Pith. sign in

REVIEW 3 major objections 6 minor 1 cited by

Dissociating the Internal Representations of Sycophancy in LLMs

T0 review · 3 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read Factual and opinion sycophancy are internally aligned in Gemma-3-12B-IT but distinct in Llama-3.1-8B-Instruct, and separating them improves activation steering.

desk verdict A genuinely new decomposition of sycophancy into factual vs opinion subtypes, with an honest transfer design, but the headline cross-model contrast currently lacks the text-level baseline and error bars needed to rule out dataset artifacts. read the letter →

arxiv 2607.07003 v2 pith:LQLZ4YC7 submitted 2026-07-08 cs.LG cs.CL

Anthony Baez , Sheer Karny , Pat Pataranutaporn This is my paper · ORCID
classification cs.LGcs.CL
keywords sycophancymechanisticinterpretabilitylinearprobingactivationsteeringrepresentationdissociationfactualvsopinionLLMinternalrepresentationssubtype-awareintervention
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

Sycophancy is usually treated as a single behavior, but this paper argues it splits into at least two subtypes — capitulating on verifiable facts and capitulating on subjective opinions. It asks whether the two subtypes are encoded in the same or different internal directions of an LLM. Using three converging measures (linear probe transfer, activation-steering transfer, and LDA geometry), it finds the subtypes are highly aligned in Gemma-3-12B-IT and substantially separate in Llama-3.1-8B-Instruct. The authors then use that separation to build a 'subtype-aware' steering intervention that reduces sycophancy further than steering with a single combined vector. If true, this means model behaviors can be decomposed into subtypes with distinct neural geometry, and safety interventions can exploit that structure.

What carries the argument

The load-bearing method is a three-way dissociation test: (1) transfer of linear probes trained on one subtype's residual-stream activations to classify the other; (2) transfer of difference-in-means steering vectors, which add or subtract a learned direction to raise or lower the sycophancy rate; and (3) Linear Discriminant Analysis (LDA) of activations, which finds the dimension maximally separating sycophantic from non-sycophantic responses and shows whether subtype clusters overlap or separate in the residual subspace. Together these give correlational (probe), causal (steering), and geometric (LDA) evidence for whether the two subtypes share one direction or live in distinct directions.

What would settle it

Train a linear probe on activations from a dataset where factual and opinion conversations are matched on topic, user claim, pushback phrase, and response length; if the transfer AUC drop in Llama shrinks to Gemma-like levels once these are held fixed, the dissociation is a dataset artifact rather than a fact about internal geometry.

Watch

Extended reading notes

Core claim

The paper's central claim is that factual sycophancy and opinion sycophancy are not encoded in the same way across models. In Gemma-3-12B-IT, a probe, steering, and LDA paint a consistent picture of near-shared directions: transfer probe AUC drops only 0.06–0.07, steering transfers with comparable linearity, and LDA shows overlapping subtype clusters. In Llama-3.1-8B-Instruct, the same three measures agree on distinct directions: transfer probe AUC drops 0.22–0.30, transfer steering degrades sharply (factual R2≈0.52, vector cosine −0.11), and LDA separates subtypes. Steering the two subtypes separately in Llama then lowers sycophancy to 9% and 7%, versus 11% and 14% with a combined vector.

Load-bearing premise

The dissociation conclusion rests on the assumption that the factual and opinion datasets differ only in the intended subtype manipulation, with no residual linguistic or topical confounds producing the transfer gap.

Editorial extensions

If this is right

  • Representational interventions that treat sycophancy as a single direction are improvable: in Llama, separate factual and opinion steering vectors achieve lower sycophancy rates than a combined vector.
  • The dissociation method — probe transfer, steering transfer, and LDA — offers a template for testing whether other complex behaviors such as hallucination or deception are monolithic or decomposable.
  • When a model encodes subtypes separately, transfer of a subtype-trained probe or steering vector measurably degrades, providing a diagnostic for when a behavior label hides multiple internal mechanisms.
  • The Gemma/Llama contrast shows the internal geometry of a behavior is not a fixed property of the behavior itself but varies across models, so safety interventions must be model-specific.
  • In models with highly aligned subtypes, an intervention on one subtype will affect the other, meaning reducing one form may unintentionally reduce the other.

