Pith. sign in

REVIEW 4 major objections 4 minor 1 cited by

When Models Refuse: Political Steerability and Feature Richness as Measures of Ideological Depth

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read This paper claims that when an LLM refuses a benign political instruction, the refusal can be a capability deficit—a shortage of internal political features—rather than a fixed safety rule, and that 'ideological depth' measured by steerabil

desk verdict A new empirical observation about refusal and SAE features, but the causal claim in the abstract is not supported by any reported ablation and the numbers don't add up. read the letter →

arxiv 2508.21448 v3 pith:C2ATC6ZS submitted 2025-08-29 cs.CL

Shariar Kabir This is my paper
classification cs.CL
keywords ideologicaldepthpoliticalsteerabilitysparseautoencodersrefusalbehaviorcapabilitydeficitfeaturerichnessactivationsteeringLLMalignment
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 examines why two similar-sized open-weight LLMs diverge when asked to argue conservative positions: one shifts its answers, the other frequently refuses. The authors define 'ideological depth' by two measurable components—steerability without failure and the richness of internal political features recovered from sparse autoencoders. Comparing the two models, the more steerable one activates roughly 7.3 times more distinct political features, and ablating a small targeted set of those features reproduces the feature-poor profile and drives up refusals. The paper concludes that refusals on benign prompts can signal a lack of representational capability rather than an active safety stance, and that ideological depth is a measurable property that predicts when a model will refuse.

What carries the argument

The central object is 'ideological depth,' operationalized by two measurable quantities: steerability without failure and feature richness. Steerability is measured through prompt-engineering conditions and activation steering—contrastive activation addition and steering target atoms—while feature richness is measured by decomposing hidden-layer activations with sparse autoencoders into sparsely activating, interpretable features. The causal claim is carried by targeted feature ablation: selecting a small set of political features by activation amplitude and frequency differences, deleting them, and observing refusal behavior rise.

What would settle it

Train or procure a sparse autoencoder for the instruction-tuned version of the less steerable model using the same dictionary-learning setup as for the more steerable model, then recount activated political features on the same 126 statements; if the 7.3x gap shrinks or disappears, the capability-deficit conclusion is an artifact of mismatched measurements.

Watch

Extended reading notes

Core claim

The paper's central claim is that ideological depth is a measurable property with two components: a model's ability to follow political instructions without breaking down (steerability), and the richness of its internal political feature vocabulary, quantified by how many distinct sparse-autoencoder features activate on political statements. Comparing two similar-sized open-weight models, the authors find large differences on both axes: the more steerable model activates about 7.3 times more distinct political features, while the other produces more refusals, especially when asked to argue conservative positions. The causal step is an ablation: removing a small, targeted set of political fea

Load-bearing premise

The paper's central comparison assumes the two models' sparse autoencoder dictionaries are equally faithful windows into political representations, and that an autoencoder trained on the base (non-instruction-tuned) version of one model faithfully measures the instruction-tuned model; if either fails, the feature-richness gap could come from the measuring instrument rather than the model's capability.

Editorial extensions

If this is right

  • Refusal rate stops being a trustworthy safety metric: a model that refuses more may simply be less capable of representing the requested viewpoint.
  • Ideological depth can be measured before deployment, giving a predictor of when a model will refuse on political topics.
  • Safety monitoring can move toward internal features: richer, more coherent feature sets make a model more inspectable and steerable.
  • Ablating political features is a reversible causal intervention, so feature richness can be experimentally manipulated to test refusal mechanisms.

