Pith. sign in

REVIEW 2 major objections 5 minor 41 references

Evaluating Steering Techniques using Human Similarity Judgments

T0 review · 2 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper shows that prompt-based steering beats task vectors, DiffMean, and sparse autoencoders on both accuracy and human alignment in a triadic similarity task, and that LLMs favor 'kind' similarity over 'size' from the start.

desk verdict A useful cognitive-science evaluation of steering methods on kind vs size, but the alignment metric and the abstract's overreach make the size-alignment result provisional. read the letter →

arxiv 2505.19333 v1 pith:IKJMA223 submitted 2025-05-25 cs.AI

Zach Studdiford , Timothy T. Rogers , Siddharth Suresh , Kushin Mukherjee This is my paper
classification cs.AI
keywords LLMsteeringtriadicsimilarityjudgmentrepresentationalalignmentprompt-basedtaskvectorssparseautoencodersProcrustescorrelationsizevskind
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

The paper argues that evaluations of LLM steering should ask not just whether a steered model gets the right answer, but whether its patterns of similarity judgments match human patterns. To test this, it runs a triadic similarity task in which humans and LLMs must pick which of two objects is most similar to a target in terms of either size or kind. Prompting, task vectors, DiffMean, and sparse autoencoders are compared on two Gemma models. The paper finds that prompt-based steering wins on both accuracy and alignment with human embeddings, and that LLMs arrive with a default affinity for kind similarity that no method fully overcomes, especially for size. If correct, it offers a cognitive benchmark for steering and suggests that the dominant axis of a model's concept space is set before any steering happens.

What carries the argument

The central object is the crowd-kernel embedding: each set of triadic judgments—human or model, under each steering condition—is converted into a two-dimensional point configuration by minimizing a triplet-loss function (Tamuz et al., 2011). The comparison then relies on squared Procrustes correlation (Gower, 1975), which measures how well pairwise distances in the model's embedding correspond to those in the human embedding after allowing arbitrary rotations, reflections, translations, and uniform scaling. The steering methods themselves are the independent variables: prompts that state the dimension, task vectors extracted from the residual stream at the final '+' token of an in-context prompt, DiffMean vectors computed as the average difference between residual activations on size vs kind prompts, and sparse-autoencoder feature directions. The two-dimensional projection and the affine-invariant comparison are what allow a fair, human-centered evaluation of steering.

What would settle it

Compute agreement between model and human choices directly on the triplet trials (e.g., percentage of trials where the model matches the modal human response), separately for size and kind. If prompt-steered models show high per-trial agreement on size despite low 2D Procrustes $r^2$, the 2D embedding is masking a real alignment; if per-trial agreement is also low, the representational-mismatch claim is supported.

Watch

Extended reading notes

Core claim

On the Round Things Dataset, which contains 46 round objects varying in kind (artifact vs plant) and continuous size, the paper reports that prompting produces higher triadic judgment accuracy than task vectors, DiffMean, and SAEs, and its derived embeddings are also closer, by squared Procrustes correlation, to embeddings derived from human judgments. The results show a dissociation between competence and alignment: prompt-steered models can predict the ground-truth size ordering nearly as well as humans, yet their size-judgment embeddings align only weakly with human size embeddings. The paper attributes this to LLMs being able to isolate a task dimension cleanly, whereas human judgments 'leak' information from the irrelevant dimension (kind when judging size, and vice versa). Neutral prompts align with human kind embeddings ($R^2 = 0.50$) but essentially not with human size embeddings ($R^2 = 0.02$), which the paper interprets as a privileged 'kind' axis in the model's representational space before steering.

Load-bearing premise

The load-bearing premise is that the alignment comparison is sensitive to what humans actually use: if projecting judgments into two dimensions discards the size axis, or if Procrustes alignment absorbs real geometric differences, then the finding that no steering method aligns with human size judgments could be an artifact rather than a fact.