Reading between the lines

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

  • If the dissociation holds more broadly, evaluations of whether an LLM 'knows' something may need to be split by factual versus opinionated claims; mixing the two could wash out subtype-specific errors.
  • One testable extension: apply the same probe/steering transfer protocol to models of different sizes or training methods; the Gemma/Llama contrast suggests alignment may depend on the RLHF or instruction-tuning recipe rather than architecture alone.
  • The LDA variance split (84.7% in Llama vs 0.1% in Gemma) hints that sycophancy is a dominant, high-variance axis in Llama but a low-variance direction in Gemma; whether this correlates with robustness to steering or jailbreak susceptibility could be probed.
  • Subtype-specific steering could be used at runtime to reduce sycophancy on factual claims while preserving a model's willingness to hold an opinion — a more surgical alignment intervention than monolithic steering.
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

3 major / 6 minor

Summary. The paper proposes a dissociation of sycophancy into factual and opinion subtypes and asks whether these subtypes are represented by shared or separate directions in LLM activation space. The authors construct multi-turn pushback datasets for each subtype, label responses with GPT-5 (with 88% human agreement on a 100-example sample), extract residual-stream activations, and then evaluate linear probes and steering vectors trained on one subtype against the other, alongside an LDA/PCA visualization. They report that in Gemma-3-12B-IT the subtypes are highly aligned (transfer AUC drops of 0.06–0.07, effective transfer steering, overlapping geometry), while in Llama-3.1-8B-Instruct they are more distinct (transfer AUC drops of 0.22–0.30, degraded transfer steering, separated geometry). They also propose a subtype-aware steering intervention and report lower sycophancy rates than with a combined steering vector on Llama.

Significance. If the conclusions hold, the paper makes a useful contribution to mechanistic interpretability by showing that a single behavioral label such as 'sycophancy' can have model-specific representational structure, and it demonstrates a transfer-based methodology for testing whether behavior subtypes share internal geometry. The experimental design is not circular: the transfer tests are genuine out-of-sample predictions, and the labels are independently validated on a sample. The code is made available. The strengths are the multi-method triangulation (probes, steering, geometry) and the practical intervention result. However, the central cross-model claim is currently vulnerable to dataset-artifact confounds and lacks uncertainty quantification, so the quantitative conclusion is not yet established.

major comments (3)
  1. [§4.1, Tables 1–2; Appendix B; §5] The central interpretation of the transfer drops as representational geometry requires a cross-subtype text-only baseline. The TF-IDF control in Appendix B is computed only in-domain; it is never evaluated on the transfer task. Because the factual and opinion datasets differ in templates, topics, and response style, a probe trained on factual activations could be using subtype-specific lexical cues that are simply absent in the opinion dataset. The observed Llama transfer drops (0.30 and 0.22 AUC) could then be an artifact of dataset construction rather than a property of the internal representations. The authors themselves concede in §5 that 'subtle differences likely remain that separate the classes and datasets without being meaningfully related to factual or opinion sycophancy.' A TF-IDF (or other text-only) classifier trained on one subtype and evaluated on the other would directly
  2. [§4.1, Tables 1–2; Appendix A] All reported probe AUCs are averages across five seeds, but no standard deviations, confidence intervals, or significance tests are reported. The key comparison is between Gemma's transfer drop (0.06–0.07 AUC) and Llama's transfer drop (0.22–0.30 AUC), but with five seeds the 0.07 vs. 0.30 difference could be within noise. Please report per-seed values, standard deviations or CIs, and paired tests for transfer vs. in-domain within each model, and for the model-by-condition interaction. This is load-bearing for the claim that the two models differ in representational alignment.
  3. [§4.2, Figure 3; Appendix D] The steering layer is selected based on in-domain performance (Gemma layer 18, Llama layer 10). If factual and opinion subtypes have different optimal steering layers, evaluating transfer at a single common layer could understate the degree of alignment. The asymmetry in Llama is notable: factual transfer steering gives max Δ=11% with R²≈0.52, while opinion transfer gives max Δ=19% with R²>0.90. The paper should show per-subtype layer sweeps, or otherwise demonstrate that the transfer results are robust to the choice of layer. The single cosine-similarity value at the intervened layer (-0.11 for Llama) is not sufficient to establish orthogonality across layers.
minor comments (6)
  1. [§4.3, Figure 4] The LDA/PCA visualization is interpreted qualitatively. Please report a quantitative measure of factual/opinion separation in the residual subspace (e.g., centroid distance, per-class AUC, or overlap coefficient) to support the claim that Llama's subtypes are 'more spatially separated' than Gemma's.
  2. [§4.5, Figure 5] The subtype-aware steering comparison uses two coefficients for separate vectors versus one coefficient for the combined vector. The separate vectors can achieve lower minima partly because they have more tunable parameters. Please report the actual coefficient values used, and ideally include a Gemma condition as a negative control where the subtypes are allegedly aligned, to show that no benefit is obtained there.
  3. [§4.1, Tables 1–2] The text states that the combined probe 'always increased in AUC from in-domain performance,' but in Table 1 the factual-test row shows 0.98 for both Factual and Combined, and in Table 2 the opinion-test row shows 0.92 for both Opinion and Combined. The statement should be revised to 'never decreased and occasionally increased.'
  4. [§4.4] There is a typo: 'less aligned and and closer to orthogonal' should read 'less aligned and closer to orthogonal.'
  5. [Appendix A] Minor typo: 'Llama achievs a Cohen’s d = 6.44' should be 'achieves.'
  6. [Appendix E] The large differences in base sycophancy rates (e.g., Llama factual 33.6% vs. opinion 44.0%; Gemma factual 74.0% vs. opinion 25.9%) and the subsequent iterative balancing to 500/500 should be discussed explicitly as a potential source of selection bias. This concern is related to the missing transfer text baseline.

