pith. machine review for the scientific record. sign in

arxiv: 2511.17408 · v4 · submitted 2025-11-21 · 💻 cs.AI · cs.LG

The Impact of Off-Policy Training Data on Probe Generalisation

Nathalie Kirch , Samuel Dower , Adrians Skapars , Helen Yannakoudakis , Ekdeep Singh Lubana , Dmitrii Krasheninnikov This is my paper

Pith reviewed 2026-05-17 20:22 UTC · model grok-4.3

classification 💻 cs.AI cs.LG
keywords probe generalisationoff-policy dataLLM monitoringdeception detectiondistribution shiftAI safetylinear probesattention probes
0
0 comments X

The pith

Off-policy training data causes the largest generalization failures in probes for intent-based LLM behaviors like strategic deception.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

This paper tests how training probes on synthetic or off-policy LLM responses affects their ability to detect eight different behaviors when applied to new models and settings. Probe performance varies with the data generation strategy, but the drop is much steeper for behaviors defined by response intent than for those defined by surface text features. The authors identify a practical check: probes that generalize well to incentivised or coerced responses also tend to perform well on on-policy examples, allowing prediction of failures even without direct access to rare real-world test data. The work concludes that current deception probes are likely to underperform in actual monitoring deployments. It also notes that off-policy data can sometimes produce more reliable probes than on-policy data drawn from a mismatched distribution.

Core claim

Training data generation strategy significantly affects probe performance across behaviors, with the magnitude varying greatly by behavior. The largest generalization failures arise for behaviors defined by response intent, such as strategic deception, rather than text-level content such as usage of lists. Successful generalization to incentivised data strongly correlates with high performance against on-policy examples and can therefore serve as a useful test for predicting generalization failures when on-policy test data is unavailable. Off-policy data can yield more reliable probes than on-policy data from a sufficiently different setting. These results lead to the prediction that current

What carries the argument

Linear and attention probes trained on mixtures of on-policy, off-policy, and incentivised responses, evaluated for generalization across eight LLM behaviors with emphasis on intent versus content distinctions.

If this is right

  • Generalization failures are substantially larger for intent-defined behaviors such as strategic deception than for text-content behaviors such as list usage.
  • Generalization success on incentivised or coerced data correlates strongly with performance on on-policy examples and can predict failures without direct on-policy test data.
  • Off-policy data from one setting can produce more reliable probes for another setting than on-policy data from a mismatched distribution.
  • Current deception probes are predicted to fail to generalize to real monitoring scenarios.
  • Better monitoring methods are needed that can handle all types of distribution shift.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The proposed correlation test with incentivised data could serve as a practical proxy for validating probes when collecting rare on-policy examples is difficult.
  • Similar generalization gaps may appear in other detection tasks that rely on synthetic data for training safety monitors.
  • The distinction between intent-based and content-based behaviors could guide the design of more robust monitoring systems that explicitly account for how the target behavior is defined.

Load-bearing premise

The observed correlation between successful generalization to incentivised data and performance on on-policy examples will reliably predict generalization failures in actual deployment settings across different models and behaviors.

What would settle it

Measuring whether a probe that generalizes successfully to incentivised deceptive responses also accurately detects strategic deception in genuine on-policy model outputs collected during real deployment or user interactions.

Figures

Figures reproduced from arXiv: 2511.17408 by Adrians Skapars, Dmitrii Krasheninnikov, Ekdeep Singh Lubana, Helen Yannakoudakis, Nathalie Kirch, Samuel Dower.

