Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

Fine-Grained Interpretation of Political Opinions in Large Language Models

T0 review · 5 major / 6 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 Four political axes untangle LLM opinions beyond left-right

desk verdict A solid, review-worthy empirical study of four-dimensional political concept vectors in LLMs, but the core disentanglement claim needs a null test that removes the label-split artifact. read the letter →

arxiv 2506.04774 v1 pith:5KSAJT2E submitted 2025-06-05 cs.CL cs.AIcs.LG

Jingyu Hu , Mengyue Yang , Mengnan Du , Weiru Liu This is my paper
classification cs.CLcs.AIcs.LG
keywords LLMpoliticalbiasfine-graineddimensionsconceptvectorsrepresentationengineeringactivationsteeringdisentanglementout-of-distributiongeneralizationinterpretability
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's central claim is that a single left–right axis is too coarse to capture how large language models encode political opinion, and that four dimensions—economic, diplomatic, civil, and society—can separate concepts that a one-dimensional view conflates. On this scheme the authors build a dataset of roughly ten thousand left/right statements, train directional concept vectors in each model layer with three representation-engineering methods, and show that same-dimension vectors stay correlated while cross-dimension correlations fade in deeper layers. They also show that these vectors detect political leanings in held-out and out-of-distribution statements more accurately than a single-axis baseline, and that adding scaled versions of the vectors into mid-layer hidden states shifts generated responses left or right. If right, the work provides a transparent, fine-grained way to probe and steer political states that ordinary response evaluation misses.

What carries the argument

The machinery is the layer-wise political concept vector: a unit direction in a transformer's hidden-state space that points along a left/right contrast within one dimension. Three learning rules are compared: CAA takes the mean difference between left and right statement embeddings, RepE takes the first principal component of those differences, and linear probing fits a logistic regression whose weight vector becomes the concept direction. All three operate on the last token's hidden state at each layer. Detection is the sign of the dot product between a test embedding and the vector; intervention adds $\alpha \vec{u}^\ell$ to the hidden state at chosen layers. The vectors' load-bearing property is their measured correlation structure: strong within-dimension, weak across dimensions in later layers, which the paper reads as disentanglement of confounded political concepts.

What would settle it

Build a test set of deliberately cross-dimensional statements—for example, a tax policy framed as a free-speech issue—and check whether the layer-28 cosine correlation between the economic 'equality' and civil 'liberty' vectors stays near zero. If the correlation rises when the test statements mix dimensions, the claimed disentanglement is a labeling artifact rather than a robust property of the learned vectors.

Watch

Extended reading notes

Core claim

The discovery is that political concept directions in LLM hidden states are not a single left/right continuum but a set of per-dimension directions that can be learned separately and that become weakly correlated across dimensions in deeper layers. The authors define left and right relative to four concepts: economic equality vs market, diplomatic globe vs nation, civil liberty vs authority, and society progress vs tradition. From contrastive statements on these dimensions, they obtain eight concept vectors per layer per model, verify internal consistency by cosine similarity, and use the vectors as linear probes and as steering additions $h^\ell + \alpha \vec{u}^\ell$. The key evidence is that single-axis vectors confuse concepts—for example, economic 'equality' (left) correlates strongly with civil 'authority' (right) at layer 8—whereas the fine-grained vectors separate them by layer 28, and that fine-grained probes retain accuracy on out-of-distribution data where an unfine-grained baseline drops.

Load-bearing premise

The load-bearing premise is that the four political dimensions do not overlap, so the training embeddings for different dimensions are already separated; if a statement about taxes is also a statement about liberty, the disentanglement seen in later layers could be an artifact of training separate vectors on separately labeled data rather than a property of the model.

Editorial extensions

If this is right

  • Concept-vector correlation analysis gives a concrete signature of political concept confounds in LLM internals and a way to check when those confounds resolve across layers.
  • Fine-grained political probes keep high detection accuracy on out-of-distribution statements, whereas single-axis probes degrade, so internal political-state auditing is more reliable when split by dimension.
  • Mid-layer activation additions of one of these vectors shift generated responses toward the targeted leaning, and with a well-chosen strength the shift happens without losing coherence, while over-intervention makes text less readable.
  • LogitLens tracing shows interventions change next-token intentions, but a late-layer 'neutral' or 'subject' guardrail can override those intentions, so final answers understate the model's internal political state.
  • The approach transfers across eight open-source models between 1B and 8B parameters, with linear probing consistently the strongest detection method.

