Pith. sign in

REVIEW 5 major objections 7 minor 3 cited by

HyperSteer: Activation Steering at Scale with Hypernetworks

T0 review · 5 major / 7 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 A single trained hypernetwork can generate steering vectors for language models on concepts it has never seen, outperforming per-concept supervised steering baselines and roughly matching prompt-based steering.

desk verdict HyperSteer is a real step forward for activation steering with a novel hypernetwork architecture and strong zero-shot generalization, but the baseline comparison and judge bias need to be addressed before the headline claims are trustworthy. read the letter →

arxiv 2506.03292 v1 pith:JQQDZSXI submitted 2025-06-03 cs.CL cs.AIcs.LG

Jiuding Sun , Sidharth Baskaran , Zhengxuan Wu , Michael Sklar , Christopher Potts , Atticus Geiger This is my paper
classification cs.CLcs.AIcs.LG
keywords activationsteeringhypernetworksvectorslanguagemodelcontrolcross-attentionzero-shotgeneralizationamortizedinferenceAxBench
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

Activation steering—modifying a language model's internal activations to control its output—usually requires either per-concept supervised training or a scalable but unreliable unsupervised search. This paper claims a third option: train one hypernetwork on thousands of natural-language steering prompts, and it will generate an effective steering vector for any new prompt, including concepts never seen during training. On Gemma-2-2B and 9B, the best HyperSteer variant beats the per-concept ReFT-r1 baseline and approaches, sometimes matches, steering-via-prompting. If true, steering becomes an amortized, language-driven control mechanism: the cost shifts from per-task data collection and training to one upfront training run that pays off as the prompt set grows.

What carries the argument

The central object is a transformer hypernetwork: a neural network that takes a steering prompt (and optionally the base prompt and the base LM's residual-stream activations) and outputs a steering vector $\Delta^s_x \in \mathbb{R}^d$, which is added to the base model's hidden state at a chosen layer. The best variant, Cross Attention, inserts masked cross-attention modules at each hypernetwork layer; queries come from the hypernetwork's residual stream and keys and values come from the base LM's residual stream at the intervention layer, so the generated vector is conditioned on how the base model is processing the actual input. The whole system is trained end-to-end with a language-modeling loss on the steered outputs against teacher-generated labels, freezing the base LM.

What would settle it

Have human annotators or an independently trained judge model rate a random sample of HyperSteer-steered, ReFT-r1-steered, and prompt-steered outputs for adherence to the steering prompt; if HyperSteer's advantage over ReFT-r1 and its parity with prompting shrink or reverse under this independent rating, the central claim is an artifact of judge bias.

Watch

Extended reading notes

Core claim

The paper's central claim is that the mapping from a natural-language steering prompt to a residual-stream steering vector can itself be learned by a hypernetwork, and that this amortized mapping outperforms the previous state-of-the-art activation steering method. Concretely, the cross-attention HyperSteer variant, trained on roughly 16,000 AxBench steering prompts, scores 0.742 held-in and 0.608 held-out on Gemma-2-2B, versus 0.509 for the ReFT-r1 baseline that is trained separately on each steering prompt; on Gemma-2-9B the held-in score reaches parity with prompting at 1.091. The authors interpret these results as combining the scalability of dictionary learning with the targeted control of supervised steering, and they report that performance grows roughly logarithmically with the number of training prompts.

Load-bearing premise

The evaluation assumes that a gpt-4o-mini judge gives an unbiased measure of steering success, but the same model generated the training labels, so the reported scores could partly reward outputs that mimic the judge's preferred style rather than actually following the steering prompt.

Editorial extensions

If this is right

  • Steering a language model to follow a new concept requires no per-concept data collection or training: a user states the concept in natural language and the hypernetwork produces the vector.
  • Compute amortizes: the TFLOPs needed per steering prompt falls as the training set grows, so adding more steering concepts makes the method cheaper relative to per-concept baselines.
  • Activation steering can approach the performance of prompt engineering while adding no extra tokens to the input and changing no model parameters.
  • The gap between activation steering and fine-tuning narrows, with HyperSteer outperforming LoRA and SFT on the AxBench held-in evaluation for Gemma-2-2B.