Figure 1
Figure 1. Figure 1: Overview of our methodology, including the four response strategies for collecting data for probe training. We evaluate each probe on natural data to determine which strategy is best. a practical heuristic: when natural exam￾ples are unavailable, using off-policy data from the target monitoring setting is more effective than using on-policy data from an unrelated setting (e.g. writing essays vs. an￾swering… view at source ↗
Figure 2
Figure 2. Figure 2: Probes generalise well for some behaviours (text-evident) but not for others (text￾ambiguous). We report test AUROC scores for linear probes, decomposed by behaviour, for all behaviours except deception and sandbagging. We evaluate probes trained on either the same (top) or different (bottom) domain as the test set data, with activations taken from Llama-3.2-3B-Instruct. Deception and sandbagging behaviour… view at source ↗
Figure 3
Figure 3. Figure 3: Shifts in the training data domain present a larger challenge to probe gener￾alisation than the choice of response strat￾egy. This figure presents the same data as [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figure 5
Figure 5. Figure 5: Deception and sandbagging probes are predicted to not generalise across re￾sponse strategy shifts. We report test AUROC scores for linear probes, for the deception and sand￾bagging behaviours. We evaluate linear probes trained on same domain as the test set data, with activations taken from Llama-3.2-3B-Instruct [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 4
Figure 4. Figure 4: Generalisation to on-policy incen￾tivised data is a good proxy for generalisation to on-policy natural data. Y-axis: the differ￾ence in test AUROC scores between the on-policy natural and off-policy probes when testing on on￾policy natural data. X-axis: the difference in test AUROC scores between the on-policy incentivised and off-policy probes when testing on on-policy in￾centivised data. Results are give… view at source ↗
Figure 6
Figure 6. Figure 6: We report test AUROC scores for linear probes, averaged across all behaviours except [PITH_FULL_IMAGE:figures/full_fig_p022_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: We report how test AUROC scores change as we increase the amount of probe training data, [PITH_FULL_IMAGE:figures/full_fig_p022_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: We report how test AUROC scores change as we add [PITH_FULL_IMAGE:figures/full_fig_p023_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Sycophancy linear probe AUROC scores for Llama-3.2-3B-Instruct. We always test on multiple [PITH_FULL_IMAGE:figures/full_fig_p023_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Deferral-to-authority linear probe AUROC scores for Llama-3.2-3B-Instruct. We test on multiple choice questions (MUL), and train on different combinations of MUL, argument (ARG) and haiku (HAI) domains. We see that for sycophancy, combining the two training domains does result in some improve￾ment, especially for on-policy natural and off-policy. However, this improvement is not significant and still does… view at source ↗
Figure 11
Figure 11. Figure 11: We report test AUROC scores for linear probes, decomposed by behaviour. We evaluate probes [PITH_FULL_IMAGE:figures/full_fig_p026_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: We report test AUROC scores for linear probes, decomposed by behaviour. We evaluate probes [PITH_FULL_IMAGE:figures/full_fig_p026_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: We report test AUROC scores for attention probes, decomposed by behaviour. We evaluate [PITH_FULL_IMAGE:figures/full_fig_p027_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: We report test AUROC scores for attention probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p027_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: We report the test AUROC scores for linear probes, across all behaviours except for [PITH_FULL_IMAGE:figures/full_fig_p028_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: We report test AUROC scores for attention probes, decomposed by behaviour. We evaluate [PITH_FULL_IMAGE:figures/full_fig_p028_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: We report test AUROC scores for linear probes, decomposed by behaviour. We evaluate probes [PITH_FULL_IMAGE:figures/full_fig_p029_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: We report test AUROC scores for linear probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p029_18.png] view at source ↗
Figure 19
Figure 19. Figure 19: We report the test AUROC scores for linear probes, across all behaviours except for [PITH_FULL_IMAGE:figures/full_fig_p030_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: We report test AUROC scores for linear probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p030_20.png] view at source ↗
Figure 21
Figure 21. Figure 21: We report test AUROC scores for attention probes, decomposed by behaviour. We evaluate [PITH_FULL_IMAGE:figures/full_fig_p031_21.png] view at source ↗
Figure 22
Figure 22. Figure 22: We report test AUROC scores for attention probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p031_22.png] view at source ↗
Figure 23