Reading between the lines

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

  • The 7.3x feature-count gap may partly reflect differences between the two SAE dictionaries and the use of a base-model autoencoder on an instruction-tuned model; a matched SAE trained on both models would settle this.
  • If refusals can be capability deficits, the same logic may apply beyond politics—refusals on math, coding, or low-resource-language instructions might also mask missing representations rather than policy.
  • A testable extension: expose the shallow model to additional political text through continued pretraining or fine-tuning and check whether refusal rates fall; if they do not, the capability-deficit story needs revision.
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

4 major / 4 minor

Summary. The paper proposes 'ideological depth' as a composite of steering success (steerability) and SAE-based political feature richness, and compares Llama-3.1-8B-Instruct against Gemma-2-9B-it. Using prompt engineering, contrastive activation addition (CAA), and SAE target-atom steering, the authors report that Gemma is more easily steered in both ideological directions, while Llama increasingly refuses conservative instructions. They further report that Gemma activates far more political SAE features (claimed as ~7.3x) and that its features are more coherent and more predictive of statement category. The abstract and discussion conclude that refusals on benign prompts can be a capability deficit rather than a safety guardrail, claiming that ablating Gemma's political features reproduces Llama's feature-poor behavior and drives up refusals. However, the manuscript does not report the promised ablation experiment; the empirical sections contain steering and feature-quality analyses only.

Significance. If the causal claim were established, this would be a notable contribution connecting mechanistic interpretability, political steerability, and refusal behavior. The study uses publicly available SAEs (GemmaScope and LlamaScope), includes a layer sweep, and validates an IRT-based ideal-point model in the appendix, which are strengths. The qualitative feature evaluation via an automated judge is also a useful attempt to move beyond raw feature counts. However, the central inference—that refusal behavior is caused by a shortage of internal political features—is currently unsupported, and several measurement choices (different SAE suites, base-vs-instruct SAE mismatch, circular feature selection) confound the comparison. The paper would require substantial additional analysis before the headline conclusion can be accepted.

major comments (4)
  1. [Abstract and Section 3.3] The abstract states that 'causally ablating a small, targeted set of political features from the former model reproduces the same feature-poor behavior and drives up refusals' (also similarly in the full-text abstract and Section 4). No ablation experiment appears in the manuscript. Section 3.3 reports feature counts, output-score interventions (Eq. 1), and qualitative evaluations, but no figure, table, or quantitative result measures refusals after ablating the selected features. Figure 9 reports steering interventions, not ablations. The causal claim is therefore unsupported as written. Either an actual ablation experiment must be added (e.g., removing or zeroing the targeted SAE features and measuring refusal rates and response distributions) or the conclusions must be restricted to correlation.
  2. [Section 3.3, '7.3x' claim] The text says: 'In total, for the gemma-2-9b-it model, we found 18458 features, whereas for the llama-3.1-8b-it, we found only 4412 activated features... On average, Gemma contained 7.3 times more features that activate for a statement.' The ratio 18458/4412 is approximately 4.18, not 7.3. This same 7.3x figure is repeated in the abstract, Section 3.3, and Section 4. This is an internal inconsistency in a headline quantitative result; it must be corrected or redefined with a clear normalization.
  3. [Section 3.2] The authors use LlamaScope SAEs, noting that 'LlamaScope only offers SAEs for the base llama-3.1-8b model, we still found it effective for steering the instruction-tuned variant.' The feature richness comparison therefore uses an SAE trained on the base model to measure features of the instruct-tuned model. This is a serious confound: the low feature count and poor feature coherence for Llama could be artifacts of dictionary mismatch rather than evidence of a capability deficit. The paper should validate that the base-model SAE is appropriate for the instruct model (e.g., reconstruction loss on instruct activations, or comparison with a matched SAE) or substantially temper the cross-model claims.
  4. [Section 3.3, feature selection and evaluation] Political features are selected using Δa and Δf computed from the same 874 training prompts that are also used to build and evaluate steering, and their 'quality' is then assessed partly by how well they steer the model. This circularity can inflate the apparent association between feature richness and steerability. A held-out set or cross-validation should be used when selecting features and evaluating their predictive/steering quality, otherwise the reported relationship may reflect fitting to the evaluation set.
minor comments (4)
  1. [Appendix D title] The appendix is titled 'Promprs Used In Feature Qualitative Analysis'; 'Promprs' should be 'Prompts'.
  2. [Section 3.3] Equation (1) uses 'interveined' instead of 'intervened', and the notation r(l*, M) should be defined more explicitly (e.g., rank among all tokens at the target layer).
  3. [Section 4] The phrase 'measurable, biological property of Large Language Models' is infelicitous; LLMs are not biological. Suggest 'structural' or 'mechanistic'.
  4. [Figure 13 caption] The caption mentions 'llama-29b-it'; this appears to be a typo for 'llama-3.1-8b-it'.