Reading between the lines

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

  • If the hypernetwork truly generalizes across prompt vocabulary, steering vectors for related concepts likely form a smooth, language-conditioned manifold; probing that geometry could connect activation steering to interpretability work on feature directions.
  • A straightforward stress test would replace the gpt-4o-mini judge with human raters or a judge from a different model family; if the advantage over ReFT-r1 and parity with prompting shrink, some of the reported gain is style-matching rather than concept following.
  • The same hypernetwork design could be trained to emit LoRA or ReFT adapters instead of single vectors, extending amortized control to parameter-efficient fine-tuning, a direction the authors list as future work.
  • Because the hypernetwork is itself roughly model-sized, its per-query inference cost is nontrivial; quantifying the break-even scale where amortized generation beats per-concept training would clarify practical deployment.
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

5 major / 7 minor

Summary. The paper introduces HyperSteer, a family of hypernetwork models that map a natural-language steering prompt (and optionally the base prompt and base-model activations) to a residual-stream steering vector for an instruction-tuned LM. The hypernetwork is trained end-to-end with a language-modeling loss on AxBench data. Experiments on Gemma-2-2B and Gemma-2-9B report that a cross-attention variant outperforms ReFT-r1 on seen steering prompts and, zero-shot, on unseen steering prompts, approaches prompt-based steering, and becomes more compute-efficient with scale. The paper also includes ablations, geometric analyses, and qualitative examples.

Significance. The central idea is original and timely: a single hypernetwork that amortizes steering-vector construction over thousands of prompts, with zero-shot generalization to new steering prompts, would be a substantial step beyond per-concept dictionary learning. The cross-attention architecture and the scaling analysis are sensible, and the paper grounds the work in the AxBench benchmark with explicit comparisons to multiple baselines. However, the strength of the empirical claims is currently constrained by evaluation-design issues: the judge and label-generation model overlap, the held-out comparison against ReFT-r1 is not directly shown, and no uncertainty estimates are reported. If these are resolved, the method would be a clear contribution to activation steering.