Figure 23. Figure 23: We report the test AUROC scores for attention probes, across all behaviours except for [PITH_FULL_IMAGE:figures/full_fig_p032_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: We report test AUROC scores for attention probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p032_24.png] view at source ↗
Figure 25
Figure 25. Figure 25: We report test AUROC scores for linear probes, decomposed by behaviour. We evaluate probes [PITH_FULL_IMAGE:figures/full_fig_p033_25.png] view at source ↗
Figure 26
Figure 26. Figure 26: We report test AUROC scores for linear probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p033_26.png] view at source ↗
Figure 27
Figure 27. Figure 27: We report the test AUROC scores for linear probes, across all behaviours except for [PITH_FULL_IMAGE:figures/full_fig_p034_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: We report test AUROC scores for linear probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p034_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: We report test AUROC scores for attention probes, decomposed by behaviour. We evaluate [PITH_FULL_IMAGE:figures/full_fig_p035_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: We report test AUROC scores for attention probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p035_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: We report the test AUROC scores for attention probes, across all behaviours except for [PITH_FULL_IMAGE:figures/full_fig_p036_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: We report test AUROC scores for attention probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p036_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: We report test AUROC scores for linear probes, decomposed by behaviour. We evaluate probes [PITH_FULL_IMAGE:figures/full_fig_p037_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: We report test AUROC scores for linear probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p037_34.png] view at source ↗
Figure 35
Figure 35. Figure 35: We report test AUROC scores for linear probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p038_35.png] view at source ↗
Figure 36
Figure 36. Figure 36: We report test AUROC scores for attention probes, decomposed by behaviour. We evaluate [PITH_FULL_IMAGE:figures/full_fig_p038_36.png] view at source ↗
Figure 37
Figure 37. Figure 37: We report test AUROC scores for attention probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p039_37.png] view at source ↗
Figure 38
Figure 38. Figure 38: We report test AUROC scores for attention probes, decomposed by behaviour, for all behaviours [PITH_FULL_IMAGE:figures/full_fig_p039_38.png] view at source ↗
Figure 39
Figure 39. Figure 39: When evaluated on the on-policy test sets, domain shift has a greater impact on probe performance than response strategy. Probes trained on a different domain showed significantly worse performance (coeff = -0.162, p<0.001). In contrast, response strategy had no significant effect on probe performance (all p>0.05) [PITH_FULL_IMAGE:figures/full_fig_p041_39.png] view at source ↗
Figure 40
Figure 40. Figure 40: When evaluated on the incentivised test sets, domain shift has a greater impact on probe performance than response strategy. Probes trained on a different domain showed significantly worse performance (coeff = -0.180, p<0.001). In contrast, response strategy had a small and non-significant effect on probe performance (coeff = 0.023, p>0.05). Interestingly, the interaction between prompted response strateg… view at source ↗
Figure 41
Figure 41. Figure 41: Domain shift has a greater impact on probe performance than response generation method, but the magnitude of this impact varies substantially across behaviours. A mixed-effects regression model reveals that probes trained on a different domain showed significantly worse performance overall (coeff = -0.272, p<0.001), with a baseline intercept of 0.916 (p<0.001). However, the domain shift effect was partial… view at source ↗
Figure 42
Figure 42. Figure 42: Domain shift has a greater impact on probe performance than response strategy, but the magnitude of this impact varies substantially across behaviours. Probes trained on a different domain showed a modest overall performance drop (coeff = -0.074, p=0.048). However, this effect varied dramatically across behaviours through significant interaction effects. Three behaviours showed substantial negative intera… view at source ↗
Figure 43
Figure 43. Figure 43: We report the decrease in test AUROC scores relative to the test AUROC scores from using [PITH_FULL_IMAGE:figures/full_fig_p049_43.png] view at source ↗
Figure 44
Figure 44. Figure 44: We report test AUROC scores for linear probes, for the [PITH_FULL_IMAGE:figures/full_fig_p049_44.png] view at source ↗
Figure 45
Figure 45. Figure 45: We report test AUROC scores for linear probes, for the [PITH_FULL_IMAGE:figures/full_fig_p050_45.png] view at source ↗
Figure 46
Figure 46. Figure 46: We report test AUROC scores for linear probes, for the [PITH_FULL_IMAGE:figures/full_fig_p050_46.png] view at source ↗
Figure 47
Figure 47. Figure 47: We report test AUROC scores for linear probes, for the [PITH_FULL_IMAGE:figures/full_fig_p050_47.png] view at source ↗
Figure 48
Figure 48. Figure 48: We report test AUROC scores for attention probes, for the [PITH_FULL_IMAGE:figures/full_fig_p051_48.png] view at source ↗
read the original abstract

