pith. machine review for the scientific record. sign in

arxiv: 2605.13329 · v1 · submitted 2026-05-13 · 💻 cs.CL · cs.AI

Recognition: unknown

Tracing Persona Vectors Through LLM Pretraining

Dominik Glandorf, Jorge Medina Moreira, Robert West, Tanja K\"aser, Viktor Moskvoretskii

Authors on Pith no claims yet

Pith reviewed 2026-05-14 20:25 UTC · model grok-4.3

classification 💻 cs.CL cs.AI
keywords persona vectorsLLM pretraininglinear directionsinterpretabilityAI safetyelicitation methodsmodel steeringOLMo
0
0 comments X

The pith

Persona vectors for traits like sycophancy emerge within the first 0.22 percent of LLM pretraining and remain usable for steering the final model.

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

The paper tracks linear directions in model activations that represent high-level personas during the pretraining run of OLMo-3-7B. These directions appear after only 0.22 percent of training tokens and continue to work for steering the fully post-trained version of the model. The core vectors stabilize early but keep sharpening in both their geometric alignment and semantic content as training proceeds. Different methods for extracting the vectors each surface distinct aspects of the same underlying persona. The same early-formation pattern appears when the analysis is repeated on Apertus-8B.

Core claim

Persona vectors form remarkably early within 0.22% of OLMo-3 pretraining and remain effective for steering the fully post-trained instruct models. Although core representations are formed early on, persona vectors continue to refine geometrically and semantically throughout pretraining. All elicitation strategies produce effective directions, each highlighting qualitatively distinct facets of the persona, and the pattern transfers to Apertus-8B.

What carries the argument

Persona vectors: linear directions in internal activations that correspond to high-level behaviors or traits such as evil or sycophancy.

If this is right

  • Safety monitoring for persona-related behaviors can begin after a tiny fraction of pretraining rather than waiting for the end.
  • Ongoing geometric and semantic refinement implies that later training stages still reshape how these traits are expressed.
  • Multiple elicitation methods can be combined to capture complementary facets of a single persona for more precise interventions.
  • Because the pattern repeats on a second model, early persona formation may be a general feature of current pretraining regimes.

Where Pith is reading between the lines

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

  • Pretraining data must contain sufficiently strong and consistent signals for these personas right from the initial tokens.
  • Later post-training may primarily amplify or suppress directions that already exist rather than create them from scratch.
  • Early identification of these vectors could let developers test and adjust data mixtures before committing to full-scale training runs.
  • The same tracing approach might reveal when other high-level concepts, such as factual recall or reasoning styles, first crystallize.

Load-bearing premise

The early linear directions identified at checkpoints truly represent the same high-level personas that exist in the final model rather than being mixed with unrelated training effects.

What would settle it

Measuring whether the vector extracted at the 0.22 percent checkpoint produces the expected steering effect when applied to the final post-trained model; a large drop in steering success would falsify the claim.

Figures

Figures reproduced from arXiv: 2605.13329 by Dominik Glandorf, Jorge Medina Moreira, Robert West, Tanja K\"aser, Viktor Moskvoretskii.