major comments (5)
  1. [Section 2, Table 1] The headline held-out claim is not directly evidenced. Table 1 reports ReFT-r1 (and all other fine-tuning and activation-steering baselines) only in the held-in columns; the held-out columns contain only Prompting and HyperSteer. The text in Section 4 states that HyperSteer 'performs better on unseen steering prompts than every supervised activation steering baseline trained and evaluated on the same steering prompt,' but no such ReFT-r1 held-out number appears in any table. Please add held-out evaluations for ReFT-r1 (and ideally the other baselines), trained per concept on the held-out steering prompts, with the same held-out base prompts and judge, so the comparison is explicit. Figure 2's caption mentions a ReFT-r1 line; if it is plotted, disclose exactly what it represents and provide the underlying values.
  2. [Section 2 / Appendix A.7.3] The evaluation is potentially confounded by the fact that the judge (gpt-4o-mini) is the same model family used to generate the training labels (Appendices A.5 and A.7.3). Because HyperSteer is optimized with a cross-entropy loss against gpt-4o-mini labels, its outputs may be stylistically closer to that model and thus receive higher judge scores for reasons unrelated to steering adherence. The manuscript itself acknowledges this possibility in A.7.3. Please provide an evaluation with a judge from a different model family (or human ratings) on at least the main comparisons in Table 1, and report whether the ranking of HyperSteer versus ReFT-r1 and prompting changes.
  3. [Section 4, Table 1] No error bars or significance tests are reported. All numbers are point estimates, and differences such as 0.608 vs. 0.762 (held-out prompting) or 0.742 vs. 0.509 (held-in ReFT-r1) should be accompanied by variance estimates over seeds or bootstrap samples. Without them, the superiority of the cross-attention variant over ReFT-r1 at scale, and the parity claim with prompting, are not fully established.
  4. [Appendix A.3.2, Figure 3] The compute-efficiency claim rests on an extrapolation. The TFLOPs curve is fit with f(c) = a + b·exp(dc) with three free parameters, and the conclusion lim_{c→∞} f(c) = a < FReFT depends on the fitted asymptote. Please report the fit parameters with confidence intervals or bootstrap ranges, include actual (not extrapolated) compute at the maximum training scale, and state how the ReFT-r1 TFLOPs value was measured, including whether hyperparameter search is accounted for.
  5. [Section 7, Limitations] There is an internal inconsistency in the limitations statement: the text says 'Due to to compute constraints we only experimented with Gemma-2-2B architectures,' but Table 1 and Section 4 report results on both Gemma-2-2B and Gemma-2-9B. Please correct this, and clarify which experiments use which model.
minor comments (7)
  1. [Section 7] The text contains 'Due to to compute constraints' with a duplicated 'to'; also the sentence beginning 'Compare' is a sentence fragment.
  2. [Section 4, Ablation Study] The sentence 'the number of decoder blocks has a greater impact on generalization to steering prompts unseen in training (+0.07) compared to steering prompts unseen in training (+0.03)' repeats 'unseen in training' twice; the second instance should presumably read 'seen in training.'
  3. [Appendix, Table 4] The mapping between short and full descriptions appears to have swapped rows: 'Mathematical Notation' is mapped to 'key phrases related to personal aspirations and career transitions' and 'Proper Nouns' to 'occurrences of mathematical symbols or notation.' Please verify the table.
  4. [Section 9, Acknowledgments] The release URL is given as '/githubstanfordnlp/axbench', which appears truncated; provide the full URL.
  5. [Figure 2] The caption mentions 'Previous best ReFT-r1 (Dictionary Learning)' but does not explain how this line was computed; state whether it is a held-in or held-out value and how it is plotted on the held-out axis.
  6. [Appendix A.6] The sentence 'The core comparisons however' is incomplete and should be finished or removed.
  7. [Throughout] The name 'HYPERSTEER' is typeset inconsistently (HYPERSTEER, HYPER STEER, HyperSteer); please unify the notation.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: HyperSteer's claims are empirical evaluations against an external benchmark, and the noted judge/label overlap is an evaluation-bias concern rather than a circular derivation.

full rationale

The paper's central claim is empirical: a hypernetwork trained end-to-end on AxBench steering prompts generates steering vectors that outperform ReFT-r1 and match prompting on held-in and held-out concepts. There is no derivation chain in which a claimed result is defined in terms of its own input. Training minimizes cross-entropy against gpt-4o-mini-generated labels (Eq. 4), while the evaluation metric is a separate gpt-4o-mini judge scoring base-prompt adherence, steering adherence, and fluency; these are not the same object, and the held-out steering prompts were not used in training. The self-citations (AxBench, HyperDAS, pyvene) supply benchmark data, baseline numbers, and implementation infrastructure rather than load-bearing premises, and AxBench is a public benchmark with external data. The compute-efficiency curve is fitted to measured TFLOPs, but it is presented as an empirical scaling observation, not as a first-principles prediction. The paper itself flags in A.7.3 that 'the LLM-as-a-judge evaluation setup (which also uses gpt-4o-mini) being biased towards outputs from the same model' is a possible explanation for its preliminary findings. This is a genuine validity threat to the evaluation, but it is not circularity: the target result is not assumed as an input, and the reported comparisons are not forced by construction.