Probing has emerged as a promising method for monitoring large language models (LLMs), enabling cheap inference-time detection of concerning behaviours. However, natural examples of many behaviours are rare, forcing researchers to rely on synthetic or off-policy LLM responses for training probes. We systematically evaluate how off-policy data influences probe generalisation across eight distinct LLM behaviours. Testing linear and attention probes across multiple LLMs, we find that training data generation strategy can significantly affect probe performance, though the magnitude varies greatly by behaviour. The largest generalisation failures arise for behaviours defined by response ``intent'' (e.g., strategic deception) rather than text-level content (e.g., usage of lists). We then propose a useful test for predicting generalisation failures in cases where on-policy test data is unavailable: successful generalisation to incentivised data (where the model was coerced) strongly correlates with high performance against on-policy examples. Based on these results, we predict that current deception probes may fail to generalise to real monitoring scenarios. We find that off-policy data can yield more reliable probes than on-policy data from a sufficiently different setting. This underscores the need for better monitoring methods that handle all types of distribution shift.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 2 minor

Summary. The manuscript presents a systematic empirical evaluation of how off-policy training data generation strategies affect the generalization performance of linear and attention probes for detecting eight LLM behaviors. Key results show that data strategy impacts probe performance with substantial variation across behaviors, that the largest generalization failures occur for intent-based behaviors such as strategic deception rather than content-based ones such as list usage, and that generalization success on incentivised/coerced data correlates with on-policy performance. This correlation is used to predict that current deception probes are likely to fail in real monitoring scenarios, while also noting that off-policy data can sometimes produce more reliable probes than on-policy data from sufficiently different settings.

Significance. If the central empirical findings hold, the work provides a useful practical test for anticipating probe generalization failures in settings where on-policy test data is unavailable. The distinction between intent-based and content-based behaviors as a predictor of generalization difficulty is a substantive contribution to LLM monitoring research. The evaluation across multiple models and behaviors offers concrete evidence that off-policy data choices matter, which is directly relevant to AI safety applications. The manuscript earns credit for its systematic design and for identifying a correlation-based heuristic that could guide future probe development.

major comments (2)
  1. [Section on proposed test / correlation analysis] Section describing the proposed generalization test: the reported correlation between probe success on incentivised/coerced synthetic data and on-policy examples is measured exclusively inside the paper's controlled generation pipeline (specific prompt templates, coercion instructions, and the eight hand-chosen behaviors). Because real deployment encounters natural shifts such as topic drift, multi-turn context, and implicit rather than explicit coercion, the correlation may be an artifact of the shared synthetic process rather than a general property; this directly affects the load-bearing prediction that deception probes will fail to generalise to real monitoring scenarios.
  2. [Results / behavior-specific analysis] Results section on behavior-specific generalization: the claim that intent-defined behaviors exhibit the largest generalization failures is central to the paper's conclusions, yet the quantitative support (performance deltas, confidence intervals, or statistical tests comparing strategic deception to list-usage behavior) is not presented with sufficient detail to confirm the distinction is robust across models and probe types.
minor comments (2)
  1. [Methods] Methods section: additional detail on the precise prompt templates, number of samples per condition, and any post-hoc filtering criteria would improve reproducibility of the off-policy data generation process.
  2. [Figures] Figure captions: several figures comparing probe performance across behaviors would benefit from explicit labeling of which curves correspond to linear versus attention probes and which data-generation conditions are shown.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their detailed and constructive feedback, which has helped us identify areas where the manuscript can be strengthened. We address each major comment below, indicating where revisions will be made to improve clarity, robustness, and the framing of our claims. Our responses focus on substantive points without altering the core empirical findings.

read point-by-point responses

  1. Referee: Section describing the proposed generalization test: the reported correlation between probe success on incentivised/coerced synthetic data and on-policy examples is measured exclusively inside the paper's controlled generation pipeline (specific prompt templates, coercion instructions, and the eight hand-chosen behaviors). Because real deployment encounters natural shifts such as topic drift, multi-turn context, and implicit rather than explicit coercion, the correlation may be an artifact of the shared synthetic process rather than a general property; this directly affects the load-bearing prediction that deception probes will fail to generalise to real monitoring scenarios.

    Authors: We agree that the observed correlation is derived from our controlled synthetic generation pipeline and does not directly test against all forms of real-world distribution shift, such as topic drift or multi-turn implicit coercion. This is a genuine limitation of the current evidence. We will revise the relevant section to frame the proposed test explicitly as a practical heuristic for anticipating generalization failures in settings that share similar synthetic characteristics, rather than claiming it as a fully general predictor. We will also expand the discussion to acknowledge that real monitoring scenarios may introduce additional shifts not captured here and suggest this as an important direction for future work. These changes will qualify the prediction regarding deception probes without overstating the current results. revision: yes

  2. Referee: Results section on behavior-specific generalization: the claim that intent-defined behaviors exhibit the largest generalization failures is central to the paper's conclusions, yet the quantitative support (performance deltas, confidence intervals, or statistical tests comparing strategic deception to list-usage behavior) is not presented with sufficient detail to confirm the distinction is robust across models and probe types.