Circularity Check

1 steps flagged · score 2.0 of 10

No significant circularity; minor definitional overlap between 'ideological depth' and refusal behavior, but the main causal claim is unsupported by an omitted ablation rather than circular.

  1. self definitional [Section 1 (Introduction), page 2]
    "Our work demonstrates that a model’s ideological depth is not an abstract quality but a measurable property determined by two key factors: Feature Richness, representing the size of its internal vocabulary of political concepts discovered via SAEs, and Steerability without Failure, the robustness with which it can follow ideological instructions without breaking down into refusal."

    Because 'Steerability without Failure' is defined as the ability to follow ideological instructions without refusal, the paper's later conclusion that models with low ideological depth refuse more is partly true by definition. However, the construct also includes Feature Richness, measured independently via SAE feature counts, so the causal claim about capability deficits does not reduce entirely to the definition. The overlap is a moderate conflation, not a full circular derivation.

full rationale

The paper defines 'ideological depth' as a combination of feature richness and refusal-free steerability, so the association between low depth and high refusal is partly tautological. This is a genuine but minor self-definitional issue: the feature-richness component is measured separately from refusal behavior via SAE activation counts, and the qualitative/quantitative feature evaluations provide independent evidence of a difference between the two models. No load-bearing step reduces to its inputs by construction: the SAE feature counts are not the same as the steering-evaluation outputs, and the output-score causality checks are distinct from the selection criteria. The paper's central causal claim, however, is seriously under-supported: the abstract and introduction promise that 'causally ablating a small, targeted set of political features' reproduces refusal behavior, but no ablation experiment is reported in the methods or results. This is an omitted proof and a validity concern, not a circularity, because the missing ablation is not derived from the paper's definitions or fitted parameters. The self-citation [13] is used only for an argumentative-pressure prompting technique and is not load-bearing. Overall, the derivation chain is not circular, but the headline causal conclusion overreaches the presented evidence.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The paper's central claim depends on comparing two different SAE dictionaries, applying a base-model SAE to an instruct-tuned model, and using arbitrary thresholds to identify 'political features.' It also assumes that any observed differences in refusal and steerability are caused by feature richness rather than by the many other differences between the two models.

free parameters (3)
  • Steering layer per model = Llama: layer 14; Gemma: layer 20
    Selected as the 'most responsive' layer from a CAA layer sweep (Figure 8). This choice affects all subsequent feature counts, steering results, and output scores.
  • SAE dictionary selection = 131K features for both models
    The authors chose the 131K-feature SAEs from GemmaScope and LlamaScope for comparability, but the two dictionaries are trained on different models and are not directly comparable.
  • Political feature inclusion thresholds = Δa > 0 and Δf > 0
    Features are classified as political if mean activation amplitude difference or frequency difference is positive between liberal and conservative prompts. No significance threshold or multiple-testing correction is applied, so the count of 'political features' is sensitive to this arbitrary criterion.
assumptions (4)
  • domain assumption Comparability of SAE feature dictionaries across models
    The paper treats raw counts of activated features in GemmaScope vs LlamaScope as comparable measures of political knowledge, ignoring that the two dictionaries are learned on different models with different distributions.
  • domain assumption Applicability of base-model SAEs to instruct-tuned variant
    The authors state that LlamaScope lacks an SAE for Llama-3.1-8b-Instruct and use the base-model SAE instead, assuming it faithfully represents the instruct model's features. This is acknowledged in Section 3.2 and could bias the feature count.
  • domain assumption Steerability differences reflect capability rather than safety training or model family
    Llama and Gemma differ in pretraining, instruction tuning, safety alignment, and architecture; the causal claim that refusal is a capability deficit assumes these other differences are not responsible for the observed refusal rates.
  • ad hoc to paper SAE features with positive Δa or Δf are political features
    The paper defines political features as those with positive activation-amplitude or frequency differences between liberal and conservative prompts, with no external validation that these are meaningfully political rather than general syntactic or distributional features.