Reading between the lines

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

  • The paper's four dimensions are author-chosen; the same method could be applied to other ideology taxonomies or national contexts, but the current data is largely U.S.-sourced, which may bias the learned vectors toward U.S. political alignments, as the authors note for healthcare.
  • The disentanglement result may partly reflect that each dimension has its own contrastively labeled training set; an unlabeled or adversarially mixed test would tell whether the model's geometry is genuinely factorized.
  • If the internal-probe method is sound, public evaluations that rely on final text could miss latent political bias, because the paper observes guardrail behavior where final outputs refuse while internal intentions were already steered.
  • Combining two or more dimension vectors in a single intervention is a natural next step that the paper lists as future work, and their interactions could be quantified with the same correlation machinery.
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 / 6 minor

Summary. The paper proposes a four-dimensional political framework (economic, diplomatic, civil, society) for learning political concept vectors from LLM internals using CAA, RepE, and linear probing on a GPT-4o-generated dataset built from AllSides news events. The central claims are (RQ1) that the resulting vectors disentangle political concept confounds, (RQ2) that they detect political leanings with good OOD generalization relative to a single left-right axis, and (RQ3) that they can intervene in LLMs to generate responses with different political leanings. Experiments are reported across eight open-source models of varying size and family.

Significance. If substantiated, the framework would be a useful step beyond single-axis political bias measurement and would provide a practical recipe for probing and steering internal political representations. The empirical footprint is substantial: eight models, three representation-engineering methods, an OOD evaluation, and a prompt-template ablation, with linear-probing detection accuracies consistently above 0.91. However, the most distinctive claim—the disentanglement of political dimensions—is currently under-tested, and the evidence presented may partly reflect the training and label structure rather than a genuine property of LLM internals. The OOD comparison and intervention evaluation also need tightening before the central claims can be accepted.

major comments (5)
  1. [Experiment Setups – Disentangled Concept Vector Learning] The central RQ1 claim is not established by the reported correlation analysis. The paper states, "Since our defined dimensions are non-overlapping, we assume the training data embeddings under different dimensions are disentangled," and then treats the decreasing cross-vector correlations in Figure 6 and Appendix Figure 11 as evidence of disentanglement. Because the eight vectors are trained on disjoint GPT-4o-generated datasets with distinct concept keywords and topic sets under each dimension, the observed decorrelation at deeper layers can be fully explained by the label-split and topic structure rather than by a genuine separation of political dimensions in the model's internal semantics. The analysis never reports correlations between full statement embeddings across dimensions, so it cannot distinguish representation geometry from a training artifact. Moreover, the "high within-dimension correlation" criterion is partly tautological: for RepE and CAA the right-leaning vector is the negative of the left-leaning vector, so the ±1 within-dimension entries in the heatmaps hold by construction. To support the claim, the paper should report cross-dimension similarities of raw activations (e.g., centroid distances, CKA, or statement-embedding correlations with topics controlled) and include a control where dimension labels are permuted or topics are shared across dimensions.
  2. [Appendix – Intervention Task] The paper itself concedes a cross-dimension confound in the intervention setting: it states that "there exist certain cases (e.g., healthcare discussions) where LLMs resist steering" and attributes this to the AllSides corpus being mainly U.S. news, so that "market-oriented concept may influence certain topics (e.g., healthcare)." This is precisely an economic-by-society confound that the framework is claimed to remove, and it undercuts the non-overlapping premise on which the disentanglement analysis is built. The authors should either explicitly bound the scope of the disentanglement claim or provide a quantitative analysis of how much the U.S.-dominated corpus shifts each dimension vector, and show that the main RQ1/RQ2 results are robust when such topics are excluded.
  3. [RQ2 – Out-of-Distribution Data (Figure 8)] The OOD comparison is not matched. In the left panel of Figure 8, each fine-grained probe is evaluated on OOD data from its own dimension while the single-axis baseline reports average performance over all four dimensions; the right panel compares a per-dimension probe against a single global baseline without clarifying whether the baseline was trained on the same combined data, with the same balance, and at the same layer selection. Without matched evaluation sets, the claim that fine-grained vectors "generalise to more diverse data distributions" is not supported. Please provide a per-dimension matched comparison, error bars or significance tests, and a precise description of the baseline's training data and size.
  4. [Experiment Setups – Data Construction] The dataset is constructed by prompting GPT-4o to generate left/right statements from concept and topic keywords, and no human validation is reported anywhere in the paper. Consequently, the in-distribution detection accuracies in Table 3 and Figure 7 may measure how well the probes reproduce GPT-4o's labeling conventions rather than an independent political ground truth. The "semantic meaning" claim in RQ2 needs at least a human-validated subset, an estimate of label noise or agreement, and an analysis of whether the OOD advantage persists when labels are re-derived from the Eight Values and Rate Your Bias instruments without GPT-4o's generation step.
  5. [RQ3 – Intervention Ability] RQ3's conclusion rests on qualitative examples: one LogitLens analysis for Mistral-7B and a few response transcripts for a single prompt. The appendix acknowledges cases where the model "resists steering," but no success rate, no systematic metric, and no evaluation across the eight models and four dimensions is reported. The strength coefficient α is apparently chosen by inspection, and the paper notes that α=2.5 degrades coherence. Without a quantitative evaluation protocol (e.g., change in the political classification of steered outputs, or a judge-based metric over many prompts and dimensions), the claim that the vectors "can intervene in LLMs to generate responses with different political leanings" is not established at the level claimed.
minor comments (6)
  1. [Related Work] There is a duplicated phrase: "LLMs show that LLMs show systematic bias in favor of certain political parties."
  2. [Table 2] The dimension notation is inconsistent: the table labels the diplomatic row as Dcivil and the civil row as Ddip, whereas the text defines Dim={eco,dip,civil,soc} with the natural correspondence.
  3. [Methodology, Eq. (4)] Equation (4) is not well-defined: the sums are both over the whole dimension dataset Dd, so SL and SR do not denote disjoint left and right subsets; the notation should be corrected to explicit left and right subsets.
  4. [Related Work / Method References] The citation for CAA is incorrect: the paper cites Belinkov (2022), a probing-classifier paper, for Contrastive Activation Addition; the relevant method reference is the Panickssery et al. (2023) entry already in the reference list.
  5. [Table 6] There are typographical errors: "regualtion" should be "regulation" and "notionalist" should presumably be "nationalist."
  6. [Figure 4] The caption does not state how the OOD samples were selected or whether the 100-sample balancing applies to all three sets; please clarify.