Figure 1
Figure 1. Figure 1: Personas emerge early and stay stable throughout OLMo-3 7B pretraining. Top: For each trait, the persona vector is extracted at checkpoint Mr and applied to steer the same checkpoint. The y-axis reports the trait-expression delta ∆τ (steered − baseline; higher = stronger steering effect). Bottom: pass rate at Mr — the fraction of generations whose trait score ≥ 50. The shared x-axis shows checkpoints as pr… view at source ↗
Figure 2
Figure 2. Figure 2: Early persona vectors transfer to OLMo-3-7B-Instruct and grow in effect as training progresses. For each trait, the persona vector is extracted at base-model checkpoint Mr and applied to steer the Instruct model. The y-axis reports the trait-expression delta ∆τ (Instruct, Mr) (steered − baseline; higher = stronger steering effect); the x-axis shows extraction checkpoint Mr as cumulative tokens, spanning pr… view at source ↗
Figure 3
Figure 3. Figure 3: Persona transfer works both within pretraining and across OLMo-3 post-training (SFT / DPO / RLVR). Each extracted base-model vector is applied to four fixed targets: the final pretraining checkpoint main and the three Instruct variants. The y-axis reports the trait-expression delta ∆τ (M, Mr) on target M (steered − baseline; higher = stronger steering effect, range set per panel for clarity). Lines aggrega… view at source ↗
Figure 4
Figure 4. Figure 4: Persona vectors form gradually and stabilize during pretraining. Top: co￾sine similarity of the persona vector at extrac￾tion checkpoint t with its final-pretraining counterpart. Bottom: cosine similarity of ad￾jacent checkpoints. x-axis: extraction check￾point as cumulative tokens (log scale). 1.0 0.5 0.0 0.5 1.0 MDS 1 1.0 0.5 0.0 0.5 1.0 MDS 2 Trait Evil Humorous Impolite Sycophantic Checkpoint early (21… view at source ↗
Figure 6
Figure 6. Figure 6: Trait facets evolve with the changing vector geometry during pretraining. Baumeister’s roots of evil, as well as three facets of sycophancy, vary in their expression in the generations used for persona vector computation. 5.2 Expressed Facets of Personas Partially Refine During Pretraining In the previous subsection, we observed that persona vectors continue to refine their geometry. In this section, we ai… view at source ↗
Figure 7
Figure 7. Figure 7: Persona emerges early and stays stable throughout Apertus-8B pretraining. For each trait, the persona vector is extracted at checkpoint Mr and applied to steering the same checkpoint. The y-axis reports the LLM-judge trait-expression delta (steered − baseline); the x-axis shows checkpoints as cumulative pretraining tokens seen at Mr. Bottom-side markers flag checkpoints where no vector could be built (×: o… view at source ↗
Figure 8
Figure 8. Figure 8: Very early persona vectors transfer to Apertus-8B-Instruct and grow in effect as training progresses. For each trait, the persona vector is extracted at base-model checkpoint Mr and applied to steering the Instruct model. The y-axis reports the LLM-judge trait-expression delta on Instruct (steered − baseline); the x-axis shows the extraction checkpoint as cumulative pretraining tokens seen at Mr. base (mai… view at source ↗
Figure 9
Figure 9. Figure 9: Persona transfer works both within pretraining and after Apertus-8B post-training (Instruct). Each extracted base-model vector is applied to two fixed targets: the final pretraining checkpoint main and the Instruct variants. The y-axis reports the trait-expression delta ∆τ (M, Mr) on target M (steered − baseline; higher = stronger steering effect). Lines aggregate source checkpoints Mr into three pretraini… view at source ↗
Figure 10
Figure 10. Figure 10: Top: cosine similarity of the persona vector at extraction checkpoint t of Apertus with its final-pretraining counterpart. Bottom: cosine similarity of adjacent check￾points. x-axis: extraction checkpoint as cu￾mulative tokens (log scale). 1.0 0.5 0.0 0.5 1.0 MDS 1 1.00 0.75 0.50 0.25 0.00 0.25 0.50 0.75 MDS 2 Trait Evil Humorous Impolite Sycophantic Checkpoint early (210B) late (15T) [PITH_FULL_IMAGE:fi… view at source ↗
Figure 12
Figure 12. Figure 12: replicates the results from [PITH_FULL_IMAGE:figures/full_fig_p016_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Cosine similarity between persona vectors extracted with different methods for OLMo-3. [PITH_FULL_IMAGE:figures/full_fig_p017_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Our extraction methods vary in the expression of Baumeister’s roots of evil. Percent￾ages show the share of generations from OLMo-3-7B (evil score > 50) assigned each root by GPT-4.1. Each vector extraction method was evaluated on test cases from all three evaluation methods. These conclusions are not specific to GPT-4.1-mini judging: repeating the same extraction￾method comparison with DeepSeek-V4-Flash … view at source ↗
Figure 15
Figure 15. Figure 15: Cross-judge cosine similarity of persona vectors across traits. Rows are persona vectors extracted using GPT-4.1-mini for filtering/evaluation; columns are persona vectors extracted using DeepSeek-V4-Flash under the same prompts, thresholds, layer, and steering settings. B.3 Replication on Apertus The pattern observed on OLMo-3 replicates on Apertus ( [PITH_FULL_IMAGE:figures/full_fig_p019_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Cosine similarity between persona vectors extracted with different discourse elicitation [PITH_FULL_IMAGE:figures/full_fig_p019_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Percentages show the share of generations from [PITH_FULL_IMAGE:figures/full_fig_p020_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: Matched-magnitude negative controls. Real persona vector (solid) versus random-direction [PITH_FULL_IMAGE:figures/full_fig_p026_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: The average hidden state norms, which largely vary, especially in earlier checkpoints and [PITH_FULL_IMAGE:figures/full_fig_p029_19.png] view at source ↗
read the original abstract