Assumptions & free parameters 1 free parameters · 3 assumptions · 0 invented entities

The central claims rest on the AxBench dataset, the gpt-4o-mini evaluation harness, and a small set of fitted/curve parameters. No new physical entities are introduced; the only fitted numbers are the compute-efficiency exponential coefficients. Domain assumptions are inherited from prior work on activation steering and hypernetworks.

free parameters (1)
  • Compute-efficiency exponential coefficients (a, b, d) = a=87.7035, b=1521.1495, d=-0.0034
    Fitted to TFLOPs-per-steering-prompt measurements at discrete training dataset sizes (c in [10, 16000]) in Appendix A.3.2. The conclusion that per-prompt cost asymptotes below ReFT-r1 depends on this exponential form and its extrapolation to infinite data.
assumptions (3)
  • domain assumption Activation steering by adding a vector to the residual stream reliably induces target behaviors in instruction-tuned LMs.
    Borrowed from prior work (Giulianelli et al.; Wu et al., 2025); the paper does not independently validate this mechanism.
  • domain assumption The AxBench evaluation protocol with a gpt-4o-mini judge yields a valid measure of steering success.
    The judge is used both to create training labels and score outputs; the paper acknowledges possible self-preference bias in A.7.3.
  • domain assumption The 16k GemmaScope-derived steering prompts are representative enough that a hypernetwork trained on them generalizes to held-out prompts.
    Generalization is the central claim; the paper offers no analysis of prompt distribution coverage.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HyperSteer: Activation Steering at Scale with Hypernetworks." pith.science (2026). https://pith.science/paper/JQQDZSXI

@misc{pith2026250603292,
  author       = {Pith},
  title        = {Pith review of: HyperSteer: Activation Steering at Scale with Hypernetworks},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JQQDZSXI}},
  note         = {Machine review of arXiv:2506.03292}
}
read the original abstract

Steering language models (LMs) by modifying internal activations is a popular approach for controlling text generation. Unsupervised dictionary learning methods, e.g., sparse autoencoders, can be scaled to produce many steering vectors, but lack guarantees on the individual efficacy of each vector and control over the coverage of relevant steering tasks. In contrast, supervised methods for constructing steering vectors are targeted and effective, but require more data collection and training for each additional steering vector produced. In this work, we introduce HyperSteer, a family of hypernetwork-based architectures which are trained end-to-end to generate steering vectors conditioned on the natural language steering prompts and the internals of the steered LM. In our evaluations, we show that scaling HyperSteer with thousands of steering prompts exceeds the performance of state-of-the-art activation steering methods, even on steering prompts never seen during training. Moreover, HyperSteer performs on par with steering-via-prompting.

Figures

Figures reproduced from arXiv: 2506.03292 by the authors.