how reviews work

0 comments
Cite this review

Pith. "Pith review of When Models Refuse: Political Steerability and Feature Richness as Measures of Ideological Depth." pith.science (2026). https://pith.science/paper/C2ATC6ZS

@misc{pith2026250821448,
  author       = {Pith},
  title        = {Pith review of: When Models Refuse: Political Steerability and Feature Richness as Measures of Ideological Depth},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/C2ATC6ZS}},
  note         = {Machine review of arXiv:2508.21448}
}
read the original abstract

Large language models (LLMs) sometimes refuse to follow benign instructions, such as declining to argue a political position or adopt a stated persona, and such refusals are commonly read as safety guardrails at work. We ask whether they can instead signal a **capability deficit**: a shortage of the internal representations a model needs to reason from the instructed perspective. To investigate, we introduce **ideological depth**, a property with two components: (i) a model's ability to follow political instructions without *failure* (steerability), and (ii) the **feature richness** of its internal political representations, measured with sparse autoencoders (SAEs). Using two widely used openweight LLMs as candidates, we compare interventions based on prompts and activation-steering, and probe political features with publicly available SAEs. We find large, systematic differences: a model that is more steerable in both ideological directions activates **~7.3x** more distinct political features, while the other model instead responds with increased refusals. Causally ablating a small, targeted set of political features from the former model reproduces the same feature-poor behavior and drives up refusals. Together, these results indicate that refusals on benign prompts can arise from **capability deficits** rather than fixed safety rules, and that ideological depth is a measurable property of LLMs that helps predict when a model will refuse.

Figures

Figures reproduced from arXiv: 2508.21448 by the authors.

Figure 2
Figure 2. Response heatmap over conditions: Llama tends to result in more refusal responses, [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 1
Figure 1. Data distribution over topics which sometimes modifies some of them, a sce￾nario known as context rot [12]. Therefore, in step 2, we collect all the categories identified by Gemini and select one statement from the original list, and again prompt the LLM to pick the topic that best matches the statement. To account for selection bias of the LLMs – where they are more likely to pick an option based on their position … view at source ↗
Figure 3
Figure 3. Candidate’s null (refusal) response rate [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (11 more)
Figure 4
Figure 4. Figure 4: Candidate’s response consistency over conditions: Models show the least consistency (most variance) in their responses while answering as a conservative or provided with conservative arguments. As shown in [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: Candidate’s response consistency over topics while responding as a conservative. Abortion Rights Climate & Environment Corporate & Economic Regulation Gun Control Healthcare Immigration & Refugees LGBTQ+ Rights Military & Defense Spending Political & Ideological Stance…
Figure 7
Figure 7. Figure 7: Estimated ideal points of the models using the IRT model, over the 9 prompting conditions. [PITH_FULL_IMAGE:figures/full_fig_p005_7.png]
Figure 8
Figure 8. Figure 8: Candidates Layer sweep for political steerability. Layers 14 and 20 were selected for the [PITH_FULL_IMAGE:figures/full_fig_p005_8.png]
Figure 9
Figure 9. Figure 9: Model behaviour due to steering Additionally, we also applied steering using SAEs, using the STA (steering target atoms) approach from Wang et al. [27], where we select target features using their activation value (amplitude) and frequencies from the SAE of the target …
Figure 10
Figure 10. Figure 10: Candidate’s response consistency over topics when steered to be conservative. [PITH_FULL_IMAGE:figures/full_fig_p006_10.png]
Figure 11
Figure 11. Figure 11: Estimated ideal points of the steered models with multiplier +1 and -1 corresponding to [PITH_FULL_IMAGE:figures/full_fig_p006_11.png]
Figure 12
Figure 12. Figure 12: Evaluation scores distribution. Dominant Themes in 14-llamascope-res-131k Dominant Themes in 20-gemmascope-res-131k [PITH_FULL_IMAGE:figures/full_fig_p008_12.png]
Figure 14
Figure 14. Figure 14: Classification Confusion Matrix. Finally, the confusion matrices in [PITH_FULL_IMAGE:figures/full_fig_p009_14.png]
Figure 16
Figure 16. Figure 16: Eigenvalue Table (Unrotated Setting) to the explained variance. Following the Kaiser criterion (eigenvalues greater than 1), eight factors were retained for rotation ( [PITH_FULL_IMAGE:figures/full_fig_p013_16.png]
Figure 17
Figure 17. Figure 17: Model Performance by Identification Strategy. Our initial model specification attempted to achieve identification solely through the speci￾fication of priors, particularly on the correlation matrices for the ideal points (Ωθ) and discrim￾ination parameters (Ωα). While…