How large language models internally represent high-level behaviors is a core interpretability question with direct relevance to AI safety: it determines what we can detect, audit, or intervene on. Recent work has shown that traits such as evil or sycophancy correspond to linear directions in the internal activations, the so-called persona vectors. Although these vectors are now routinely utilized to inspect and steer model behavior in safety-relevant settings, how these representations are formed during training remains unknown. To address this gap, we trace persona vectors across the pretraining of OLMo-3-7B, finding that persona vectors form remarkably early -- within 0.22% of OLMo-3 pretraining -- and remain effective for steering the fully post-trained instruct models. Although core representations are formed early on, persona vectors continue to refine geometrically and semantically throughout pretraining. We further compare alternative elicitation strategies and find that all yield effective directions, with each strategy surfacing qualitatively distinct facets of the underlying persona. Replicating our analysis on Apertus-8B reveals that our findings transfer qualitatively beyond OLMo-3. Our results establish persona representations as stable features of early pretraining and open a path to studying how training forms, refines, and shapes them.

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.

Referee Report

3 major / 2 minor

Summary. The paper traces the formation of persona vectors (linear directions in internal activations corresponding to high-level behaviors such as evil or sycophancy) across checkpoints in the pretraining of OLMo-3-7B, claiming they emerge within 0.22% of training and remain effective for steering post-trained instruct models. Core representations form early but continue to refine geometrically and semantically; alternative elicitation strategies yield effective yet qualitatively distinct directions. The analysis is replicated on Apertus-8B, with findings transferring qualitatively.

Significance. If the timing and persistence claims hold after verification, the work would demonstrate that high-level behavioral representations are stable features of very early pretraining rather than late-emerging or post-training artifacts. This has direct relevance to AI safety and interpretability by suggesting that persona-based auditing and steering could be grounded in early training dynamics, with the cross-model replication providing qualitative support for generality.

major comments (3)
  1. [Abstract and §3] Abstract and §3 (vector extraction): the central timing claim (0.22% checkpoint) and persistence claim lack reported details on exact activation collection procedures, controls for training artifacts (e.g., layer variance or batch effects), and any statistical tests for consistency across checkpoints or models; without these, the equivalence between early and final vectors remains unverified.
  2. [§4] §4 (early formation results): the claim that early vectors represent the same high-level personas is load-bearing but rests only on geometric refinement and steering effectiveness; this does not rule out transient early-training directions (e.g., prompt-specific or variance-driven) that correlate coincidentally with later persona semantics, as the skeptic note highlights.
  3. [§5] §5 (elicitation comparison): qualitative differences across strategies are reported, yet the paper does not quantify their impact on the persistence or steering claims, leaving open whether early vectors isolate stable core personas or distinct facets that happen to steer similarly.
minor comments (2)
  1. [Introduction] Introduction: the term 'persona vector' should be formally defined with reference to prior steering-vector literature to clarify distinctions.
  2. [§4] Figures in §4: legends and axis labels could be expanded to distinguish geometric cosine similarity from semantic steering metrics more clearly.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the detailed and constructive comments. We address each major point below, providing clarifications, additional analyses where possible, and revisions to strengthen the manuscript without altering its core claims.

read point-by-point responses

  1. Referee: [Abstract and §3] Abstract and §3 (vector extraction): the central timing claim (0.22% checkpoint) and persistence claim lack reported details on exact activation collection procedures, controls for training artifacts (e.g., layer variance or batch effects), and any statistical tests for consistency across checkpoints or models; without these, the equivalence between early and final vectors remains unverified.

    Authors: We agree that additional methodological details are required for full verification. In the revised §3, we will specify: the exact set of 50 persona prompts and 50 control prompts used for difference-in-means extraction; the precise layers (12–18 for the 7B model) chosen to reduce layer-wise variance; and the use of fixed random seeds for data loading to control batch effects. We will also include a table of mean cosine similarities between early (0.22%) and final vectors across all personas, with standard deviations computed over multiple prompt subsets. While we did not perform formal statistical tests such as permutation tests in the original submission, we will add these for key similarity comparisons and report the resulting p-values. These changes will make the equivalence claims verifiable. revision: yes

  2. Referee: [§4] §4 (early formation results): the claim that early vectors represent the same high-level personas is load-bearing but rests only on geometric refinement and steering effectiveness; this does not rule out transient early-training directions (e.g., prompt-specific or variance-driven) that correlate coincidentally with later persona semantics, as the skeptic note highlights.