Editorial extensions

If this is right

  • If prompting is indeed the superior steering route, practitioners should prefer carefully worded prompts over internal interventions such as task vectors, DiffMean, and SAE feature addition when the goal is human-like behavior.
  • The default kind axis implies that LLMs need no steering to judge kind similarity, so future steering work should focus on dimensions like size that are not privileged; otherwise, accuracy gains may hide a persistent representational bias.
  • Accuracy and alignment can diverge sharply: a model can match human accuracy on size while its internal organization is unlike a human's, so benchmarks that report only task performance will miss this.
  • The 15-example in-context variant sometimes lowers alignment relative to the zero-shot prompt, so adding demonstrations is not automatically beneficial for alignment with human geometry.

Reading between the lines

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

  • If the kind bias is fixed during pre-training, then the most direct way to make LLMs human-like on size may be to change the training distribution or use fine-tuning on human size judgments, not inference-time vector steering; this is not tested in the paper.
  • The leaked-dimension account predicts a specific pattern in human choices: on size trials, errors should tend toward the kind-similar item, and on kind trials, errors toward the size-similar item; a direct test on the existing triplet data could support or complicate this explanation.
  • Because the Procrustes alignment is affine-invariant, it ignores whether the model's axes are stretched differently than a human's; a metric that preserves axis scales, or one computed at higher dimensionality, might rank the steering methods differently.
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

2 major / 5 minor

Summary. The paper evaluates four LLM steering techniques (prompting, task vectors, DiffMean, and sparse autoencoders) on gemma2-27b and gemma2-9b using a triadic similarity judgment task drawn from the Round Things Dataset. For 'size' and 'kind' dimensions, the authors measure task accuracy against ground-truth labels and representational alignment to human embeddings estimated from crowd-sourced judgments, using 2D crowd-kernel embeddings and squared Procrustes correlation. They report that prompting outperforms other methods in accuracy, that prompt-based methods are generally better aligned to human representations, that neutral LLM representations are biased toward 'kind' over 'size', and that no steering method produces size embeddings well-aligned with humans. They interpret the size misalignment as reflecting humans' tendency to leak kind information into size judgments, which LLMs do not.

Significance. If the results hold, the paper provides a valuable cognitive-science-grounded benchmark for steering evaluation and supports the practical conclusion that simple prompting is competitive with or superior to more complex representational interventions. A clear strength is the use of a human-derived similarity space as an external benchmark, and the transparent reporting of regression coefficients and p-values for accuracy differences is commendable. The paper also honestly acknowledges limitations in model and method coverage. However, the headline claims about alignment rest on a single metric whose construct validity needs additional support, and one central claim, that prompting wins on both accuracy and alignment, is overstated relative to the reported statistics.

major comments (2)
  1. [Abstract; Section 4.2; Section 5] The abstract and conclusion claim that prompt-based steering outperforms other methods 'both in terms of steering accuracy and model-to-human alignment,' but Section 4.2 states that for size judgments only the zero-shot size prompt has a higher representational alignment than other methods. This implies that not all prompt-based conditions show significantly higher alignment for size, so the 'both accuracy and alignment' conclusion is not supported for the full prompting family. Please revise the abstract and conclusion to specify the conditions under which alignment improvements are significant, and report the corresponding statistics.
  2. [Section 3.3; Figures 1 and 5; Section 4.3] The squared Procrustes correlation between full 2D crowd-kernel embeddings is the sole basis for the alignment results and for the central negative claim in Section 4.3 that no steering method yields size alignments. Because the crowd-kernel algorithm fits two dimensions to all judged-similarity variance, the second dimension of each embedding is not necessarily task-relevant. For size, a roughly one-dimensional ground-truth scale, several model embeddings in Figure 5 (e.g., prompt_size_icl_9b, prompt_size_9b, sae_size_9b) have near-zero variance on Dim2, whereas human size embeddings show a second dimension that the paper attributes to kind leakage. After optimal Procrustes alignment, r2 is computed jointly over both dimensions, so low size r2 may reflect mismatch in an unconstrained second axis rather than poor alignment of the size axis itself. Conversely, for kind, a binary category split could inflate r2 even when within-category structure differs. To support the strong claims, the authors should report the variance explained by each embedding dimension per condition, compute alignment restricted to the task-relevant dimension (e.g., the coordinate along the size gradient after Procrustes), and check robustness to embedding dimensionality (e.g., 3D or 4D). Without such analyses, the headline 'struggled with size alignment' and the method ranking for alignment do not follow.
minor comments (5)
  1. [Appendix A.2] The text contains an unresolved citation placeholder 'GemmaScope (?)'; please supply the appropriate reference.
  2. [References] The reference for Zou et al. (2024) cites arXiv:2501.17148, which is actually the AxBench paper by Wu et al. (2025); please correct this citation to avoid misattribution.
  3. [Appendix A.3] The appendix heading 'A.3' is followed by a raw list of condition names and the Figure 4 caption without an introductory sentence; please add a brief description of what the pairwise Procrustes correlation figure shows.
  4. [Section 4.2] The sentence 'only the zero-shot size prompt has a higher representational alignment than other methods' is ambiguous; please clarify whether the comparison is within size conditions, within prompting conditions, or across all conditions.
  5. [Throughout] There are several typos and formatting issues, including 'evaluted' in Section 3.2, 'privilged' in Section 5, and the figure axis label 'MnDiam' should be expanded in captions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the comparison against human judgments is externally grounded and no claimed result reduces to a fitted parameter or self-citation.