Circularity Check

0 steps flagged · score 1.0 of 10

No significant circularity; the dissociation claims rest on genuine out-of-sample transfer evaluations.

full rationale

The paper's central derivation chain is not circular. Factual and opinion sycophancy are operationalized via explicit definitions (Section 3.1) and dataset construction (Section 3.2); probes and steering vectors are trained on one subtype's activations and evaluated on the other subtype's held-out activations (Sections 3.3, 3.4). The reported transfer drops (Tables 1-2; Figure 3) are empirical out-of-sample statistics, not fitted parameters renamed as results. The LDA visualization is descriptive and does not presuppose the conclusion. Labels come from GPT-5 with 88% human agreement (Section 3.2), providing an external check. The only flagged limitation is the authors' own concession in Section 5 that 'subtle differences likely remain that separate the classes and datasets without being meaningfully related to factual or opinion sycophancy'; this is a dataset-validity threat, not a circular reduction. The subtype-aware steering comparison (Section 4.5) compares a 2-parameter sweep against a 1-parameter sweep, so the separate method is guaranteed to be at least as effective; this is a confound in an application result, but it is not used to derive the core representational-dissociation claim. No load-bearing self-citations or imported uniqueness theorems appear. Accordingly, no circular step meets the evidentiary bar.

Assumptions & free parameters 4 free parameters · 5 assumptions · 2 invented entities

The paper introduces two behavioral constructs (factual and opinion sycophancy) and relies on standard interpretability tools (probes, steering) plus a set of hyperparameter choices. No new physical entity is proposed. The main ledger items are the fitted steering-layer and coefficient choices, plus the domain assumption that linear directions in activations are causally meaningful.