    Authors: We acknowledge that the manuscript would benefit from more detailed quantitative support for the distinction between intent-based and content-based behaviors. In the revised version, we will add tables or figures reporting performance deltas with confidence intervals across models and probe types, along with appropriate statistical comparisons (e.g., paired tests or effect size measures) between representative behaviors such as strategic deception and list usage. This will allow readers to assess the robustness of the pattern directly. The underlying data and analysis code will also be made available to support verification. revision: yes

Circularity Check

0 steps flagged

No circularity: purely empirical study with independent experimental results

full rationale

This paper conducts a systematic empirical evaluation of how off-policy training data affects linear and attention probe generalization across eight LLM behaviors. All claims rest on direct experimental comparisons of probe performance under different data generation strategies (on-policy, off-policy, incentivised/coerced), with correlations reported as observed patterns rather than derived from any equations or self-referential definitions. No mathematical derivations, fitted parameters renamed as predictions, or load-bearing self-citations appear in the provided text or abstract. The proposed test (correlation between incentivised-data generalization and on-policy performance) is presented as an empirical observation that can be independently verified or falsified on new models and behaviors, keeping the central findings self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 2 axioms · 0 invented entities

The work relies on standard assumptions from the LLM probing literature and introduces an empirical correlation as a predictive tool.

axioms (2)
  • domain assumption Probes can be trained on synthetic or off-policy responses to detect LLM behaviors
    Core premise of the probing setup described in the abstract.
  • ad hoc to paper Generalization to incentivised data correlates with on-policy performance
    The proposed test is presented as a useful predictor based on observed results.

pith-pipeline@v0.9.0 · 5529 in / 1389 out tokens · 39088 ms · 2026-05-17T20:22:06.521255+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 45 canonical work pages · 17 internal anchors

  1. [1]

    Linear Control of Test Awareness Reveals Differential Compliance in Reasoning Models , May 2025

    Sahar Abdelnabi and Ahmed Salem. Linear Control of Test Awareness Reveals Differential Compliance in Reasoning Models , May 2025. URL http://arxiv.org/abs/2505.14617. arXiv:2505.14617 [cs]

    work page arXiv 2025

  2. [2]

    Invariant Risk Minimization

    Martin Arjovsky, Léon Bottou, Ishaan Gulrajani, and David Lopez-Paz. Invariant risk minimization, 2020. URL http://arxiv.org/abs/1907.02893

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  3. [3]

    Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback

    Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, Nicholas Joseph, Saurav Kadavath, Jackson Kernion, Tom Conerly, Sheer El-Showk, Nelson Elhage, Zac Hatfield-Dodds, Danny Hernandez, Tristan Hume, Scott Johnston, Shauna Kravec, Liane Lovitt, Neel Nanda, Catherine Olsson, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2022

  4. [4]

    Deception in LLMs : Self - Preservation and Autonomous Goals in Large Language Models , January 2025

    Sudarshan Kamath Barkur, Sigurd Schacht, and Johannes Scholl. Deception in LLMs : Self - Preservation and Autonomous Goals in Large Language Models , January 2025. URL http://arxiv.org/abs/2501.16513. arXiv:2501.16513 [cs]

    work page arXiv 2025

  5. [5]

    Bowman, and David Duvenaud

    Joe Benton, Misha Wagner, Eric Christiansen, Cem Anil, Ethan Perez, Jai Srivastav, Esin Durmus, Deep Ganguli, Shauna Kravec, Buck Shlegeris, Jared Kaplan, Holden Karnofsky, Evan Hubinger, Roger Grosse, Samuel R. Bowman, and David Duvenaud. Sabotage evaluations for frontier models. URL http://arxiv.org/abs/2410.21514

    work page arXiv

  6. [6]

    Chalmers

    David J. Chalmers. Propositional Interpretability in Artificial Intelligence , January 2025. URL http://arxiv.org/abs/2501.15740. arXiv:2501.15740 [cs]

    work page arXiv 2025

  7. [7]