full rationale

The paper's derivation chain is self-contained with respect to its target claims. Human embeddings are estimated from independent human triplet judgments, and model embeddings are estimated from LLM triplet judgments using the same crowd-kernel procedure; the squared Procrustes correlation is then computed between these independently estimated spaces (Section 3.3). The alignment result is therefore an external benchmark, not a quantity forced by the paper's own equations. Steering-method hyperparameters, such as layer selection for task vectors, DiffMean, and SAEs, are chosen on held-out accuracy rather than on alignment, and the alignment comparison uses judgments collected under each method, so no prediction reduces to a fit. Self-citations to the Round Things Dataset (Giallanza et al., 2024), triadic-judgment methodology, and prior alignment work are used as inputs or related work, not as a uniqueness argument or as definitional support for the present conclusions. The skeptical concern about the 2D Procrustes r2 metric is a construct-validity issue, not circularity, because the human-benchmark target is independent of the model outputs being scored.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central results depend on an off-the-shelf embedding and alignment pipeline, on external human data not included in the preprint, and on an untested interpretive assumption about neutral prompts. Layer selection and embedding dimension are data-driven choices that affect the comparisons. None of these are fatal, but they mean the results are only as strong as the unshipped pipeline.

free parameters (3)
  • steering layer l*_d for Task Vector, DiffMean, and SAE = not reported; selected per model, dimension, and method on held-out accuracy
    Appendix A.2 states that the best-performing layer for each dimension is used for downstream embedding extraction. This data-driven choice can change both accuracy and alignment results, and it is not fixed in advance.
  • in-context example count n = 15 triplets
    Appendix A.2 sets n=15 for in-context prompts and for task-vector induction. This value is chosen by hand; no sensitivity analysis is reported, so the relative ranking of methods could depend on it.
  • embedding dimension = 2
    Section 3.3 constructs two-dimensional embeddings for both humans and models; the Procrustes alignment results depend on this dimensionality choice, which is inherited from prior work but not validated here.
assumptions (5)
  • standard math Crowd-kernel triplet embedding estimation returns a reliable 2D geometric map from at least 2,500 triplet judgments per condition.
    Used in Section 3.3 to estimate embeddings; reliability of 2D recovery for 46 concepts is assumed from prior work (Jamieson et al., 2015), not verified here.
  • standard math Squared Procrustes correlation after affine alignment is a valid measure of model-to-human representational alignment.
    Section 3.3 Alignment Analysis uses this to make all headline comparisons; the choice of tolerance for affine transforms can change which differences count as misalignment.
  • domain assumption Human similarity judgments for the Round Things Dataset are a valid benchmark for human semantic organization.
    Human data are external (Giallanza et al., 2024); the paper gives no participant details or raw judgments, yet all alignment conclusions are relative to these embeddings.
  • ad hoc to paper Neutral prompt behavior reflects the model's default representational bias.
    Section 4.1 interprets high neutral-prompt accuracy for kind and low accuracy for size as a privileged kind axis; the ambiguous neutral prompt may itself cue taxonomic responding, so the interpretation is not uniquely determined.
  • domain assumption Triplet construction where kind and size answers are mutually exclusive yields a fair accuracy measure.
    Section 3.1 states triplets were constructed so kind judgments are mutually exclusive from size judgments, simplifying the accuracy metric but removing natural cases where both dimensions compete, which may be exactly where human alignment fails.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Evaluating Steering Techniques using Human Similarity Judgments." pith.science (2026). https://pith.science/paper/IKJMA223