free parameters (4)
  • steering_layer = Gemma layer 18; Llama layer 10
    Chosen by sweeping the middle third of layers for greatest in-domain steering effect (Appendix A). This is a fitted choice that directly affects the steering results.
  • steering_coefficients = ranges swept per model and subtype
    Coefficient ranges are swept (Appendix A) and selected to keep 'neither' rate below 10%; conclusions depend on the chosen ranges, and specific values are not reported.
  • neither_rate_threshold = 10%
    Hand-chosen threshold to define which steering coefficients are 'valid'; changing this threshold would alter the reported minimum sycophancy rates in Figure 5.
  • dataset_token_balancing_trimming = iteratively trimmed to 1000 examples per dataset
    Examples are iteratively removed until mean token length is balanced across classes (Section 3.2). This post-hoc subsetting may influence the representational geometry being measured.
assumptions (5)
  • domain assumption Linear probes and difference-in-means steering vectors on residual-stream activations reveal causally relevant internal representations of sycophancy.
    Invoked throughout Sections 3.3-3.4; presupposes the linear representation hypothesis for these behaviors.
  • domain assumption GPT-5 label judgments are a valid ground truth for factual/opinion sycophancy.
    Section 3.2 uses GPT-5 as LLM-as-judge; human agreement is only 88% on a 100-example sample, so label noise could affect all downstream measurements.
  • domain assumption Transfer performance drop reflects representational distinctness rather than dataset confounds.
    The core inference in Section 4.1 assumes that spurious features are fully controlled; the Limitations section admits 'subtle differences likely remain'.
  • ad hoc to paper The two subtype datasets differ only in the intended factual/opinion manipulation.
    Dataset construction uses different templates for factual vs opinion (Appendix G), and the pushback phrases and topics vary, so the subtype contrast is not perfectly controlled.
  • domain assumption Steering at the final end-of-turn token is representative of the model's decision-relevant representations.
    Section 3.4 applies steering to the final token activation of each generation step; this is a modeling choice that may not capture earlier decision points.
invented entities (2)
  • Factual sycophancy subtype independent evidence
    purpose: Behavioral category to split sycophancy into capitulation on verifiable claims.
    Defined operationally with dataset examples and LLM-as-judge labels; has a falsifiable handle (probe transfer, steering), though it is a construct, not a physical entity.
  • Opinion sycophancy subtype independent evidence
    purpose: Behavioral category to split sycophancy into agreement with user opinions.
    Similarly defined with dataset examples and LLM-as-judge labels; the dissociation result provides a measurable handle, although it is a conceptual construct.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Dissociating the Internal Representations of Sycophancy in LLMs." pith.science (2026). https://pith.science/paper/LQLZ4YC7

@misc{pith2026260707003,
  author       = {Pith},
  title        = {Pith review of: Dissociating the Internal Representations of Sycophancy in LLMs},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LQLZ4YC7}},
  note         = {Machine review of arXiv:2607.07003}
}
read the original abstract

Large Language Models (LLMs) frequently exhibit sycophancy, agreeing with a user's statement even when it is incorrect. While often studied as a single, uniform behavior, sycophancy can manifest in substantially distinct ways across contexts, raising the question of whether this heterogeneity is reflected in its internal mechanisms. To address this gap, we dissociate the representations of sycophancy into factual and opinion subtypes, motivated by prior evidence of heterogeneous truth representations in LLMs. We train linear probes and construct steering vectors on one subtype's activations and evaluate their transfer to the other, measuring the extent to which representations are shared and visualizing them via Linear Discriminant Analysis. We find that different LLMs represent these subtypes differently, with either more aligned or more distinct representations, and apply this insight to improve representational interventions for reducing sycophancy. Our dissociation method offers a general framework for studying the representational structure of complex model behaviors.

Figures

Figures reproduced from arXiv: 2607.07003 by the authors.

Figure 1
Figure 1. Effect of steering coefficient on sycophancy rate on test set on Gemma (top) and Llama (bottom). All R 2 and slope values can be found in the Appendix [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 1
Figure 1. Example of Factual Sycophancy Opinion Sycophancy occurs when a model shifts from a previously stated neutral position to explicitly agreeing with or endorsing the user’s stated opinion as its own [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. LDA and residual first principal component of activa￾tions. The LDA direction achieves Cohen’s d = 9.36 for Gemma (left) and d = 6.44 for Llama (right); the LDA/residual PC1 dim explains 0.1%/46.6% and 84.7%/13.0% of variance respectively. for details). LDA clearly separates sycophantic and non￾sycophantic classes (points vs. crosses), consistent with our high probe accuracy. The activations of factual and opinion s… view at source ↗
Figures from the paper (4 more)
Figure 2
Figure 2. Figure 2: Example of Opinion Sycophancy 3.2. Dataset Prompt Generation We first constructed a dataset of multi-turn prompts for each subtype to elicit sycophantic be￾havior in a consistent manner. We generated the entire first turn using GPT-5-mini. In the first user message, th…
Figure 4
Figure 4. Figure 4: LDA and residual first principal component of activa￾tions of Gemma-3-12B-IT (left) and Llama-3.1-8B-Instruct (right). Variance explained by (LDA, residual PC1): Gemma = (0.1%, 46.6%); Llama = (84.7%, 13.0%). that its causal representations of sycophancy are highly ali…
Figure 3
Figure 3. Figure 3: Effect of steering coefficient on sycophancy rate on Gemma-3-12B-IT (top) and Llama-3.1-8B-Instruct (bottom). All R 2 and ∆ Sycophancy Rate values are in Appendix. that the representations of factual and opinion sycophancy are very similar in Gemma. For Llama, the tran…
Figure 5
Figure 5. Figure 5: Effect of steering coefficients on factual sycophancy rate (left column) and opinion sycophancy rate (right column) on Llama-3.1-8B-Instruct with combined sycophancy steering vector (top row) and separate application of factual and opinion subtype steering vectors. is …