Discussion (0). Sign in 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. Auditing Alignment Controllability in LLMs via Political Axes

    cs.CY 2026-07 conditional novelty 6.0 of 10

    On a 63,700-response Political Compass stress test of seven frontier LLMs, system-prompt framing dominates model identity, and steerability needs dispersion, symmetry, saturation, and refusal-floor metrics.

Reference graph

Works this paper leans on

39 extracted references · 28 canonical work pages · cited by 1 Pith paper

  1. [1]

    Aldahoul, H

    N. Aldahoul, H. Ibrahim, M. Varvello, A. Kaufman, T. Rahwan, and Y . Zaki. Large language models are often politically extreme, usually ideologically inconsistent, and persuasive even in informational contexts. arXiv preprint arXiv:2505.04171, 2025

    arXiv 2025

  2. [2]

    Ameisen, J

    E. Ameisen, J. Lindsey, A. Pearce, W. Gurnee, N. L. Turner, B. Chen, C. Citro, D. Abra- hams, S. Carter, B. Hosmer, J. Marcus, M. Sklar, A. Templeton, T. Bricken, C. McDougall, H. Cunningham, T. Henighan, A. Jermyn, A. Jones, A. Persic, Z. Qi, T. Ben Thompson, S. Zimmerman, K. Rivoire, T. Conerly, C. Olah, and J. Batson. Circuit tracing: Reveal- ing compu...

    work page 2025

  3. [3]

    D. Arad, A. Mueller, and Y . Belinkov. Saes are good for steering – if you select the right features, 2025. URL https://arxiv.org/abs/2505.20063

    arXiv 2025

  4. [4]

    E. M. Bender, T. Gebru, A. McMillan-Major, and S. Shmitchell. On the dangers of stochastic parrots: Can language models be too big? In Proceedings of the 2021 ACM conference on fairness, accountability, and transparency, pages 610–623, 2021. 10

    work page 2021

  5. [5]

    Bricken, A

    T. Bricken, A. Templeton, J. Batson, B. Chen, A. Jermyn, T. Conerly, N. Turner, C. Anil, C. Deni- son, A. Askell, R. Lasenby, Y . Wu, S. Kravec, N. Schiefer, T. Maxwell, N. Joseph, Z. Hatfield- Dodds, A. Tamkin, K. Nguyen, B. McLean, J. E. Burke, T. Hume, S. Carter, T. Henighan, and C. Olah. Towards monosemanticity: Decomposing language models with dictio...

    work page 2023

  6. [6]

    Caliskan, J

    A. Caliskan, J. J. Bryson, and A. Narayanan. Semantics derived automatically from language corpora contain human-like biases. Science, 356(6334):183–186, 2017

    work page 2017

  7. [7]

    Chiang and H.-y

    C.-H. Chiang and H.-y. Lee. A closer look into using large language models for automatic evaluation. In H. Bouamor, J. Pino, and K. Bali, editors, Findings of the Association for Computational Linguistics: EMNLP 2023, pages 8928–8942, Singapore, Dec. 2023. Association for Computational Linguistics. doi: 10.18653/v1/2023.findings-emnlp.599. URL https: //ac...

    doi:10.18653/v1/2023.findings-emnlp.599 2023

  8. [8]

    Clinton, S

    J. Clinton, S. Jackman, and D. Rivers. The statistical analysis of roll call data. American Political Science Review, 98(2):355–370, 2004

    work page 2004

Show all 39 references
  1. [9]