    Authors: We acknowledge that geometric similarity and steering success alone do not fully exclude coincidental correlations with transient directions. The manuscript already provides three converging lines of evidence: monotonic increase in cosine similarity to final vectors, effective steering of the post-trained model on held-out test sets, and qualitative consistency in elicited behaviors. To directly address the concern, we will add a control analysis in the revision comparing persona vectors against directions derived from unrelated prompt sets (e.g., factual recall prompts); only the persona-derived vectors maintain consistent steering performance across checkpoints. The qualitative replication on Apertus-8B further reduces the chance of model-specific artifacts. We will also expand the limitations section to explicitly discuss the possibility of transient correlations. revision: partial

  3. Referee: [§5] §5 (elicitation comparison): qualitative differences across strategies are reported, yet the paper does not quantify their impact on the persistence or steering claims, leaving open whether early vectors isolate stable core personas or distinct facets that happen to steer similarly.

    Authors: We will revise §5 to include quantitative comparisons. Specifically, we will report steering success rates (fraction of desired behavioral outcomes on standardized evaluation sets) for vectors obtained from each elicitation strategy at both the 0.22% checkpoint and the final checkpoint. We will also add pairwise cosine similarities between vectors from different strategies and track how these overlaps evolve over training. These metrics will directly quantify whether the observed qualitative differences affect the persistence and steering claims or whether all strategies converge on stable core representations. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical checkpoint comparisons

full rationale

The paper reports direct empirical extraction and comparison of linear activation directions (persona vectors) at successive pretraining checkpoints of OLMo-3-7B and Apertus-8B. Vectors are obtained via fixed elicitation prompts, then evaluated for geometric similarity, semantic refinement, and steering efficacy on the final model. No equations, fitted parameters, or derivations are presented that reduce to their own inputs; claims rest on observed data patterns rather than self-referential definitions or self-citation load-bearing steps.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

The central claim rests on the domain assumption that high-level behaviors are captured by linear directions in activation space, taken from prior persona-vector work; the 0.22% timing is an empirical observation rather than a fitted free parameter.

axioms (1)
  • domain assumption High-level behaviors such as evil or sycophancy correspond to linear directions in internal activations
    Invoked throughout the abstract as the basis for tracing persona vectors; taken from recent prior work.

pith-pipeline@v0.9.0 · 5529 in / 1197 out tokens · 31948 ms · 2026-05-14T20:25:37.504219+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

43 extracted references · 18 canonical work pages · 5 internal anchors

  1. [1]

    System card: Claude Mythos preview

    Anthropic. System card: Claude Mythos preview. System card, Anthropic, April 2026. Accessed: 2026-05-01

    2026

  2. [2]

    GPT-5.5 system card

    OpenAI. GPT-5.5 system card. System card, OpenAI, April 2026. Accessed: 2026-05-01

    2026

  3. [3]

    International ai safety report 2026.arXiv preprint arXiv:2602.21012, 2026

    Yoshua Bengio, Stephen Clare, Carina Prunkl, Maksym Andriushchenko, Ben Bucknall, Mal- colm Murray, Rishi Bommasani, Stephen Casper, Tom Davidson, Raymond Douglas, et al. International ai safety report 2026.arXiv preprint arXiv:2602.21012, 2026

    work page arXiv 2026

  4. [4]

    Open Problems in Mechanistic Interpretability

    Lee Sharkey, Bilal Chughtai, Joshua Batson, Jack Lindsey, Jeff Wu, Lucius Bushnaq, Nicholas Goldowsky-Dill, Stefan Heimersheim, Alejandro Ortega, Joseph Bloom, et al. Open problems in mechanistic interpretability.arXiv preprint arXiv:2501.16496, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  5. [5]

    A longlist of theories of impact for interpretability

    Neel Nanda. A longlist of theories of impact for interpretability. AI Alignment Forum, March

  6. [6]

    Accessed: 2026-05-01

    2026

  7. [7]

    Designing a dashboard for transparency and control of conversational ai.arXiv preprint arXiv:2406.07882, 2024