    Yik Siu Chan, Zheng-Xin Yong, and Stephen H. Bach. Can We Predict Alignment Before Models Finish Thinking ? Towards Monitoring Misaligned Reasoning Models , July 2025. URL http://arxiv.org/abs/2507.12428. arXiv:2507.12428 [cs]

    work page arXiv 2025

  8. [8]

    Cost-effective constitutional classifiers via representation re-use, 2025

    Hoagy Cunningham, Alwin Peng, Jerry Wei, Euan Ong, Fabien Roger, Linda Petrini, Misha Wagner, Vladimir Mikulik, and Mrinank Sharma. Cost-effective constitutional classifiers via representation re-use, 2025. URL https://alignment.anthropic.com/2025/cheap-monitors/

    work page 2025

  9. [9]

    DeepSeek-V3 Technical Report

    DeepSeek - AI , Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, Damai Dai, Daya Guo, Dejian Yang, Deli Chen, Dongjie Ji, Erhang Li, Fangyun Lin, Fucong Dai, Fuli Luo, Guangbo Hao, Guanting Chen, Guowei Li, H. Zhang, Han Bao, Hanwei Xu, Haocheng Wang, Haowei Zhang, Honghui Ding, H...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  10. [10]

    Enhancing Chat Language Models by Scaling High-quality Instructional Conversations

    Ning Ding, Yulin Chen, Bokai Xu, Yujia Qin, Zhi Zheng, Shengding Hu, Zhiyuan Liu, Maosong Sun, and Bowen Zhou. Enhancing Chat Language Models by Scaling High -quality Instructional Conversations , May 2023. URL http://arxiv.org/abs/2305.14233. arXiv:2305.14233 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  11. [11]

    Hierarchical Neural Story Generation

    Angela Fan, Mike Lewis, and Yann Dauphin. Hierarchical Neural Story Generation , May 2018. URL http://arxiv.org/abs/1805.04833. arXiv:1805.04833 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  12. [12]

    Monitoring Latent World States in Language Models with Propositional Probes , December 2024

    Jiahai Feng, Stuart Russell, and Jacob Steinhardt. Monitoring Latent World States in Language Models with Propositional Probes , December 2024. URL http://arxiv.org/abs/2406.19501. arXiv:2406.19501 [cs]

    work page arXiv 2024

  13. [13]

    Detecting strategic deception using linear probes

    Nicholas Goldowsky-Dill, Bilal Chughtai, Stefan Heimersheim, and Marius Hobbhahn. Detecting Strategic Deception Using Linear Probes , February 2025. URL http://arxiv.org/abs/2502.03407. arXiv:2502.03407 [cs]

    work page arXiv 2025

  14. [14]

    Probing the Robustness of Large Language Models Safety to Latent Perturbations , June 2025

    Tianle Gu, Kexin Huang, Zongqi Wang, Yixu Wang, Jie Li, Yuanqi Yao, Yang Yao, Yujiu Yang, Yan Teng, and Yingchun Wang. Probing the Robustness of Large Language Models Safety to Latent Perturbations , June 2025. URL http://arxiv.org/abs/2506.16078. arXiv:2506.16078 [cs]

    work page arXiv 2025

  15. [15]

    What makes a convincing argument? Empirical analysis and detecting attributes of convincingness in Web argumentation

    Ivan Habernal and Iryna Gurevych. What makes a convincing argument? Empirical analysis and detecting attributes of convincingness in Web argumentation. In Proceedings of the 2016 Conference on Empirical Methods in Natural Language Processing , pages 1214--1223, Austin, Texas, 2016. Association for Computational Linguistics. URL https://aclweb.org/antholog...

    work page 2016

  16. [16]

    Measuring Massive Multitask Language Understanding

    Dan Hendrycks, Collin Burns, Steven Basart, Andy Zou, Mantas Mazeika, Dawn Song, and Jacob Steinhardt. Measuring Massive Multitask Language Understanding , January 2021. URL http://arxiv.org/abs/2009.03300. arXiv:2009.03300 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2021

  17. [17]

