way to edit specific behavioral traits without retraining or prompting, which could have implications for alignment and personalized AI systems.

major comments (3)

[Abstract] The abstract states that the method 'shows promise' and 'demonstrates' effects but supplies no quantitative results, error bars, benchmark numbers, or ablation data. This is load-bearing for the central claim of demonstrating controllable steering, as the empirical support is not visible.
[Linear weighting heuristic description] The linear weighting heuristic with grid search is described as balancing personality and performance without stating whether the search is performed on held-out data or risks fitting to the evaluation itself. This raises a potential circularity issue for the performance preservation claim.
[Identification of latent directions] The SAE and contrastive activation analysis identify directions that differ across trait-conditioned activations, but this does not establish that additive interventions on these directions will causally control the OCEAN traits rather than merely correlating with them. This is the weakest assumption underlying the steering method.

minor comments (1)

[Notation] The formalization of the additive steering vector could benefit from explicit equations to clarify the intervention process.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback. We address each major comment below, proposing revisions where they strengthen the manuscript without misrepresenting our results.

read point-by-point responses

Referee: [Abstract] The abstract states that the method 'shows promise' and 'demonstrates' effects but supplies no quantitative results, error bars, benchmark numbers, or ablation data. This is load-bearing for the central claim of demonstrating controllable steering, as the empirical support is not visible.

Authors: We agree that the abstract would be strengthened by including key quantitative results. In the revision we will incorporate specific benchmark numbers, personality expression metrics, and error bars drawn from the experimental sections to make the central claims concrete. revision: yes
Referee: [Linear weighting heuristic description] The linear weighting heuristic with grid search is described as balancing personality and performance without stating whether the search is performed on held-out data or risks fitting to the evaluation itself. This raises a potential circularity issue for the performance preservation claim.

Authors: This concern is valid given the current description. The grid search was performed on a held-out validation split separate from the reported test benchmarks. We will revise the methods section to explicitly document this procedure and confirm that evaluation metrics reflect performance on unseen data. revision: yes
Referee: [Identification of latent directions] The SAE and contrastive activation analysis identify directions that differ across trait-conditioned activations, but this does not establish that additive interventions on these directions will causally control the OCEAN traits rather than merely correlating with them. This is the weakest assumption underlying the steering method.

Authors: The identification step is correlational, but the paper's central evidence for causal control comes from the subsequent intervention experiments, which apply the vectors and measure resulting changes in generated text personality scores. We will add a short discussion paragraph clarifying this distinction and the evidential role of the steering results. revision: partial

Circularity Check

0 steps flagged

No significant circularity in derivation chain

full rationale

The abstract and description outline identifying directions via SAEs plus contrastive activation analysis, formalizing additive steering vectors, applying shifts, and using grid search for linear weights to balance traits and performance. No equations or steps are shown that reduce by construction to their own inputs, no self-citations are load-bearing on uniqueness or ansatzes, and no fitted parameters are relabeled as independent predictions. The optimization step is presented as a heuristic without evidence it collapses the central claim to the identification data itself. The derivation remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Only the abstract is available, so the ledger is necessarily incomplete; the method implicitly rests on the assumption that SAEs recover causally meaningful directions.

axioms (1)

domain assumption Sparse autoencoders recover disentangled, causally relevant features from LLM residual stream activations.
Invoked when the paper states that SAEs identify directions corresponding to target OCEAN traits.

pith-pipeline@v0.9.1-grok · 5672 in / 1216 out tokens · 45409 ms · 2026-06-30T09:42:33.072648+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

19 extracted references · 16 canonical work pages · 4 internal anchors

[1]

Arditi, A., Obeso, O., Syed, A., Paleka, D., Panickssery, N., Gurnee, W., and Nanda, N

doi: 10.3389/fncom.2021.654315. Arditi, A., Obeso, O., Syed, A., Paleka, D., Panickssery, N., Gurnee, W., and Nanda, N. Refusal in language models is mediated by a single direction,

work page doi:10.3389/fncom.2021.654315 2021
[2]

URL https://arxiv.org/abs/2406.11717. Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., Turner, N., Anil, C., Denison, C., Askell, A., Lasenby, R., Wu, Y ., Kravec, S., Schiefer, N., Maxwell, T., Joseph, N., Hatfield-Dodds, Z., Tamkin, A., Nguyen, K., McLean, B., Burke, J. E., Hume, T., Carter, S., Henighan, T., and Olah, C. Towar...

work page internal anchor Pith review Pith/arXiv arXiv
[3]