Figure 1
Figure 1. The state-of-the-art HYPERSTEER Model: A transformer hypernetwork uses self attention to process a steering prompt and uses a cross attention module to read from the residual stream of a base LM run on a second prompt. The hypernetwork outputs a steering vector that is added to the base LM residual stream. vised, e.g., Sparse Autoencoders (SAEs) (Hernan￾dez et al., 2022; Cunningham et al., 2023; Bricken et al., 2023… view at source ↗
Figure 2
Figure 2. Performance on steering prompts that have [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. As the number of steering prompts in our [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (16 more)
Figure 4
Figure 4. Figure 4: Concept10 t-SNE analysis of 2500 steering vectors from HYPERSTEER (cross attention), 250 per concept. 150 100 50 0 50 100 150 200 PC 1 100 50 0 50 100 PC 2 Concept Structured Data Entries Personal Identity References Time References Java Interface References Legal Term…
Figure 5
Figure 5. Figure 5: Concept10 PCA analysis (2 components) 2500 steering vectors from HYPERSTEER (cross at￾tention), 250 per concept. cross-attention heatmaps across layers and heads (N = 20 layers). A key takeaway is that all query (concept) to￾kens all tend to attend to the same or a few…
Figure 6
Figure 6. Figure 6: Pairwise cosine similarities of steering vectors, averaged within each steering prompt, for our two [PITH_FULL_IMAGE:figures/full_fig_p010_6.png]
Figure 7
Figure 7. Figure 7: Concept10 perplexity distribution on data labels sampled from gpt-4o-mini. We ran preliminary experiments by training on labels from both distributions, and find steering performance is still better with the gpt-4o-mini labels. We suspect that this could be a result of…
Figure 11
Figure 11. Figure 11: Layer 10 attention map [PITH_FULL_IMAGE:figures/full_fig_p011_11.png]
Figure 10
Figure 10. Figure 10: Layer 5 attention map [PITH_FULL_IMAGE:figures/full_fig_p011_10.png]
Figure 13
Figure 13. Figure 13: Successful steering and instruction following by H [PITH_FULL_IMAGE:figures/full_fig_p012_13.png]
Figure 14
Figure 14. Figure 14: Successful steering and instruction following by H [PITH_FULL_IMAGE:figures/full_fig_p012_14.png]
Figure 15
Figure 15. Figure 15: Successful steering and instruction following by H [PITH_FULL_IMAGE:figures/full_fig_p012_15.png]
Figure 16
Figure 16. Figure 16: Successful steering and instruction following by H [PITH_FULL_IMAGE:figures/full_fig_p013_16.png]
Figure 17
Figure 17. Figure 17: Successful steering and instruction following by H [PITH_FULL_IMAGE:figures/full_fig_p013_17.png]
Figure 18
Figure 18. Figure 18: Failed steering by HYPERSTEER, but successful instruction following [PITH_FULL_IMAGE:figures/full_fig_p013_18.png]
Figure 19
Figure 19. Figure 19: Failed steering and instruction following by H [PITH_FULL_IMAGE:figures/full_fig_p014_19.png]
Figure 20
Figure 20. Figure 20: Failed steering by HYPERSTEER. Steering Prompt: key phrases related to personal aspirations and career transitions Input Prompt How can you determine if a restaurant is popular based on online reviews? HYPERSTEER Generation Transitioning from understanding how to nurt…
Figure 21
Figure 21. Figure 21: Somewhat successful steering by HYPERSTEER, but failed to follow the instruction [PITH_FULL_IMAGE:figures/full_fig_p014_21.png]
Figure 22
Figure 22. Figure 22: Failed steering by HYPERSTEER, but successful instruction following [PITH_FULL_IMAGE:figures/full_fig_p015_22.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 3 Pith papers

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

  1. Resist and Update: Counterfactual Report Coordinates for Incentive-Compatible LLMs

    cs.AI 2026-07 conditional novelty 6.5 of 10

    A two-pass counterfactual report-coordinate clamp jointly achieves perfect resist-and-update scores on a Bayesian-witness benchmark by holding LLM reports to an incentive-neutralized causal contract.

  2. Policy Gradient Steering: Interventions from Behavioral Objectives

    cs.LG 2026-07 conditional novelty 6.0 of 10

    PGS builds a removable activation offset from return-weighted action-score gradients and steers frozen policies across gridworld, chess, and football.

  3. Persistent Sparse Autoencoders: Learning Feature Timescales in Language Models

    cs.LG 2026-07 conditional novelty 6.0 of 10

    Persistent SAEs learn per-feature persistence coefficients from reconstruction, splitting features into fast local detectors and slow topic-tracking states that retain prompt-injection signals over long contexts.

Reference graph

Works this paper leans on

26 extracted references · 6 canonical work pages · cited by 3 Pith papers

  1. [1]

    Trenton Bricken, Adly Templeton, Joshua Batson, Brian Chen, Adam Jermyn, Tom Conerly, Nick Turner, Cem Anil, Carson Denison, Amanda Askell, Robert Lasenby, Yifan Wu, Shauna Kravec, Nicholas Schiefer, Tim Maxwell, Nicholas Joseph, Zac Hatfield-Dodds, Alex Tamkin, Karina Nguyen, and 6 others. 2023. Towards monosemanticity: Decomposing language models with d...