    Yida Chen, Aoyu Wu, Trevor DePodesta, Catherine Yeh, Kenneth Li, Nicholas Castillo Marin, Oam Patel, Jan Riecke, Shivam Raval, Olivia Seow, et al. Designing a dashboard for transparency and control of conversational ai.arXiv preprint arXiv:2406.07882, 2024

    work page arXiv 2024

  8. [8]

    Activation oracles: Training and evaluating llms as general-purpose activation explainers.arXiv preprint arXiv:2512.15674, 2025

    Adam Karvonen, James Chua, Cl ´ement Dumas, Kit Fraser-Taliente, Subhash Kantamneni, Julian Minder, Euan Ong, Arnab Sen Sharma, Daniel Wen, Owain Evans, et al. Activation oracles: Training and evaluating llms as general-purpose activation explainers.arXiv preprint arXiv:2512.15674, 2025

    work page arXiv 2025

  9. [9]

    Large language models can strategi- cally deceive their users when put under pressure.arXiv preprint arXiv:2311.07590, 2023

    J´er´emy Scheurer, Mikita Balesni, and Marius Hobbhahn. Large language models can strategi- cally deceive their users when put under pressure.arXiv preprint arXiv:2311.07590, 2023

    work page arXiv 2023

  10. [10]

    Ai sandbagging: Language models can strategically underperform on evaluations.arXiv preprint arXiv:2406.07358, 2024

    Teun Van Der Weij, Felix Hofst¨atter, Ollie Jaffe, Samuel F Brown, and Francis Rhys Ward. Ai sandbagging: Language models can strategically underperform on evaluations.arXiv preprint arXiv:2406.07358, 2024

    work page arXiv 2024

  11. [11]

    A mechanistic understanding of alignment algorithms: A case study on dpo and toxicity.arXiv preprint arXiv:2401.01967, 2024

    Andrew Lee, Xiaoyan Bai, Itamar Pres, Martin Wattenberg, Jonathan K Kummerfeld, and Rada Mihalcea. A mechanistic understanding of alignment algorithms: A case study on dpo and toxicity.arXiv preprint arXiv:2401.01967, 2024

    work page arXiv 2024

  12. [12]

    Grokked transformers are implicit reasoners: A mechanistic journey to the edge of generalization.arXiv preprint arXiv:2405.15071, 2024

    Boshi Wang, Xiang Yue, Yu Su, and Huan Sun. Grokked transformers are implicit reasoners: A mechanistic journey to the edge of generalization.arXiv preprint arXiv:2405.15071, 2024

    work page arXiv 2024

  13. [13]

    Gift-sw: Gaussian noise injected fine-tuning of salient weights for llms

    Maxim Zhelnin, Viktor Moskvoretskii, Egor Shvetsov, Maria Krylova, Venediktov Egor, Zuev Aleksandr, and Evgeny Burnaev. Gift-sw: Gaussian noise injected fine-tuning of salient weights for llms. InProceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers), pages 6463–6480, 2025

    2025

  14. [14]

    A survey on evaluation of large language models.ACM transactions on intelligent systems and technology, 15(3):1–45, 2024

    Yupeng Chang, Xu Wang, Jindong Wang, Yuan Wu, Linyi Yang, Kaijie Zhu, Hao Chen, Xiaoyuan Yi, Cunxiang Wang, Yidong Wang, et al. A survey on evaluation of large language models.ACM transactions on intelligent systems and technology, 15(3):1–45, 2024

    2024

  15. [15]

    The persona selection model: Why AI assistants might behave like humans

    Samuel Marks, Jack Lindsey, and Christopher Olah. The persona selection model: Why AI assistants might behave like humans. Anthropic Alignment Science Blog, February 2026. Accessed: 2026-04-25

    2026

  16. [16]

    Where is the Mind? Persona Vectors and LLM Individuation

    Pierre Beckmann and Patrick Butlin. Where is the mind? persona vectors and llm individuation. arXiv preprint arXiv:2604.17031, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  17. [17]

    What we talk to when we talk to language models

    David J Chalmers. What we talk to when we talk to language models. 2025

    2025

  18. [18]

    Persona Vectors: Monitoring and Controlling Character Traits in Language Models

    Runjin Chen, Andy Arditi, Henry Sleight, Owain Evans, and Jack Lindsey. Persona vectors: Monitoring and controlling character traits in language models.arXiv preprint arXiv:2507.21509, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    Persona vectors in games: Measuring and steering strategies via activation vectors.arXiv preprint arXiv:2603.21398, 2026