@misc{pith2026250519333,
  author       = {Pith},
  title        = {Pith review of: Evaluating Steering Techniques using Human Similarity Judgments},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/IKJMA223}},
  note         = {Machine review of arXiv:2505.19333}
}
read the original abstract

Current evaluations of Large Language Model (LLM) steering techniques focus on task-specific performance, overlooking how well steered representations align with human cognition. Using a well-established triadic similarity judgment task, we assessed steered LLMs on their ability to flexibly judge similarity between concepts based on size or kind. We found that prompt-based steering methods outperformed other methods both in terms of steering accuracy and model-to-human alignment. We also found LLMs were biased towards 'kind' similarity and struggled with 'size' alignment. This evaluation approach, grounded in human cognition, adds further support to the efficacy of prompt-based steering and reveals privileged representational axes in LLMs prior to steering.

Figures

Figures reproduced from arXiv: 2505.19333 by the authors.

Figure 1
Figure 1. Representational geometry of the concepts in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Steering accuracy (top row) and alignment of [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Overview of the triadic judgment task in humans (left) and LLM steering methods (right). [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Full procrustes correlations for all methods [PITH_FULL_IMAGE:figures/full_fig_p009_4.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 27 canonical work pages

  1. [1]

    URL: " 'urlintro :=

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type volume year eprint doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRINGS urlintro eprinturl eprintpr...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, et al. 2023. https://transformer-circuits.pub/2023/monosemantic-features Towards monosemanticity: Decomposing language models with dictionary learning . Transformer Circuits Thread

    work page 2023

  4. [4]

    Jonathan D Cohen, Kevin Dunbar, and James L McClelland. 1990. On the control of automatic processes: a parallel distributed processing account of the stroop effect. Psychological review, 97(3):332

    work page 1990

  5. [5]

    Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. 2023. https://doi.org/10.48550/arXiv.2309.08600 Sparse autoencoders find highly interpretable features in language models . arXiv preprint arXiv:2309.08600

    arXiv doi:10.48550/arxiv.2309.08600 2023

  6. [6]

    Jos De Bruin, Thomas Bourguignon, Mouhamadou Biran, Walid Saoud, Pierre Morizet-Mahoudeaux, Karine Tasso, Nicolas Chanez, Laurent Perrinet, Kathinka Evers, Claire Montfroy, et al. 2024. Strong and weak alignment of large language models with human values. Scientific Reports, 14(1):17428

    work page 2024

  7. [7]

    Leo Gao, Tom Dupré la Tour, Henk Tillman, Gabriel Goh, Rajan Troll, Alec Radford, Ilya Sutskever, Jan Leike, and Jeffrey Wu. 2024. https://doi.org/10.48550/arXiv.2406.04093 Scaling and evaluating sparse autoencoders . arXiv preprint arXiv:2406.04093

    arXiv doi:10.48550/arxiv.2406.04093 2024

  8. [8]

    Tyler Giallanza, Declan Campbell, Jonathan D Cohen, and Timothy T Rogers. 2024. An integrated model of semantics and control. Psychological Review

    work page 2024

Show all 41 references
  1. [9]

    John C Gower. 1975. Generalized procrustes analysis. Psychometrika, 40:33--51

    1975

  2. [10]

    Martin N Hebart, Charles Y Zheng, Francisco Pereira, and Chris Ian Baker. 2020. Revealing the multidimensional mental representations of natural objects underlying human similarity judgments. Nature Human Behaviour, 4(11):1173--1185

    2020

  3. [11]

    Roee Hendel, Mor Geva, and Amir Globerson. 2023. https://doi.org/10.48550/arXiv.2310.15916 In-context learning creates task vectors . arXiv preprint arXiv:2310.15916. Accepted at Findings of EMNLP 2023