Circularity Check

3 steps flagged · score 4.0 of 10

RQ1's disentanglement evidence is partly self-confirming: same-dimension concept-vector correlations are ±1 by construction, and the four dimensions' non-overlap is assumed as a premise rather than established; layer-dependent decorrelation and external OOD detection keep the central claim partially empirical.

  1. self definitional [Methodology, 'Political Concept Vector Learning' (Eq. 4, RepE direction convention); Experiment Setups, 'Disentangled Concept Vector Learning'; RQ1, 'Correlation Analysis']
    "Unless otherwise specified, we set the default positive direction of ⃗ uℓ to present lean left concept CL, and its opposite as lean right concept CR. ... The concept vectors are disentangled if they show high correlations within the same dimension and weak correlations across different dimensions."

    For CAA (Eq. 4), RepE, and linear probing, the left and right vectors of one dimension are the same direction with opposite signs (Eq. 4's mean-difference u⃗ℓ, and the single logistic weight w^ℓ_c from Eq. 5-6), so the cosine similarity between same-dimension pairs is identically ±1 at every layer of every model. Appendix Figure 11 confirms this: Globe-Nation, Liberty-Authority, Progress-Tradition, and Equality-Market are exactly -1.00/1.00 from layer 1 through layer 32. The paper's first disentanglement criterion ('high correlations within the same dimension') is therefore satisfied by construction and carries no information about whether LLM internals separate the four political dimensions.

  2. self definitional [Experiment Setups, 'Disentangled Concept Vector Learning'; RQ1, 'Correlation Analysis']
    "Since our defined dimensions are non-overlapping, we assume the training data embeddings under different dimensions are disentangled. By learning concept vectors across these distinct dimensions, we obtain disentangled representations that address the concept confounds. ... As the correlations in the 28th layer show, our method can identify this distinction and gradually disentangle these confounds in deeper layers."

    The non-overlap of the four author-defined dimensions is asserted, and the deep-layer half of RQ1's conclusion ('gradually disentangle these confounds') restates that assertion: disentanglement of training-data embeddings is the premise used to justify learning separate vectors, then reported as the finding.

1 more flagged steps
  1. other [Appendix, 'Intervention Task' (self-admitted limitation)]
    "there exist certain cases (e.g., healthcare discussions) where LLMs resist steering. One possible cause relates to the data source: although we collected many events (news) on diverse topics from AllSides, they are mainly U.S. news. ... therefore, if U.S. economic news dominates the corpus, market-oriented concept may influence certain topics (e.g., healthcare)."