Cunningham, H., Ewart, A., Riggs, L., Huben, R., and Sharkey, L

https: //transformer-circuits.pub/2023/ monosemantic-features/index.html. Cunningham, H., Ewart, A., Riggs, L., Huben, R., and Sharkey, L. Sparse autoencoders find highly interpretable features in language models,

2023
[4]

URL https:// arxiv.org/abs/2309.08600. Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., Hatfield-Dodds, Z., Lasenby, R., Drain, D., Chen, C., Grosse, R., McCandlish, S., Kaplan, J., Amodei, D., Wattenberg, M., and Olah, C. Toy models of superposition.Transformer Circuits Thread,

work page internal anchor Pith review Pith/arXiv arXiv
[5]

Fatemi, S., Hu, Y ., and Mousavi, M

https://transformer-circuits.pub/ 2022/toy_model/index.html. Fatemi, S., Hu, Y ., and Mousavi, M. A comparative analysis of instruction fine-tuning llms for financial text classifica- tion,

2022
[6]

Scaling and evaluating sparse autoencoders

URL https: //arxiv.org/abs/2406.04093. Hagendorff, T., Dasgupta, I., Binz, M., Chan, S. C. Y ., Lampinen, A., Wang, J. X., Akata, Z., and Schulz, E. Machine psychology,

work page internal anchor Pith review Pith/arXiv arXiv
[7]

org/abs/2303.13988

URL https://arxiv. org/abs/2303.13988. Hilliard, A., Munoz, C., Wu, Z., and Koshiyama, A. S. Eliciting personality traits in large language models,

work page arXiv
[8]

Jiang, H., Zhang, X., Cao, X., Breazeal, C., Roy, D., and Kabbara, J

URLhttps://arxiv.org/abs/2402.08341. Jiang, H., Zhang, X., Cao, X., Breazeal, C., Roy, D., and Kabbara, J. Personallm: Investigating the ability of large 13 Mechanistic Personality Analysis of LLMs language models to express personality traits,

work page arXiv
[9]

Kerz, E., Qiao, Y ., Zanwar, S., and Wiechmann, D

URL https://arxiv.org/abs/2305.02547. Kerz, E., Qiao, Y ., Zanwar, S., and Wiechmann, D. Push- ing on personality detection from verbal behavior: A transformer meets text contours of psycholinguistic fea- tures. In Barnes, J., De Clercq, O., Barriere, V ., Tafreshi, S., Alqahtani, S., Sedoc, J., Klinger, R., and Balahur, A. (eds.),Proceedings of the 12th ...

work page arXiv
[10]

doi: 10.18653/v1/2022.wassa-1.17

Association for Computational Linguis- tics. doi: 10.18653/v1/2022.wassa-1.17. URL https: //aclanthology.org/2022.wassa-1.17/. Lee, S., Lim, S., Han, S., Oh, G., Chae, H., Chung, J., Kim, M., woo Kwak, B., Lee, Y ., Lee, D., Yeo, J., and Yu, Y . Do llms have distinct and consistent personality? trait: Personality testset designed for llms with psychomet- rics,

work page doi:10.18653/v1/2022.wassa-1.17 2022
[11]

Peters, H

URLhttps://arxiv.org/abs/2502.08265. Peters, H. and Matz, S. C. Large language models can infer psychological dispositions of social media users.PNAS Nexus, 3(6), May

work page arXiv
[12]

Reynolds, L

doi: 10.1093/pnasnexus/pgae231. Reynolds, L. and McDonell, K. Prompt program- ming for large language models: Beyond the few-shot paradigm,

work page doi:10.1093/pnasnexus/pgae231
[13]

Serapio-Garc´ıa, G., Safdari, M., Crepy, C., Sun, L., Fitz, S., Romero, P., Abdulhai, M., Faust, A., and Matari´c, M

URL https://arxiv.org/abs/ 2102.07350. Serapio-Garc´ıa, G., Safdari, M., Crepy, C., Sun, L., Fitz, S., Romero, P., Abdulhai, M., Faust, A., and Matari´c, M. Personality traits in large language models,

work page arXiv
[14]

Personality traits in large language models,

URL https://arxiv.org/abs/2307.00184. Sorokovikova, A., Fedorova, N., Rezagholi, S., and Yamshchikov, I. P. Llms simulate big five personality traits: Further evidence,

work page arXiv
[15]

org/abs/2402.01765

URL https://arxiv. org/abs/2402.01765. Templeton, A., Conerly, T., Marcus, J., Lindsey, J., Bricken, T., Chen, B., Pearce, A., Citro, C., Ameisen, E., Jones, A., Cunningham, H., Turner, N. L., McDougall, C., MacDiarmid, M., Freeman, C. D., Sumers, T. R., Rees, E., Batson, J., Jermyn, A., Carter, S., Olah, C., and Henighan, T. Scaling monosemanticity: Ex- ...