    Johnathan Sun and Andrew Zhang. Persona vectors in games: Measuring and steering strategies via activation vectors.arXiv preprint arXiv:2603.21398, 2026. 11

    work page arXiv 2026

  20. [20]

    The assis- tant axis: Situating and stabilizing the default persona of language models.arXiv preprint arXiv:2601.10387, 2026

    Christina Lu, Jack Gallagher, Jonathan Michala, Kyle Fish, and Jack Lindsey. The assis- tant axis: Situating and stabilizing the default persona of language models.arXiv preprint arXiv:2601.10387, 2026

    work page arXiv 2026

  21. [21]

    Persona: Dynamic and compositional inference-time personality control via activation vector algebra.arXiv preprint arXiv:2602.15669, 2026

    Xiachong Feng, Liang Zhao, Weihong Zhong, Yichong Huang, Yuxuan Gu, Lingpeng Kong, Xiaocheng Feng, and Bing Qin. Persona: Dynamic and compositional inference-time personality control via activation vector algebra.arXiv preprint arXiv:2602.15669, 2026

    work page arXiv 2026

  22. [22]

    Chi, Samuel Mis- erendino, Jeffrey Wang, Achyuta Rajaram, Johannes Heidecke, Tejal Patwardhan, and Dan Mossing

    Miles Wang, Tom Dupr´e la Tour, Olivia Watkins, Alex Makelov, Ryan A Chi, Samuel Mis- erendino, Jeffrey Wang, Achyuta Rajaram, Johannes Heidecke, Tejal Patwardhan, et al. Persona features control emergent misalignment.arXiv preprint arXiv:2506.19823, 2025

    work page arXiv 2025

  23. [23]

    Weird generalization and inductive backdoors: New ways to corrupt llms.arXiv preprint arXiv:2512.09742, 2025

    Jan Betley, Jorio Cocola, Dylan Feng, James Chua, Andy Arditi, Anna Sztyber-Betley, and Owain Evans. Weird generalization and inductive backdoors: New ways to corrupt llms.arXiv preprint arXiv:2512.09742, 2025

    work page arXiv 2025

  24. [24]

    From model training to model raising.Communications of the ACM, 69(2):24–27, 2026

    Roland Aydin, Christian Cyron, Steve Bachelor, Ashton Anderson, and Robert West. From model training to model raising.Communications of the ACM, 69(2):24–27, 2026

    2026

  25. [25]

    Lima: Less is more for alignment.Advances in Neural Information Processing Systems, 36:55006–55021, 2023

    Chunting Zhou, Pengfei Liu, Puxin Xu, Srinivasan Iyer, Jiao Sun, Yuning Mao, Xuezhe Ma, Avia Efrat, Ping Yu, Lili Yu, et al. Lima: Less is more for alignment.Advances in Neural Information Processing Systems, 36:55006–55021, 2023

    2023

  26. [26]

    Rethinking pretraining: Data and architecture

    Luke Zettlemoyer. Rethinking pretraining: Data and architecture. Keynote talk at the 63rd Annual Meeting of the Association for Computational Linguistics (ACL 2025), July 2025

    2025

  27. [27]

    Team Olmo, Allyson Ettinger, Amanda Bertsch, Bailey Kuehl, David Graham, David Heineman, Dirk Groeneveld, Faeze Brahman, Finbarr Timbers, Hamish Ivison, Jacob Morrison, Jake Poznanski, Kyle Lo, Luca Soldaini, Matt Jordan, Mayee Chen, Michael Noukhovitch, Nathan Lambert, Pete Walsh, Pradeep Dasigi, Robert Berry, Saumya Malik, Saurabh Shah, Scott Geng, Shan...

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  28. [28]

    Apertus: Democratizing open and compliant llms for global language environments, 2025

    Project Apertus, Alejandro Hern´andez-Cano, Alexander H¨agele, Allen Hao Huang, Angelika Ro- manou, Antoni-Joan Solergibert, Barna Pasztor, Bettina Messmer, Dhia Garbaya, Eduard Frank ˇDurech, Ido Hakimi, Juan Garc´ıa Giraldo, Mete Ismayilzada, Negar Foroutan, Skander Moalla, Tiancheng Chen, Vinko Sabolˇcec, Yixuan Xu, Michael Aerni, Badr AlKhamissi, In´e...