    2023

  2. [2]

    Karl Cobbe, Vineet Kosaraju, Mohammad Bavarian, Mark Chen, Heewoo Jun, Lukasz Kaiser, Matthias Plappert, Jerry Tworek, Jacob Hilton, Reiichiro Nakano, Christopher Hesse, and John Schulman. 2021. Training verifiers to solve math word problems. arXiv preprint arXiv:2110.14168

    arXiv 2021

  3. [3]

    Mike Conover, Matt Hayes, Ankit Mathur, Jianwei Xie, Jun Wan, Sam Shah, Ali Ghodsi, Patrick Wendell, Matei Zaharia, and Reynold Xin. 2023. https://www.databricks.com/blog/2023/04/12/dolly-first-open-commercially-viable-instruction-tuned-llm Free dolly: Introducing the world's first truly open instruction-tuned llm

    2023

  4. [4]

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

    arXiv 2023

  5. [5]

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

    arXiv 2024

  6. [6]

    Mario Giulianelli, Jack Harding, Florian Mohnert, Dieuwke Hupkes, and Willem H. Zuidema. 2018. https://doi.org/10.18653/V1/W18-5426 Under the hood: Using diagnostic classifiers to investigate and improve how language models track agreement information . In Proceedings of the Workshop: Analyzing and Interpreting Neural Networks for NLP, BlackboxNLP EMNLP 2...

    doi:10.18653/v1/w18-5426 2018

  7. [7]

    David Ha, Andrew Dai, and Quoc V. Le. 2016. https://arxiv.org/abs/1609.09106 Hypernetworks . Preprint, arXiv:1609.09106

    arXiv 2016

  8. [8]

    Evan Hernandez, Sarah Schwettmann, David Bau, Teona Bagashvili, Antonio Torralba, and Jacob Andreas. 2022. https://openreview.net/forum?id=NudBMY-tzDr Natural language descriptions of deep visual features . In ICLR

    work page 2022

Show all 26 references
  1. [9]

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, and 1 others. 2022. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3

    2022

  2. [10]

    Kenneth Li, Oam Patel, Fernanda Vi \'e gas, Hanspeter Pfister, and Martin Wattenberg. 2023 a . https://openreview.net/forum?id=aLLuYpn83y Inference-time intervention: Eliciting truthful answers from a language model . In Thirty-seventh Conference on Neural Information Processi...

    2023

  3. [11]

    Hashimoto

    Xuechen Li, Tianyi Zhang, Yann Dubois, Rohan Taori, Ishaan Gulrajani, Carlos Guestrin, Percy Liang, and Tatsunori B. Hashimoto. 2023 b . Alpacaeval: An automatic evaluator of instruction-following models. https://github.com/tatsu-lab/alpaca_eval

    2023

  4. [12]

    Dragan, Rohin Shah, and Neel Nanda

    Tom Lieberum, Senthooran Rajamanoharan, Arthur Conmy, Lewis Smith, Nicolas Sonnerat, Vikrant Varma, J \' a nos Kram \' a r, Anca D. Dragan, Rohin Shah, and Neel Nanda. 2024. https://doi.org/10.48550/ARXIV.2408.05147 Gemma scope: Open sparse autoencoders everywhere all at once ...

    2024 doi arXiv

  5. [13]

    Samuel Marks, Can Rager, Eric J Michaud, Yonatan Belinkov, David Bau, and Aaron Mueller. 2025. https://openreview.net/forum?id=I4e82CIDxv Sparse feature circuits: Discovering and editing interpretable causal graphs in language models . In The Thirteenth International Conferenc...