    This is the authors' own concession that the load-bearing premise fails: the economic dimension's vector is pulled toward market-oriented content on a society-dimension topic (healthcare), i.e., a cross-dimension confound that the framework was claimed to remove. It confirms that 'our defined dimensions are non-overlapping' is an assumption, not an established property of the data, so the RQ1 disentanglement result rests on the very premise it claims to demonstrate. The passage is weighed here per the review rule that self-admitted limitations count as in-scope evidence.

full rationale

The paper's derivation chain is: four author-defined dimensions with GPT-4o-generated left/right statements per dimension → concept vectors per layer (CAA/RepE/linear probing) → correlation analysis (RQ1) → detection including OOD (RQ2) → intervention (RQ3). The genuinely empirical, non-circular parts are substantial: the correlation analysis shows a layer-dependent evolution (cross-dimension correlations of 0.85-0.98 at layers 1-8 collapsing to roughly 0.0-0.17 by layer 28), which is NOT forced by construction — if separate training forced decorrelation, the early layers would already be clean. Detection is evaluated on held-out data plus genuinely external OOD sets (8values quiz, Rate Your Bias), and the fine-grained-vs-single-axis OOD comparison is an independent benchmark. Intervention is a standard activation-steering demonstration. No load-bearing self-citation chain exists: self-citations (Zhao et al. 2024; Jin et al. 2025; Hu, Liu, Du 2024) are methodological or contextual, no uniqueness theorem is imported from prior author work, and the four-axis framework is explicitly credited to Eight Values rather than disguised as new. The circularity is confined to RQ1's validation framing: (1) the criterion's 'high within-dimension correlation' half is identically ±1 by construction of the vector-learning equations, as Figure 11 shows at every layer; (2) the non-overlap of dimensions is assumed as a premise ('we assume the training data embeddings under different dimensions are disentangled') and the deep-layer conclusion partly restates it; and (3) the appendix concedes a U.S.-news-driven economic-society confound, contradicting the premise. These self-definitional elements affect the headline disentanglement claim, but the layer-dependence and the external OOD results keep the central contribution partially independent. Score 4 reflects moderate, partial circularity, not a forced derivation.

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

The central claims rest on author-defined political dimensions, the linear representation hypothesis, and LLM-generated labels without human validation. No new physical or ontological entities are introduced; the 'concept vectors' are standard learned representations.

free parameters (3)
  • Four-dimensional political concept definitions
    The dimensions (economic, diplomatic, civil, society) and the paired left/right concepts (e.g., equality/market) are chosen by hand. All downstream vectors and correlation analyses are defined relative to this choice.
  • Intervention strength alpha = αL=2, αR=1.5, 2, 2.5 in examples
    The steering effect and output coherence depend strongly on the manually chosen alpha; no systematic tuning or search is reported.
  • Multi-layer intervention range = layers 15 to 25
    The effective layer range for multi-layer intervention is selected based on single-layer experiments; this choice affects the intervention outcomes.
assumptions (4)
  • domain assumption The linear representation hypothesis: political concepts are encoded as linear directions in LLM hidden states.
    Invoked in Methodology, Political Concept Vector Learning, citing Park et al. 2023 and Elhage et al. 2022. The whole vector learning procedure depends on this assumption.
  • ad hoc to paper The four dimensions are non-overlapping, so training data embeddings under different dimensions are disentangled.
    Stated in Experiment Setups, Disentangled Concept Vector Learning. This is load-bearing for the disentanglement claim; if false, the reported separation may be an artifact of the training setup.
  • domain assumption GPT-4o-generated statements correctly represent the intended left/right political concepts.
    Dataset construction uses GPT-4o to generate statements from news headlines with author-defined concept prompts; no human validation is reported. If GPT-4o's own political biases skew the labels, the vectors inherit them.
  • domain assumption The OOD datasets (8values, Rate Your Bias) are a valid external benchmark for political leaning.
    Used in RQ2 to claim generalization; these are online questionnaires, not a gold-standard academic benchmark.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Fine-Grained Interpretation of Political Opinions in Large Language Models." pith.science (2026). https://pith.science/paper/5KSAJT2E

@misc{pith2026250604774,
  author       = {Pith},
  title        = {Pith review of: Fine-Grained Interpretation of Political Opinions in Large Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5KSAJT2E}},
  note         = {Machine review of arXiv:2506.04774}
}
read the original abstract

Studies of LLMs' political opinions mainly rely on evaluations of their open-ended responses. Recent work indicates that there is a misalignment between LLMs' responses and their internal intentions. This motivates us to probe LLMs' internal mechanisms and help uncover their internal political states. Additionally, we found that the analysis of LLMs' political opinions often relies on single-axis concepts, which can lead to concept confounds. In this work, we extend the single-axis to multi-dimensions and apply interpretable representation engineering techniques for more transparent LLM political concept learning. Specifically, we designed a four-dimensional political learning framework and constructed a corresponding dataset for fine-grained political concept vector learning. These vectors can be used to detect and intervene in LLM internals. Experiments are conducted on eight open-source LLMs with three representation engineering techniques. Results show these vectors can disentangle political concept confounds. Detection tasks validate the semantic meaning of the vectors and show good generalization and robustness in OOD settings. Intervention Experiments show these vectors can intervene in LLMs to generate responses with different political leanings.