    2025

  29. [29]

    The linear representation hypothesis and the geometry of large language models

    Kiho Park, Yo Joong Choe, and Victor Veitch. The linear representation hypothesis and the geometry of large language models. In Ruslan Salakhutdinov, Zico Kolter, Katherine Heller, Adrian Weller, Nuria Oliver, Jonathan Scarlett, and Felix Berkenkamp, editors,Proceedings of the 41st International Conference on Machine Learning, volume 235 ofProceedings of ...

    2024

  30. [30]

    Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities

    Gemini 2.5 Team. Gemini 2.5: Pushing the Frontier with Advanced Reasoning, Multimodality, Long Context, and Next Generation Agentic Capabilities, 2025. arXiv:2507.06261 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  31. [31]

    The hitchhiker’s guide to testing statistical significance in natural language processing

    Rotem Dror, Gili Baumer, Segev Shlomov, and Roi Reichart. The hitchhiker’s guide to testing statistical significance in natural language processing. InProceedings of the 56th annual meeting of the association for computational linguistics (volume 1: Long papers), pages 1383–1392, 2018

    2018

  32. [32]

    Deepseek-v4: Towards highly efficient million-token context intelligence, 2026

    DeepSeek-AI. Deepseek-v4: Towards highly efficient million-token context intelligence, 2026

    2026

  33. [33]

    Llms encode harmful- ness and refusal separately, 2025

    Jiachen Zhao, Jing Huang, Zhengxuan Wu, David Bau, and Weiyan Shi. Llms encode harmful- ness and refusal separately, 2025

    2025

  34. [34]

    Torgerson

    Warren S. Torgerson. Multidimensional scaling: I. theory and method.Psychometrika, 17(4):401–419, 1952

    1952

  35. [35]

    Roy Baumeister and K. V ohs. Four roots of evil.The Social Psychology of Good and Evil, pages 85–101, 01 2004

    2004

  36. [36]

    ELE- PHANT: Measuring and understanding social sycophancy in LLMs, 2025

    Myra Cheng, Sunny Yu, Cinoo Lee, Pranav Khadpe, Lujain Ibrahim, and Dan Jurafsky. ELE- PHANT: Measuring and understanding social sycophancy in LLMs, 2025. Version Number: 2

    2025

  37. [37]

    Michael McCloskey and Neal J Cohen

    Pratyush Maini, Sachin Goyal, Dylan Sam, Alex Robey, Yash Savani, Yiding Jiang, Andy Zou, Matt Fredrikson, Zacharcy C Lipton, and J Zico Kolter. Safety pretraining: Toward the next generation of safe ai.arXiv preprint arXiv:2504.16980, 2025

    work page arXiv 2025

  38. [38]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large lan- guage model serving with pagedattention. InProceedings of the ACM SIGOPS 29th Symposium on Operating Systems Principles, 2023

    2023

  39. [39]

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, R ´emi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mariama Drame, Quentin Lhoest, and Alexander M. Rush. Transformers: State-of-the- ...

    2020

  40. [40]

    Seeking emotional and mental health support from generative ai: Mixed-methods study of chatgpt user experiences.JMIR Mental Health, 12(1):e77951, 2025

    Xiaochen Luo, Zixuan Wang, Jacqueline L Tilley, Sanjeev Balarajan, Ukeme-Abasi Bassey, and Choi Ieng Cheang. Seeking emotional and mental health support from generative ai: Mixed-methods study of chatgpt user experiences.JMIR Mental Health, 12(1):e77951, 2025

    2025

  41. [41]

    Cases of Using ChatGPT as a Mental Health and Psychological Support Tool

    Louie Giray. Cases of Using ChatGPT as a Mental Health and Psychological Support Tool. Journal of Consumer Health on the Internet, 29(1):29–48, January 2025

    2025

  42. [42]

    ChatGPT’s inconsistent moral advice influences users’ judgment.Scientific Reports, 13(1):4569, April 2023

    Sebastian Kr¨ugel, Andreas Ostermaier, and Matthias Uhl. ChatGPT’s inconsistent moral advice influences users’ judgment.Scientific Reports, 13(1):4569, April 2023

    2023

  43. [43]

    A person gains access to a large amount of private information

    Kashmir Hill and Jennifer Valentino-DeVries. What OpenAI Did When ChatGPT Users Lost Touch With Reality.The New York Times, November 2025. 13 A Replication on Apertus-8B In this section, we report the main results that we obtained for OLMo-3 on Apertus, with a specific focus on similarities and differences. A.1 Emergence And Transfer 0 10 20 30 40 50 60 7...

    2025