    2023 doi arXiv

  4. [12]

    Michael C Hout, Arryn Robbins, Hayward J Godwin, Gemma Fitzsimmons, and Collin Scarince. 2022. Visual and semantic similarity norms for a photographic image stimulus set containing recognizable objects, animals and scenes. Journal of Open Psychology Data, 10(1)

    2022

  5. [13]

    Gabriel Ilharco, Samuel Kerr, Douwe Kiela, Mitchell Wortsman, Tim Dettmers, Maarten Sap, Jialin Schominski, Xingliang Chen, Wenhao Zhao, Ludwig Schmidt, et al. 2023. Taskventures: Venturing into the land of large language model task vectors. arXiv preprint arXiv:2310.15916

    2023 arXiv

  6. [14]

    Kevin G Jamieson, Lalit Jain, Chris Fernandez, Nicholas J Glattard, and Rob Nowak. 2015. Next: A system for real-world development, evaluation, and application of active learning. Advances in neural information processing systems, 28

    2015

  7. [15]

    Kenneth Li, Oam Patel, Fernanda Vi\'egas, Hanspeter Pfister, and Martin Wattenberg. 2023. https://doi.org/10.48550/arXiv.2306.03341 Inference‑time intervention: Eliciting truthful answers from a language model . arXiv preprint arXiv:2306.03341

    2023 doi arXiv

  8. [16]

    Drew Linsley, Ivan F Rodriguez, Thomas Fel, Michael Arcaro, Saloni Sharma, Margaret Livingstone, and Thomas Serre. 2023. Performance-optimized deep neural networks are evolving into worse models of inferotemporal visual cortex. arXiv preprint arXiv:2306.03779

    2023 arXiv

  9. [17]

    Laria Reynolds Rishi Liu. 2021. Prompt programming for large language models: Beyond the few-shot paradigm. In Extended Abstracts of the 2021 CHI Conference on Human Factors in Computing Systems, pages 1--7

    2021

  10. [18]

    Wenhao Liu, Xiaohua Wang, Zihan Ye, Jingwei Zhang, Hanchao Tang, Zhi Yang, Chuanyang Wang, Zhicheng Xu, Yiqi Zhou, Xiaocheng Wu, et al. 2023. Aligning large language models with human preferences through representation engineering. arXiv preprint arXiv:2312.15997

    2023 arXiv

  11. [19]

    Samuel Marks and Max Tegmark. 2024. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. arXiv preprint arXiv:2310.06824

    2024 arXiv

  12. [20]

    Earl K Miller and Jonathan D Cohen. 2001. An integrative theory of prefrontal cortex function. Annual review of neuroscience, 24(1):167--202

    2001

  13. [21]

    Daniel Mirman, Jon-Frederick Landrigan, and Allison E Britt. 2017. Taxonomic and thematic semantic systems. Psychological Bulletin, 143(5):499--520

    2017

  14. [22]

    Kushin Mukherjee and Timothy T Rogers. 2025. Using drawings and deep neural networks to characterize the building blocks of human visual similarity. Memory & Cognition, 53(1):219--241

    2025

  15. [23]

    Lukas Muttenthaler, Jonas Dippel, Lorenz Linhardt, Robert A Vandermeulen, and Simon Kornblith. 2023 a . Human alignment of neural network representations. In International Conference on Learning Representations

    2023

  16. [24]

    Lukas Muttenthaler, Lorenz Linhardt, Jonas Dippel, Robert A Vandermeulen, Katherine Hermann, Andrew Lampinen, and Simon Kornblith. 2023 b . Improving neural network representations using human similarity judgments. In Advances in Neural Information Processing Systems, volume 36

    2023

  17. [25]

    Lukas Muttenthaler, Charles Y Zheng, Patrick McClure, Robert A Vandermeulen, Martin N Hebart, and Francisco Pereira. 2022. Vice: Variational interpretable concept embeddings. Advances in Neural Information Processing Systems, 35:33661--33675

    2022

  18. [26]

    Bernat Ortiz and Joan Lasenby. 2023. Task vectors: Compositional task arithmetic for zero-shot task adaptation. arXiv preprint arXiv:2310.15213

    2023 arXiv

  19. [27]

    Nina Panickssery, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Matt Turner. 2024. https://aclanthology.org/2024.acl-long.828.pdf Steering llama 2 via contrastive activation addition . In Proceedings of the 62nd Annual Meeting of the Association for Comp...