Figures

Figures reproduced from arXiv: 2506.04774 by the authors.

Figure 1
Figure 1. An Example of Left-Right Concept Confounds [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. The Proposed Fine-grained Political Learning Within LLMs’ Internal States [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. An Illustration of Constructing Left-leaning and Right-leaning Statements at Economic Dimension [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Data Distribution Comparisons Among In￾distribution Dtrain, Dtest and Out-of-distribution (OOD) Data to the held-out Dtest, and OOD data consists of state￾ments based on the Eight Values Questionnaire4 and Rate Your Bias5 [PITH_FULL_IMAGE:figures/full_fig_p006_4.png]
Figure 5
Figure 5. Figure 5: Representations at Different Layers of LLMs (Blue = left-leaning, Red = right-leaning). The first three subplots are [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Concept Vector Correlation Analysis at ℓ = {8, 28} on Llama3-8B. (Left Two) Concept vectors learned with linear probing. (Right Two) Concept vectors learned with RepE. confounds in deeper layers. These fine-grained concept vec￾tors exhibit strong within-dimension corre…
Figure 7
Figure 7. Figure 7: Detection Performance of Linear Probing Across Layers in Llama3-8B. (First two) Performance on [PITH_FULL_IMAGE:figures/full_fig_p008_7.png]
Figure 8
Figure 8. Figure 8: Detection Performance Comparison on OOD Data Between Our Fine-Grained Probes (the first four rows) and the [PITH_FULL_IMAGE:figures/full_fig_p009_8.png]
Figure 9
Figure 9. Figure 9: Distribution Shift on Single Layer Intervention. (Left Three) Intervened at different layers [PITH_FULL_IMAGE:figures/full_fig_p009_9.png]
Figure 10
Figure 10. Figure 10: LogitLens Visualization on Mistral-7B. It shows the top candidate tokens for the next word, based on the hidden [PITH_FULL_IMAGE:figures/full_fig_p010_10.png]
Figure 11
Figure 11. Figure 11: RepE Political Concept Vectors Correlations across Layers (Llama3-8B) [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]
Figure 12
Figure 12. Figure 12: Single Layer Intervention. The circles ◦ refer to the original data distribution, while the crosses × show the distribu￾tion after the intervention on the original data. The colors denote political leaning (left/right). When the intervention direction opposes the data…

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. Agent Security Needs Redefinition through a Holistic Framework

    cs.CR 2026-07 conditional novelty 6.0 of 10

    Agent security should be redefined around four contextual authorization properties instead of the content of the action performed.

Reference graph

Works this paper leans on

55 extracted references · 20 canonical work pages · cited by 1 Pith paper

  1. [1]

    Agiza, A.; Mostagir, M.; and Reda, S. 2024. Politune: Analyzing the impact of data selection and fine-tuning on economic and political biases in large language models. In Proceedings of the AAAI/ACM Conference on AI, Ethics, and Society, volume 7, 2--12

    work page 2024

  2. [2]

    Ball, S.; Kreuter, F.; and Panickssery, N. 2024. Understanding jailbreak success: A study of latent space dynamics in large language models. arXiv preprint arXiv:2406.09289

    arXiv 2024

  3. [3]

    Belinkov, Y. 2022. Probing classifiers: Promises, shortcomings, and advances. Computational Linguistics, 48(1): 207--219

    2022

  4. [4]

    M.; Gebru, T.; McMillan-Major, A.; and Shmitchell, S

    Bender, E. M.; Gebru, T.; McMillan-Major, A.; and Shmitchell, S. 2021. On the Dangers of Stochastic Parrots: Can Language Models Be Too Big? In Proceedings of the 2021 ACM Conference on Fairness, Accountability, and Transparency . New York: Association for Computer Machinery – ACM

    work page 2021

  5. [5]

    Bhalla, U.; Srinivas, S.; Ghandeharioun, A.; and Lakkaraju, H. 2024. Towards unifying interpretability and control: Evaluation via intervention. arXiv preprint arXiv:2411.04430

    arXiv 2024

  6. [6]

    C.; Straitouri, E.; Tsirtsis, S.; and Gomez-Rodriguez, M

    Chatzi, I.; Benz, N. C.; Straitouri, E.; Tsirtsis, S.; and Gomez-Rodriguez, M. 2024. Counterfactual token generation in large language models. arXiv preprint arXiv:2409.17027

    arXiv 2024

  7. [7]