    Elhage, T

    N. Elhage, T. Hume, C. Olsson, N. Schiefer, T. Henighan, S. Kravec, Z. Hatfield-Dodds, R. Lasenby, D. Drain, C. Chen, R. Grosse, S. McCandlish, J. Kaplan, D. Amodei, M. Wattenberg, and C. Olah. Toy models of superposition, 2022. URL https://arxiv.org/abs/2209. 10652

    2022

  2. [10]

    Gurnee and M

    W. Gurnee and M. Tegmark. Language models represent space and time. arXiv preprint arXiv:2310.02207, 2023

    2023 arXiv

  3. [11]

    Z. He, W. Shu, X. Ge, L. Chen, J. Wang, Y . Zhou, F. Liu, Q. Guo, X. Huang, Z. Wu, Y .-G. Jiang, and X. Qiu. Llama scope: Extracting millions of features from llama-3.1-8b with sparse autoencoders, 2024. URL https://arxiv.org/abs/2410.20526

    2024 arXiv

  4. [12]

    K. Hong, A. Troynikov, and J. Huber. Context rot: How increasing input tokens impacts llm performance. Technical report, Chroma, July 2025. URL https://research.trychroma. com/context-rot

    2025

  5. [13]

    Kabir, K

    S. Kabir, K. Esterling, and Y . Dong. Do words reflect beliefs? evaluating belief depth in large language models, 2025. URL https://arxiv.org/abs/2504.17052

    2025 arXiv

  6. [14]

    J. Kim, J. Evans, and A. Schein. Linear representations of political perspective emerge in large language models, 2025. URL https://arxiv.org/abs/2503.02080

    2025 arXiv

  7. [15]

    Lieberum, S

    T. Lieberum, S. Rajamanoharan, A. Conmy, L. Smith, N. Sonnerat, V . Varma, J. Kramar, A. Dragan, R. Shah, and N. Nanda. Gemma scope: Open sparse autoencoders everywhere all at once on gemma 2. In Y . Belinkov, N. Kim, J. Jumelet, H. Mohebbi, A. Mueller, and H. Chen, editors, P...

    2024 doi

  8. [16]

    Lindsey, W

    J. Lindsey, W. Gurnee, E. Ameisen, B. Chen, A. Pearce, N. L. Turner, C. Citro, D. Abrahams, S. Carter, B. Hosmer, J. Marcus, M. Sklar, A. Templeton, T. Bricken, C. McDougall, H. Cun- ningham, T. Henighan, A. Jermyn, A. Jones, A. Persic, Z. Qi, T. B. Thompson, S. Zimmerman, K. ...

    2025

  9. [17]

    Linegar, R

    M. Linegar, R. Kocielnik, and R. M. Alvarez. Large language models and political science. Frontiers in Political Science, 5:1257092, 2023

    2023

  10. [18]

    N. Nanda. Attribution patching: Activation patching at industrial scale. URL: https://www. neelnanda. io/mechanistic-interpretability/attribution-patching, 2023. 11

    2023

  11. [19]

    T.-Q. Peng, K. Yang, S. Lee, H. Li, Y . Chu, Y . Lin, and H. Liu. Unpacking political bias in large language models: Insights across topic polarization, 2025. URL https://arxiv.org/abs/ 2412.16746v1

    2025

  12. [20]

    Perez, S

    E. Perez, S. Ringer, K. Lukosiute, K. Nguyen, E. Chen, S. Heiner, C. Pettit, C. Olsson, S. Kundu, S. Kadavath, et al. Discovering language model behaviors with model-written evaluations. In Findings of the association for computational linguistics: ACL 2023, pages 13387–13434, 2023

    2023

  13. [21]

    K. T. Poole and H. Rosenthal. A spatial model for legislative roll call analysis. American journal of political science, pages 357–384, 1985

    1985

  14. [22]

    Rimsky, N

    N. Rimsky, N. Gabrieli, J. Schulz, M. Tong, E. Hubinger, and A. Turner. Steering llama 2 via contrastive activation addition. In L.-W. Ku, A. Martins, and V . Srikumar, editors,Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: L...