    2024

  20. [28]

    Steven T Piantadosi and Felix Hill. 2021. Performance vs. competence in human–machine comparisons. Proceedings of the National Academy of Sciences, 118(43):e1905334118

    2021

  21. [29]

    Matthew A Lambon Ralph, Elizabeth Jefferies, Karalyn Patterson, and Timothy T Rogers. 2017. The neural and computational bases of semantic cognition. Nature reviews neuroscience, 18(1):42--55

    2017

  22. [30]

    Timothy T Rogers. 2024. Generalization and abstraction: Human memory as a magic library. The Oxford Handbook of Human Memory, Two Volume Pack: Foundations and Applications, page 172

    2024

  23. [31]

    Timothy T Rogers and James L McClelland. 2004. Semantic cognition: A parallel distributed processing approach. MIT press

    2004

  24. [32]

    Andrew M Saxe, James L McClelland, and Surya Ganguli. 2019. A mathematical theory of semantic development in deep neural networks. Proceedings of the National Academy of Sciences, 116(23):11537--11546

    2019

  25. [33]

    Scott Sievert, Robert Nowak, and Timothy T Rogers. 2023. Efficiently learning relative similarity embeddings with crowdsourcing. Journal of open source software, 8(84)

    2023

  26. [34]

    Nishant Subramani, Nivedita Suresh, and Matthew E. Peters. 2022. https://doi.org/10.18653/v1/2022.findings-acl.48 Extracting latent steering vectors from pretrained language models . In Findings of the Association for Computational Linguistics (ACL), pages 566--581

    2022 doi

  27. [35]

    Ilia Sucholutsky, Lukas Muttenthaler, Adrian Weller, Andi Peng, Andreea Bobu, Been Kim, Bradley C Love, Erin Grant, Iris Groen, Jascha Achterberg, et al. 2023. Getting aligned on representational alignment. arXiv preprint arXiv:2310.13018

    2023 arXiv

  28. [36]

    Siddharth Suresh, Wei-Chun Huang, Kushin Mukherjee, and Timothy T. Rogers. 2024. https://openreview.net/forum?id=iE5aXw3RFd Categories vs semantic features: What shape the similarities people discern in photographs of objects? In ICLR 2024 Workshop on Representational Alignment

    2024

  29. [37]

    Siddharth Suresh, Kushin Mukherjee, Xizheng Yu, Wei-Chun Huang, Lisa Padua, and Timothy Rogers. 2023. Conceptual structure coheres in human cognition but not in large language models. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pa...

    2023

  30. [38]

    Omer Tamuz, Ce Liu, Serge Belongie, Ohad Shamir, and Adam Tauman Kalai. 2011. Adaptively learning the crowd kernel. arXiv preprint arXiv:1105.1033

    2011 arXiv

  31. [39]

    Adly Templeton, Tom Conerly, Jonathan Marcus, Jack Lindsey, Trenton Bricken, Brian Chen, et al. 2024. Scaling monosemanticity: Extracting interpretable features from claude 3 sonnet. Transformer Circuits Thread

    2024

  32. [40]

    V\'azquez, Ulisse Mini, and Monte MacDiarmid

    Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. V\'azquez, Ulisse Mini, and Monte MacDiarmid. 2023. https://doi.org/10.48550/arXiv.2308.10248 Steering language models with activation engineering . arXiv preprint arXiv:2308.10248. Introduces Activation ...

    2023 doi arXiv

  33. [42]

    Andy Zou, Zifan Wang, Roger Grosse, Jason Wei, Jacob Adler, Minsuk Chen, Gregory DeSalvo, Steven Geiger, Noah Oppenheim, Aniruddh Venkatesh, et al. 2024. Representation engineering: A top-down approach to ai alignment. arXiv preprint arXiv:2501.17148

    2024 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.