    C.; Patel, O.; Riecke, J.; Raval, S.; Seow, O.; et al

    Chen, Y.; Wu, A.; DePodesta, T.; Yeh, C.; Li, K.; Marin, N. C.; Patel, O.; Riecke, J.; Raval, S.; Seow, O.; et al. 2024. Designing a dashboard for transparency and control of conversational AI. arXiv preprint arXiv:2406.07882

    arXiv 2024

  8. [8]

    Chu, Z.; Wang, Y.; Li, L.; Wang, Z.; Qin, Z.; and Ren, K. 2024. A causal explainable guardrails for large language models. In Proceedings of the 2024 on ACM SIGSAC Conference on Computer and Communications Security, 1136--1150

    work page 2024

Show all 55 references
  1. [9]

    Dathathri, S.; Madotto, A.; Lan, J.; Hung, J.; Frank, E.; Molino, P.; Yosinski, J.; and Liu, R. 2019. Plug and play language models: A simple approach to controlled text generation. arXiv preprint arXiv:1912.02164

    2019 arXiv

  2. [10]

    Duetting, P.; Mirrokni, V.; Paes Leme, R.; Xu, H.; and Zuo, S. 2024. Mechanism design for large language models. In Proceedings of the ACM Web Conference 2024, 144--155

    2024

  3. [11]

    Elhage, N.; Hume, T.; Olsson, C.; Schiefer, N.; Henighan, T.; Kravec, S.; Hatfield-Dodds, Z.; Lasenby, R.; Drain, D.; Chen, C.; et al. 2022. Toy models of superposition. arXiv preprint arXiv:2209.10652

    2022 arXiv

  4. [12]

    Goldowsky-Dill, N.; Chughtai, B.; Heimersheim, S.; and Hobbhahn, M. 2025. Detecting Strategic Deception Using Linear Probes. arXiv preprint arXiv:2502.03407

    2025 arXiv

  5. [13]

    Grattafiori, A.; Dubey, A.; Jauhri, A.; Pandey, A.; Kadian, A.; Al-Dahle, A.; Letman, A.; Mathur, A.; Schelten, A.; Vaughan, A.; et al. 2024. The llama 3 herd of models. arXiv preprint arXiv:2407.21783

    2024 arXiv

  6. [14]

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

    2023 arXiv

  7. [15]

    He, J.; Gong, Y.; Lin, Z.; Wei, C.; Zhao, Y.; and Chen, K. 2024. Llm factoscope: Uncovering llms’ factual discernment through measuring inner states. In Findings of the Association for Computational Linguistics ACL 2024, 10218--10230

    2024

  8. [16]

    Hu, J.; Liu, W.; and Du, M. 2024. Strategic Demonstration Selection for Improved Fairness in LLM In-Context Learning. In Al-Onaizan, Y.; Bansal, M.; and Chen, Y.-N., eds., Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 7460--7475. Miami...

    2024

  9. [17]

    Huang, R.; and Wang, S. 2025. Steering LLMs' Behavior with Concept Activation Vectors. In The Fourth Blogpost Track at ICLR 2025

    2025

  10. [18]

    Im, S.; and Li, Y. 2025. A Unified Understanding and Evaluation of Steering Methods. arXiv preprint arXiv:2502.02716

    2025

  11. [19]

    Ji, Z.; Chen, D.; Ishii, E.; Cahyawijaya, S.; Bang, Y.; Wilie, B.; and Fung, P. 2024. Llm internal states reveal hallucination risk faced with a query. arXiv preprint arXiv:2407.03282

    2024 arXiv

  12. [20]

    Q.; Sablayrolles, A.; Mensch, A.; Bamford, C.; Chaplot, D

    Jiang, A. Q.; Sablayrolles, A.; Mensch, A.; Bamford, C.; Chaplot, D. S.; de las Casas, D.; Bressand, F.; Lengyel, G.; Lample, G.; Saulnier, L.; Lavaud, L. R.; Lachaux, M.-A.; Stock, P.; Scao, T. L.; Lavril, T.; Wang, T.; Lacroix, T.; and Sayed, W. E. 2023. Mistral 7B. arXiv:2310.06825

    2023 arXiv

  13. [21]

    Jin, M.; Yu, Q.; Huang, J.; Zeng, Q.; Wang, Z.; Hua, W.; Zhao, H.; Mei, K.; Meng, Y.; Ding, K.; Yang, F.; Du, M.; and Zhang, Y. 2025. Exploring Concept Depth: How Large Language Models Acquire Knowledge and Concept at Different Layers? In Proceedings of the 31st International ...