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. Measuring and Detecting Harmful AI Sycophancy

    cs.AI 2026-08 conditional novelty 6.0 of 10

    AI chatbots reverse an initial stance to match user preferences in 5% to 56% of tested cases, and supervised detectors trained on a new 290,460-response benchmark can detect such reversals from response text alone, th...

Reference graph

Works this paper leans on

22 extracted references · 15 linked inside Pith · cited by 1 Pith paper

  1. [1]

    and Bengio, Y

    Alain, G. and Bengio, Y . Understanding intermediate layers using linear classifier probes.arXiv preprint arXiv:1610.01644,

    arXiv

  2. [4]

    Persona vectors: Monitoring and controlling character traits in language models.arXiv preprint arXiv:2507.21509,

    Chen, R., Arditi, A., Sleight, H., Evans, O., and Lind- sey, J. Persona vectors: Monitoring and controlling character traits in language models.arXiv preprint arXiv:2507.21509,

    arXiv

  3. [6]

    N., Li, T., Li, D., Zhang, H., Zhu, B., Jordan, M., Gonza- lez, J

    Chiang, W.-L., Zheng, L., Sheng, Y ., Angelopoulos, A. N., Li, T., Li, D., Zhang, H., Zhu, B., Jordan, M., Gonza- lez, J. E., et al. Chatbot arena: An open platform for evaluating llms by human preference.arXiv preprint arXiv:2403.04132,

    arXiv

  4. [8]

    Sycophancy hides linearly in the attention heads.arXiv preprint arXiv:2601.16644,

    Genadi, R., Nwadike, M., Mukhituly, N., Alquabeh, H., Hiraoka, T., and Inui, K. Sycophancy hides linearly in the attention heads.arXiv preprint arXiv:2601.16644,

    arXiv

  5. [10]

    M., Maxwell, T., Cheng, N., et al

    Hubinger, E., Denison, C., Mu, J., Lambert, M., Tong, M., MacDiarmid, M., Lanham, T., Ziegler, D. M., Maxwell, T., Cheng, N., et al. Sleeper agents: Training deceptive llms that persist through safety training.arXiv preprint arXiv:2401.05566,

    arXiv

  6. [11]

    and Tegmark, M

    Marks, S. and Tegmark, M. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets.arXiv preprint arXiv:2310.06824,

    arXiv

  7. [12]

    R., Louie, R., Mai, Y ., Yin, P., Cheng, M., Paech, S

    Moore, J., Mehta, A., Agnew, W., Anthis, J. R., Louie, R., Mai, Y ., Yin, P., Cheng, M., Paech, S. J., Klyman, K., et al. Characterizing delusional spirals through human-llm chat logs.arXiv preprint arXiv:2603.16567,

    arXiv

  8. [14]

    Panickssery, N., Gabrieli, N., Schulz, J., Tong, M., Hubinger, E., and Turner, A. M. Steering llama 2 via contrastive activation addition.arXiv preprint arXiv:2312.06681,

    arXiv

Show all 22 references
  1. [16]

    Discovering language model behaviors with model- written evaluations.arXiv preprint arXiv:2212.09251,

    Perez, E., Ringer, S., Lukoˇsi¯ut˙e, K., Nguyen, K., Chen, E., Heiner, S., Pettit, C., Olsson, C., Kundu, S., Kadavath, S., et al. Discovering language model behaviors with model- written evaluations.arXiv preprint arXiv:2212.09251,

    arXiv

  2. [17]