work page arXiv
[16]

Wang, S., Li, R., Chen, X., Yuan, Y ., Wong, D

URL https: //transformer-circuits.pub/2024/ scaling-monosemanticity/index.html. Wang, S., Li, R., Chen, X., Yuan, Y ., Wong, D. F., and Yang, M. Exploring the impact of personality traits on llm bias and toxicity,

2024
[17]

Exploring the impact of personality traits on llm bias and toxicity,

URL https://arxiv. org/abs/2502.12566. Wei, J., Tay, Y ., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., Chi, E. H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J., and Fedus, W. Emergent abili- ties of large language models,

work page arXiv
[18]

Emergent Abilities of Large Language Models

URL https: //arxiv.org/abs/2206.07682. Widiger, T. A. and Crego, C. The five factor model of personality structure: an update.World Psychiatry, 18 (3):271–272,

work page internal anchor Pith review Pith/arXiv arXiv
[19]

doi: https://doi.org/10.1002/wps. 20658. URL https://onlinelibrary.wiley. com/doi/abs/10.1002/wps.20658. 14 Mechanistic Personality Analysis of LLMs A. Additional Results 15 Mechanistic Personality Analysis of LLMs Table 4.Full Comparison of Tuned-Up vs. Tuned-Down Responses for All Five Traits. Trait Tuned Up Response Excerpt Tuned Down Response Excerpt ...

work page doi:10.1002/wps

[1] [1]

Arditi, A., Obeso, O., Syed, A., Paleka, D., Panickssery, N., Gurnee, W., and Nanda, N

doi: 10.3389/fncom.2021.654315. Arditi, A., Obeso, O., Syed, A., Paleka, D., Panickssery, N., Gurnee, W., and Nanda, N. Refusal in language models is mediated by a single direction,

work page doi:10.3389/fncom.2021.654315 2021

[2] [2]

URL https://arxiv.org/abs/2406.11717. Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., Turner, N., Anil, C., Denison, C., Askell, A., Lasenby, R., Wu, Y ., Kravec, S., Schiefer, N., Maxwell, T., Joseph, N., Hatfield-Dodds, Z., Tamkin, A., Nguyen, K., McLean, B., Burke, J. E., Hume, T., Carter, S., Henighan, T., and Olah, C. Towar...

work page internal anchor Pith review Pith/arXiv arXiv

[3] [3]

Cunningham, H., Ewart, A., Riggs, L., Huben, R., and Sharkey, L

https: //transformer-circuits.pub/2023/ monosemantic-features/index.html. Cunningham, H., Ewart, A., Riggs, L., Huben, R., and Sharkey, L. Sparse autoencoders find highly interpretable features in language models,

2023

[4] [4]

URL https:// arxiv.org/abs/2309.08600. Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., Hatfield-Dodds, Z., Lasenby, R., Drain, D., Chen, C., Grosse, R., McCandlish, S., Kaplan, J., Amodei, D., Wattenberg, M., and Olah, C. Toy models of superposition.Transformer Circuits Thread,

work page internal anchor Pith review Pith/arXiv arXiv

[5] [5]

Fatemi, S., Hu, Y ., and Mousavi, M

https://transformer-circuits.pub/ 2022/toy_model/index.html. Fatemi, S., Hu, Y ., and Mousavi, M. A comparative analysis of instruction fine-tuning llms for financial text classifica- tion,

2022

[6] [6]

Scaling and evaluating sparse autoencoders

URL https: //arxiv.org/abs/2406.04093. Hagendorff, T., Dasgupta, I., Binz, M., Chan, S. C. Y ., Lampinen, A., Wang, J. X., Akata, Z., and Schulz, E. Machine psychology,

work page internal anchor Pith review Pith/arXiv arXiv

[7] [7]

org/abs/2303.13988

URL https://arxiv. org/abs/2303.13988. Hilliard, A., Munoz, C., Wu, Z., and Koshiyama, A. S. Eliciting personality traits in large language models,

work page arXiv

[8] [8]

Jiang, H., Zhang, X., Cao, X., Breazeal, C., Roy, D., and Kabbara, J

URLhttps://arxiv.org/abs/2402.08341. Jiang, H., Zhang, X., Cao, X., Breazeal, C., Roy, D., and Kabbara, J. Personallm: Investigating the ability of large 13 Mechanistic Personality Analysis of LLMs language models to express personality traits,

work page arXiv

[9] [9]