    2025

  14. [22]

    Ju, C.; Shi, W.; Liu, C.; Ji, J.; Zhang, J.; Zhang, R.; Zhu, J.; Xu, J.; Yang, Y.; Han, S.; et al. 2025. Benchmarking Multi-National Value Alignment for Large Language Models. arXiv preprint arXiv:2504.12911

    2025 arXiv

  15. [23]

    Kantamneni, S.; Engels, J.; Rajamanoharan, S.; Tegmark, M.; and Nanda, N. 2025. Are sparse autoencoders useful? a case study in sparse probing. arXiv preprint arXiv:2502.16681

    2025 arXiv

  16. [24]

    Kim, J.; Evans, J.; and Schein, A. 2025. Linear Representations of Political Perspective Emerge in Large Language Models. arXiv preprint arXiv:2503.02080

    2025 arXiv

  17. [25]

    Li, K.; Patel, O.; Vi \'e gas, F.; Pfister, H.; and Wattenberg, M. 2023. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36: 41451--41530

    2023

  18. [26]

    Marks, S.; and Tegmark, M. 2023. The geometry of truth: Emergent linear structure in large language model representations of true/false datasets. arXiv preprint arXiv:2310.06824

    2023 arXiv

  19. [27]

    Marks, S.; Treutlein, J.; Bricken, T.; Lindsey, J.; Marcus, J.; Mishra-Sharma, S.; Ziegler, D.; Ameisen, E.; Batson, J.; Belonax, T.; et al. 2025. Auditing language models for hidden objectives. arXiv preprint arXiv:2503.10965

    2025 arXiv

  20. [28]

    Motoki, F.; Pinho Neto, V.; and Rodrigues, V. 2024. More human than human: measuring ChatGPT political bias. Public Choice, 198(1): 3--23

    2024

  21. [29]

    nostalgebraist. 2020. Interpreting GPT: The Logit Lens

    2020

  22. [30]

    Ousidhoum, N.; Zhao, X.; Fang, T.; Song, Y.; and Yeung, D.-Y. 2021. Probing toxic content in large pre-trained language models. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Lan...

    2021

  23. [31]

    Panickssery, N.; Gabrieli, N.; Schulz, J.; Tong, M.; Hubinger, E.; and Turner, A. M. 2023. Steering llama 2 via contrastive activation addition. arXiv preprint arXiv:2312.06681

    2023 arXiv

  24. [32]

    J.; and Veitch, V

    Park, K.; Choe, Y. J.; and Veitch, V. 2023. The linear representation hypothesis and the geometry of large language models. arXiv preprint arXiv:2311.03658

    2023 arXiv

  25. [33]

    Paschalides, D.; Pallis, G.; and Dikaiakos, M. D. 2025. Probing the Subtle Ideological Manipulation of Large Language Models. arXiv preprint arXiv:2504.14287

    2025 arXiv

  26. [34]

    P.; Li, Y.; and Evans, J

    Piao, J.; Lu, Z.; Gao, C.; Xu, F.; Santos, F. P.; Li, Y.; and Evans, J. 2025. Emergence of human-like polarization among large language model agents. arXiv preprint arXiv:2501.05171

    2025 arXiv

  27. [35]

    Potter, Y.; Lai, S.; Kim, J.; Evans, J.; and Song, D. 2024. Hidden Persuaders: LLMs' Political Leaning and Their Influence on Voters. arXiv preprint arXiv:2410.24190

    2024 arXiv

  28. [36]

    S.; and Krueger, D

    Pres, I.; Ruis, L.; Lubana, E. S.; and Krueger, D. 2024. Towards Reliable Evaluation of Behavior Steering Interventions in LLMs. arXiv preprint arXiv:2410.17245

    2024 arXiv

  29. [37]

    o ttger, P.; Hofmann, V.; Pyatkin, V.; Hinck, M.; Kirk, H. R.; Sch \

    R \"o ttger, P.; Hofmann, V.; Pyatkin, V.; Hinck, M.; Kirk, H. R.; Sch \"u tze, H.; and Hovy, D. 2024. Political compass or spinning arrow? towards more meaningful evaluations for values and opinions in large language models. arXiv preprint arXiv:2402.16786

    2024 arXiv

  30. [38]

    Rozado, D. 2024. The political preferences of LLMs. PloS one, 19(7): e0306621

    2024

  31. [39]

    Subramani, N.; Suresh, N.; and Peters, M. E. 2022. Extracting latent steering vectors from pretrained language models. arXiv preprint arXiv:2205.05124

    2022 arXiv

  32. [40]