    2024 doi

  15. [23]

    Röttger, V

    P. Röttger, V . Hofmann, V . Pyatkin, M. Hinck, H. Kirk, H. Schütze, and D. Hovy. Political compass or spinning arrow? towards more meaningful evaluations for values and opinions in large language models. In Proceedings of the 62nd Annual Meeting of the Association for Computa...

    2024

  16. [24]

    Santurkar, E

    S. Santurkar, E. Durmus, F. Ladhak, C. Lee, P. Liang, and T. Hashimoto. Whose opinions do language models reflect? In International Conference on Machine Learning, pages 29971– 30004. PMLR, 2023

    2023

  17. [25]

    Templeton, T

    A. Templeton, T. Conerly, J. Marcus, J. Lindsey, T. Bricken, B. Chen, A. Pearce, C. Citro, E. Ameisen, A. Jones, H. Cunningham, N. L. Turner, C. McDougall, M. MacDiarmid, C. D. Freeman, T. R. Sumers, E. Rees, J. Batson, A. Jermyn, S. Carter, C. Olah, and T. Henighan. Scaling m...

    2024

  18. [26]

    E. Todd, M. L. Li, A. S. Sharma, A. Mueller, B. C. Wallace, and D. Bau. Function vectors in large language models, 2024. URL https://arxiv.org/abs/2310.15213

    2024 arXiv

  19. [27]

    M. Wang, Z. Xu, S. Mao, S. Deng, Z. Tu, H. Chen, and N. Zhang. Beyond prompt engineering: Robust behavior control in llms via steering target atoms, 2025. URL https://arxiv.org/ abs/2505.20322

    2025 arXiv

  20. [28]

    Zheng, H

    C. Zheng, H. Zhou, F. Meng, J. Zhou, and M. Huang. Large language models are not robust multiple choice selectors. In ICLR, 2024

    2024

  21. [29]

    A. Zou, L. Phan, S. Chen, J. Campbell, P. Guo, R. Ren, A. Pan, X. Yin, M. Mazeika, A.-K. Dombrowski, S. Goel, N. Li, M. J. Byun, Z. Wang, A. Mallen, S. Basart, S. Koyejo, D. Song, M. Fredrikson, J. Z. Kolter, and D. Hendrycks. Representation engineering: A top-down approach to...

    2025 arXiv

  22. [30]

    and IDEAL scores [8]. 13 B.1 Model Specification The core of our model is a two-parameter logistic D-dimensional IRT, where the probability of a candidate j voting "yes" on item k is given by the liklihood function: P (yjk = 1) = logit−1 (θj · αk − βk) yjk ∼ Bernoulli (logit−1...

  23. [31]

    Test Statement

    Read the original "Test Statement" carefully

  24. [32]

    Extracted Features

    Analyze the provided "Extracted Features."

  25. [33]

    Extracted Features,

    Based ONLY on the "Extracted Features," classify the "Test Statement" into one of the following categories: {category_list}

  26. [34]

    Rate your confidence in this classification on a scale of 1 to 5, where 1 is a pure guess and 5 is highly confident

  27. [35]

    {statement}

    Provide a brief, one-sentence justification for your classification, explaining which features were most influential. **STRICT INSTRUCTION:** Your classification MUST be based *solely* on the provided list of features, not on your prior knowledge of the statement itself. **INP...

  28. [36]

    List of Activated Features

    Carefully review the "List of Activated Features."

  29. [37]

    - **1:** No clear theme

    On a scale of 1 to 5, rate the **thematic coherence** of the feature set. - **1:** No clear theme. The features seem random and unrelated. - **3:** A weak theme is present, but many features are unrelated to the core concept. - **5:** Highly coherent. All features clearly rela...

  30. [38]

    Critique of social welfare spending

    In one phrase or sentence, describe the primary theme that unifies these features (e.g., "Critique of social welfare spending" or "Analysis of conditional legal language")

  31. [39]

    coherence_score

    Provide a brief justification for your score, noting any outlier features that do not fit the main theme. **INPUT:** * **List of Activated Features (from Llama/Gemma):** {feature_str} 16 **OUTPUT FORMAT (JSON):** ‘‘‘json { "coherence_score": <1-5 integer>, "primary_theme": "Yo...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.