Kerz, E., Qiao, Y ., Zanwar, S., and Wiechmann, D

URL https://arxiv.org/abs/2305.02547. Kerz, E., Qiao, Y ., Zanwar, S., and Wiechmann, D. Push- ing on personality detection from verbal behavior: A transformer meets text contours of psycholinguistic fea- tures. In Barnes, J., De Clercq, O., Barriere, V ., Tafreshi, S., Alqahtani, S., Sedoc, J., Klinger, R., and Balahur, A. (eds.),Proceedings of the 12th ...

work page arXiv

[10] [10]

doi: 10.18653/v1/2022.wassa-1.17

Association for Computational Linguis- tics. doi: 10.18653/v1/2022.wassa-1.17. URL https: //aclanthology.org/2022.wassa-1.17/. Lee, S., Lim, S., Han, S., Oh, G., Chae, H., Chung, J., Kim, M., woo Kwak, B., Lee, Y ., Lee, D., Yeo, J., and Yu, Y . Do llms have distinct and consistent personality? trait: Personality testset designed for llms with psychomet- rics,

work page doi:10.18653/v1/2022.wassa-1.17 2022

[11] [11]

Peters, H

URLhttps://arxiv.org/abs/2502.08265. Peters, H. and Matz, S. C. Large language models can infer psychological dispositions of social media users.PNAS Nexus, 3(6), May

work page arXiv

[12] [12]

Reynolds, L

doi: 10.1093/pnasnexus/pgae231. Reynolds, L. and McDonell, K. Prompt program- ming for large language models: Beyond the few-shot paradigm,

work page doi:10.1093/pnasnexus/pgae231

[13] [13]

Serapio-Garc´ıa, G., Safdari, M., Crepy, C., Sun, L., Fitz, S., Romero, P., Abdulhai, M., Faust, A., and Matari´c, M

URL https://arxiv.org/abs/ 2102.07350. Serapio-Garc´ıa, G., Safdari, M., Crepy, C., Sun, L., Fitz, S., Romero, P., Abdulhai, M., Faust, A., and Matari´c, M. Personality traits in large language models,

work page arXiv

[14] [14]

Personality traits in large language models,

URL https://arxiv.org/abs/2307.00184. Sorokovikova, A., Fedorova, N., Rezagholi, S., and Yamshchikov, I. P. Llms simulate big five personality traits: Further evidence,

work page arXiv

[15] [15]

org/abs/2402.01765

URL https://arxiv. org/abs/2402.01765. Templeton, A., Conerly, T., Marcus, J., Lindsey, J., Bricken, T., Chen, B., Pearce, A., Citro, C., Ameisen, E., Jones, A., Cunningham, H., Turner, N. L., McDougall, C., MacDiarmid, M., Freeman, C. D., Sumers, T. R., Rees, E., Batson, J., Jermyn, A., Carter, S., Olah, C., and Henighan, T. Scaling monosemanticity: Ex- ...

work page arXiv

[16] [16]

Wang, S., Li, R., Chen, X., Yuan, Y ., Wong, D

URL https: //transformer-circuits.pub/2024/ scaling-monosemanticity/index.html. Wang, S., Li, R., Chen, X., Yuan, Y ., Wong, D. F., and Yang, M. Exploring the impact of personality traits on llm bias and toxicity,

2024

[17] [17]

Exploring the impact of personality traits on llm bias and toxicity,

URL https://arxiv. org/abs/2502.12566. Wei, J., Tay, Y ., Bommasani, R., Raffel, C., Zoph, B., Borgeaud, S., Yogatama, D., Bosma, M., Zhou, D., Metzler, D., Chi, E. H., Hashimoto, T., Vinyals, O., Liang, P., Dean, J., and Fedus, W. Emergent abili- ties of large language models,

work page arXiv

[18] [18]

Emergent Abilities of Large Language Models

URL https: //arxiv.org/abs/2206.07682. Widiger, T. A. and Crego, C. The five factor model of personality structure: an update.World Psychiatry, 18 (3):271–272,

work page internal anchor Pith review Pith/arXiv arXiv

[19] [19]

doi: https://doi.org/10.1002/wps. 20658. URL https://onlinelibrary.wiley. com/doi/abs/10.1002/wps.20658. 14 Mechanistic Personality Analysis of LLMs A. Additional Results 15 Mechanistic Personality Analysis of LLMs Table 4.Full Comparison of Tuned-Up vs. Tuned-Down Responses for All Five Traits. Trait Tuned Up Response Excerpt Tuned Down Response Excerpt ...

work page doi:10.1002/wps