    N.; Banayeeanzade, A.; Bolourani, A.; Kian, M.; Jia, R.; and Gratch, J

    Tak, A. N.; Banayeeanzade, A.; Bolourani, A.; Kian, M.; Jia, R.; and Gratch, J. 2025. Mechanistic Interpretability of Emotion Inference in Large Language Models. arXiv preprint arXiv:2502.05489

    2025 arXiv

  33. [41]

    Tan, B. C. Z.; and Lee, R. K.-W. 2025. Unmasking Implicit Bias: Evaluating Persona-Prompted LLM Responses in Power-Disparate Social Scenarios. arXiv preprint arXiv:2503.01532

    2025 arXiv

  34. [42]

    Team, G.; Kamath, A.; Ferret, J.; Pathak, S.; Vieillard, N.; Merhej, R.; Perrin, S.; Matejovicova, T.; Ram \'e , A.; Rivi \`e re, M.; et al. 2025. Gemma 3 technical report. arXiv preprint arXiv:2503.19786

    2025 arXiv

  35. [43]

    M.; Thiergart, L.; Leech, G.; Udell, D.; Vazquez, J

    Turner, A. M.; Thiergart, L.; Leech, G.; Udell, D.; Vazquez, J. J.; Mini, U.; and MacDiarmid, M. 2023 a . Activation addition: Steering language models without optimization. arXiv, arXiv:2308

    2023

  36. [44]

    M.; Thiergart, L.; Leech, G.; Udell, D.; Vazquez, J

    Turner, A. M.; Thiergart, L.; Leech, G.; Udell, D.; Vazquez, J. J.; Mini, U.; and MacDiarmid, M. 2023 b . Steering language models with activation engineering. arXiv preprint arXiv:2308.10248

    2023 arXiv

  37. [45]

    Vogel, T. 2024. repeng

    2024

  38. [46]

    Wan, Y.; Pu, G.; Sun, J.; Garimella, A.; Chang, K.-W.; and Peng, N. 2023. Kelly is a Warm Person, Joseph is a Role Model : Gender Biases in LLM -Generated Reference Letters. In Bouamor, H.; Pino, J.; and Bali, K., eds., Findings of the Association for Computational Linguistics...

    2023

  39. [47]

    D.; Cislak, A.; and Schmidt, P

    Wojcik, A. D.; Cislak, A.; and Schmidt, P. 2021. ‘The left is right’: Left and right political orientation across Eastern and Western Europe. The Social Science Journal, 1--17

    2021

  40. [48]

    D.; and Potts, C

    Wu, Z.; Arora, A.; Geiger, A.; Wang, Z.; Huang, J.; Jurafsky, D.; Manning, C. D.; and Potts, C. 2025. AxBench: Steering LLMs? Even Simple Baselines Outperform Sparse Autoencoders. arXiv:2501.17148

    2025 arXiv

  41. [49]

    D.; Manning, C

    Wu, Z.; Geiger, A.; Arora, A.; Huang, J.; Wang, Z.; Goodman, N. D.; Manning, C. D.; and Potts, C. 2024. pyvene: A library for understanding and improving pytorch models via interventions. arXiv preprint arXiv:2403.07809

    2024 arXiv

  42. [50]

    Xu, Z.; Huang, R.; Chen, C.; and Wang, X. 2024. Uncovering safety risks of large language models through concept activation vector. Advances in Neural Information Processing Systems, 37: 116743--116782

    2024

  43. [51]

    Yang, A.; Li, A.; Yang, B.; Zhang, B.; Hui, B.; Zheng, B.; Yu, B.; Gao, C.; Huang, C.; Lv, C.; Zheng, C.; Liu, D.; Zhou, F.; Huang, F.; Hu, F.; Ge, H.; Wei, H.; Lin, H.; Tang, J.; Yang, J.; Tu, J.; Zhang, J.; Yang, J.; Yang, J.; Zhou, J.; Zhou, J.; Lin, J.; Dang, K.; Bao, K.; ...

    2025 arXiv

  44. [52]

    Zhao, H.; Zhao, H.; Shen, B.; Payani, A.; Yang, F.; and Du, M. 2024. Beyond single concept vector: Modeling concept subspace in llms with gaussian distribution. arXiv preprint arXiv:2410.00153

    2024 arXiv

  45. [53]

    Zou, A.; Phan, L.; Chen, S.; Campbell, J.; Guo, P.; Ren, R.; Pan, A.; Yin, X.; Mazeika, M.; Dombrowski, A.-K.; et al. 2023. Representation engineering: A top-down approach to ai transparency. arXiv preprint arXiv:2310.01405

    2023 arXiv

  46. [54]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  47. [55]

    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.