    2025

  6. [14]

    Samuel Marks and Max Tegmark. 2023. https://doi.org/10.48550/ARXIV.2310.06824 The geometry of truth: Emergent linear structure in large language model representations of true/false datasets . CoRR, abs/2310.06824

    2023 doi arXiv

  7. [15]

    OpenAI, :, Aaron Hurst, Adam Lerer, Adam P. Goucher, Adam Perelman, Aditya Ramesh, Aidan Clark, AJ Ostrow, Akila Welihinda, Alan Hayes, Alec Radford, Aleksander Mądry, Alex Baker-Whitcomb, Alex Beutel, Alex Borzunov, Alex Carney, Alex Chow, Alex Kirillov, and 401 others. 2024....

    2024 arXiv

  8. [16]

    Jason Phang, Yi Mao, Pengcheng He, and Weizhu Chen. 2023. https://proceedings.mlr.press/v202/phang23a.html H yper T uning: Toward adapting large language models without back-propagation . In Proceedings of the 40th International Conference on Machine Learning, volume 202 of Pr...

    2023

  9. [17]

    Nina Rimsky, Nick Gabrieli, Julian Schulz, Meg Tong, Evan Hubinger, and Alexander Turner. 2024. https://doi.org/10.18653/v1/2024.acl-long.828 Steering llama 2 via contrastive activation addition . In Proceedings of the 62nd Annual Meeting of the Association for Computational L...

    2024 doi

  10. [18]

    Morgane Rivi \` e re, Shreya Pathak, Pier Giuseppe Sessa, Cassidy Hardin, Surya Bhupatiraju, L \' e onard Hussenot, Thomas Mesnard, Bobak Shahriari, Alexandre Ram \' e , Johan Ferret, Peter Liu, Pouya Tafti, Abe Friesen, Michelle Casbon, Sabela Ramos, Ravin Kumar, Charline Le ...

    2024 doi arXiv

  11. [19]

    Jiuding Sun, Jing Huang, Sidharth Baskaran, Karel D'Oosterlinck, Christopher Potts, Michael Sklar, and Atticus Geiger. 2025. https://openreview.net/forum?id=6fDjUoEQvm Hyper DAS : Towards automating mechanistic interpretability with hypernetworks . In The Thirteenth Internatio...

    2025

  12. [20]

    Alexander Matt Turner, Lisa Thiergart, David Udell, Gavin Leech, Ulisse Mini, and Monte MacDiarmid. 2023. https://doi.org/10.48550/ARXIV.2308.10248 Activation addition: Steering language models without optimization . CoRR, abs/2308.10248

    2023 doi arXiv

  13. [21]

    Laurens van der Maaten and Geoffrey Hinton. 2008. http://jmlr.org/papers/v9/vandermaaten08a.html Visualizing data using t-sne . Journal of Machine Learning Research, 9(86):2579--2605

    2008

  14. [22]

    Manning, and Christopher Potts

    Zhengxuan Wu, Aryaman Arora, Atticus Geiger, Zheng Wang, Jing Huang, Dan Jurafsky, Christopher D. Manning, and Christopher Potts. 2025. https://arxiv.org/abs/2501.17148 AxBench : Steering llms? even simple baselines outperform sparse autoencoders . Preprint, arXiv:2501.17148

    2025 arXiv

  15. [23]

    Zhengxuan Wu, Aryaman Arora, Zheng Wang, Atticus Geiger, Dan Jurafsky, Christopher D Manning, and Christopher Potts. 2024 a . https://openreview.net/forum?id=fykjplMc0V Re FT : Representation finetuning for language models . In The Thirty-eighth Annual Conference on Neural Inf...

    2024

  16. [24]

    Zhengxuan Wu, Atticus Geiger, Aryaman Arora, Jing Huang, Zheng Wang, Noah Goodman, Christopher Manning, and Christopher Potts. 2024 b . https://aclanthology.org/2024.naacl-demo.16 pyvene: A library for understanding and improving P y T orch models via interventions . In Procee...

    2024

  17. [25]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list...

  18. [26]

    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 gl...

Pith tools

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