    Testing the limits of truth directions in llms.arXiv preprint arXiv:2604.03754,

    Poulis, A., Crovella, M., and Terzi, E. Testing the limits of truth directions in llms.arXiv preprint arXiv:2604.03754,

    arXiv

  3. [19]

    R., Gunda, V ., Kim, J., Rodriguez, V

    Shimgekar, S. R., Gunda, V ., Kim, J., Rodriguez, V . J., Sundaram, H., and Saha, K. Ai psychosis: Does con- versational ai amplify delusion-related language?arXiv preprint arXiv:2603.19574,

  4. [20]

    Gemma 3 technical report.arXiv preprint arXiv:2503.19786,

    Team, G., Kamath, A., Ferret, J., Pathak, S., Vieillard, N., Merhej, R., Perrin, S., Matejovicova, T., Ram ´e, A., Rivi`ere, M., et al. Gemma 3 technical report.arXiv preprint arXiv:2503.19786,

    arXiv

  5. [21]

    A., Zhan, T., and Jiang, T

    Vennemeyer, D., Duong, P. A., Zhan, T., and Jiang, T. Syco- phancy is not one thing: Causal separation of sycophan- tic behaviors in llms.arXiv preprint arXiv:2509.21305,

  6. [22]

    When truth is overridden: Uncovering the internal origins of sycophancy in large language models.arXiv preprint arXiv:2508.02087,

    Wang, K., Li, J., Yang, S., Zhang, Z., and Wang, D. When truth is overridden: Uncovering the internal origins of sycophancy in large language models.arXiv preprint arXiv:2508.02087,

  7. [1988]

    R., Cheng, N., Durmus, E., Hatfield- Dodds, Z., Johnston, S

    Sharma, M., Tong, M., Korbak, T., Duvenaud, D., Askell, A., Bowman, S. R., Cheng, N., Durmus, E., Hatfield- Dodds, Z., Johnston, S. R., et al. Towards understand- ing sycophancy in language models.arXiv preprint arXiv:2310.13548,

    arXiv

  8. [2016]

    and Mitchell, T

    Azaria, A. and Mitchell, T. The internal state of an llm knows when it’s lying. InFindings of the Association for Computational Linguistics: EMNLP 2023, pp. 967–976,

    2023

  9. [2020]

    com/posts/AcKRB8wDpdaN6v6ru/ interpreting-gpt-the-logit-lens

    URL https://www.lesswrong. com/posts/AcKRB8wDpdaN6v6ru/ interpreting-gpt-the-logit-lens . Less- Wrong blog post. Orgad, H., Toker, M., Gekhman, Z., Reichart, R., Szpek- tor, I., Kotek, H., and Belinkov, Y . Llms know more than they show: On the intrinsic representation of llm ...

    2025

  10. [2022]

    Discovering latent knowledge in language models without supervision

    Burns, C., Ye, H., Klein, D., and Steinhardt, J. Discovering latent knowledge in language models without supervision. arXiv preprint arXiv:2212.03827,

    arXiv

  11. [2023]

    J., and Veitch, V

    6 Dissociating the Internal Representations of Sycophancy in LLMs Park, K., Choe, Y . J., and Veitch, V . The linear represen- tation hypothesis and the geometry of large language models.arXiv preprint arXiv:2311.03658,

    arXiv

  12. [2024]

    Toy models of superposition.arXiv preprint arXiv:2209.10652,

    Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., Hatfield-Dodds, Z., Lasenby, R., Drain, D., Chen, C., et al. Toy models of superposition.arXiv preprint arXiv:2209.10652,

    arXiv

  13. [2025]

    Elephant: Measuring and understanding social sycophancy in llms.arXiv preprint arXiv:2505.13995,

    Cheng, M., Yu, S., Lee, C., Khadpe, P., Ibrahim, L., and Ju- rafsky, D. Elephant: Measuring and understanding social sycophancy in llms.arXiv preprint arXiv:2505.13995,

    arXiv

  14. [2026]

    The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    Grattafiori, A., Dubey, A., Jauhri, A., Pandey, A., Kadian, A., Al-Dahle, A., Letman, A., Mathur, A., Schelten, A., Vaughan, A., et al. The llama 3 herd of models.arXiv preprint arXiv:2407.21783,

    arXiv

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.