    Ministral-8b-instruct-2410 model card

    Albert Jiang, Alexandre Abou Chahine, Alexandre Sablayrolles, Alexis Tacnet, Alodie Boissonnet, Alok Kothari, Amélie Héliou, Andy Lo, Anna Peronnin, Antoine Meunier, Antoine Roux, Antonin Faure, Aritra Paul, Arthur Darcet, Arthur Mensch, Audrey Herblin-Stoop, Augustin Garreau, Austin Birky, Avinash Sooriyarachchi, Baptiste Rozière, Barry Conklin, Bastien ...

    work page 2024

  18. [18]

    Albert Q. Jiang, Alexandre Sablayrolles, Arthur Mensch, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Florian Bressand, Gianna Lengyel, Guillaume Lample, Lucile Saulnier, Lélio Renard Lavaud, Marie-Anne Lachaux, Pierre Stock, Teven Le Scao, Thibaut Lavril, Thomas Wang, Timothée Lacroix, and William El Sayed. Mistral 7b, 2023. URL http://arxiv...

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  19. [19]

    Albert Q. Jiang, Alexandre Sablayrolles, Antoine Roux, Arthur Mensch, Blanche Savary, Chris Bamford, Devendra Singh Chaplot, Diego de las Casas, Emma Bou Hanna, Florian Bressand, Gianna Lengyel, Guillaume Bour, Guillaume Lample, Lélio Renard Lavaud, Lucile Saulnier, Marie-Anne Lachaux, Pierre Stock, Sandeep Subramanian, Sophia Yang, Szymon Antoniak, Teven...

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  20. [20]

    Wildteaming at scale: From in-the-wild jailbreaks to (adversarially) safer language models, 2024a

    Liwei Jiang, Kavel Rao, Seungju Han, Allyson Ettinger, Faeze Brahman, Sachin Kumar, Niloofar Mireshghallah, Ximing Lu, Maarten Sap, Yejin Choi, and Nouha Dziri. WildTeaming at Scale : From In -the- Wild Jailbreaks to ( Adversarially ) Safer Language Models , June 2024 c . URL http://arxiv.org/abs/2406.18510. arXiv:2406.18510 [cs]

    work page arXiv 2024

  21. [21]

    HiddenDetect : Detecting Jailbreak Attacks against Large Vision - Language Models via Monitoring Hidden States , June 2025

    Yilei Jiang, Xinyan Gao, Tianshuo Peng, Yingshui Tan, Xiaoyong Zhu, Bo Zheng, and Xiangyu Yue. HiddenDetect : Detecting Jailbreak Attacks against Large Vision - Language Models via Monitoring Hidden States , June 2025. URL http://arxiv.org/abs/2502.14744. arXiv:2502.14744 [cs]

    work page arXiv 2025

  22. [22]

    What Features in Prompts Jailbreak LLMs ? Investigating the Mechanisms Behind Attacks , May 2025

    Nathalie Kirch, Constantin Weisser, Severin Field, Helen Yannakoudakis, and Stephen Casper. What Features in Prompts Jailbreak LLMs ? Investigating the Mechanisms Behind Attacks , May 2025. URL http://arxiv.org/abs/2411.03343. arXiv:2411.03343 [cs]

    work page arXiv 2025

  23. [23]

    The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning

    Nathaniel Li, Alexander Pan, Anjali Gopal, Summer Yue, Daniel Berrios, Alice Gatti, Justin D. Li, Ann-Kathrin Dombrowski, Shashwat Goel, Long Phan, Gabriel Mukobi, Nathan Helm-Burger, Rassin Lababidi, Lennart Justen, Andrew B. Liu, Michael Chen, Isabelle Barrass, Oliver Zhang, Xiaoyuan Zhu, Rishub Tamirisa, Bhrugu Bharathi, Adam Khoja, Zhenqi Zhao, Ariel ...

    work page internal anchor Pith review Pith/arXiv arXiv

  24. [24]

    Simple probes can catch sleeper agents, April 2024

    Monte MacDiarmid, Timothy Maxwell, Nicholas Schiefer, Jesse Mu, Jared Kaplan, David Duvenaud, Sam Bowman, Alex Tamkin, Ethan Perez, Mrinank Sharma, Carson Denison, and Evan Hubinger. Simple probes can catch sleeper agents, April 2024. URL https://www.anthropic.com/news/probes-catch-sleeper-agents

    work page 2024

  25. [25]

    HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal

    Mantas Mazeika, Long Phan, Xuwang Yin, Andy Zou, Zifan Wang, Norman Mu, Elham Sakhaee, Nathaniel Li, Steven Basart, Bo Li, David Forsyth, and Dan Hendrycks. HarmBench : A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal , February 2024. URL http://arxiv.org/abs/2402.04249

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  26. [26]

    Detecting High - Stakes Interactions with Activation Probes , June 2025

    Alex McKenzie, Urja Pawar, Phil Blandfort, William Bankes, David Krueger, Ekdeep Singh Lubana, and Dmitrii Krasheninnikov. Detecting High - Stakes Interactions with Activation Probes , June 2025. URL http://arxiv.org/abs/2506.10805. arXiv:2506.10805 [cs] version: 1

    work page arXiv 2025

  27. [27]

    Llama 3.2: Revolutionizing edge ai and vision (connect 2024)

    Meta AI . Llama 3.2: Revolutionizing edge ai and vision (connect 2024). https://ai.meta.com/blog/llama-3-2-connect-2024-vision-edge-mobile-devices/, September 2024

    work page 2024

  28. [28]

    Probing evaluation awareness of language models.arXiv preprint arXiv:2507.01786,

    Jord Nguyen, Khiem Hoang, Carlo Leonardo Attubato, and Felix Hofstätter. Probing and Steering Evaluation Awareness of Language Models , July 2025. URL http://arxiv.org/abs/2507.01786. arXiv:2507.01786 [cs] version: 2

    work page arXiv 2025

  29. [29]

    Gpt-5 system card

    OpenAI . Gpt-5 system card. Technical report, OpenAI, August 13 2025. URL https://cdn.openai.com/gpt-5-system-card.pdf

    work page 2025

  30. [30]

    Benchmarking Deception Probes via Black -to- White Performance Boosts , August 2025

    Avi Parrack, Carlo Leonardo Attubato, and Stefan Heimersheim. Benchmarking Deception Probes via Black -to- White Performance Boosts , August 2025. URL http://arxiv.org/abs/2507.12691. arXiv:2507.12691 [cs]

    work page arXiv 2025

  31. [31]

    Qwen2.5 Technical Report

    Qwen, An Yang, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chengyuan Li, Dayiheng Liu, Fei Huang, Haoran Wei, Huan Lin, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jingren Zhou, Junyang Lin, Kai Dang, Keming Lu, Keqin Bao, Kexin Yang, Le Yu, Mei Li, Mingfeng Xue, Pei Zhang, Qin Zhu, Rui Men, Runji Lin, Tianhao Li, Ti...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    Coup probes: Catching catastrophes with probes trained off-policy, November 2023

    Fabien Roger. Coup probes: Catching catastrophes with probes trained off-policy, November 2023. URL https://www.lesswrong.com/posts/WCj7WgFSLmyKaMwPR/coup-probes-catching-catastrophes-with-probes-trained-off

    work page 2023

  33. [33]

    Large language models can strategically deceive their users when put under pressure.arXiv preprint arXiv:2311.07590, 2023

    Jérémy Scheurer, Mikita Balesni, and Marius Hobbhahn. Large Language Models can Strategically Deceive their Users when Put Under Pressure , July 2024. URL http://arxiv.org/abs/2311.07590. arXiv:2311.07590 [cs]

    work page arXiv 2024

  34. [34]

    Stress testing deliberative alignment for anti-scheming training, 2025

    Bronson Schoen, Evgenia Nitishinskaya, Mikita Balesni, Axel Højmark, Felix Hofstätter, Jérémy Scheurer, Alexander Meinke, Jason Wolfe, Teun van der Weij, Alex Lloyd, Nicholas Goldowsky-Dill, Angela Fan, Andrei Matveiakin, Rusheb Shah, Marcus Williams, Amelia Glaese, Boaz Barak, Wojciech Zaremba, and Marius Hobbhahn. Stress testing deliberative alignment f...

    work page arXiv 2025

  35. [35]

    Towards Understanding Sycophancy in Language Models

    Mrinank Sharma, Meg Tong, Tomasz Korbak, David Duvenaud, Amanda Askell, Samuel R. Bowman, Newton Cheng, Esin Durmus, Zac Hatfield-Dodds, Scott R. Johnston, Shauna Kravec, Timothy Maxwell, Sam McCandlish, Kamal Ndousse, Oliver Rausch, Nicholas Schiefer, Da Yan, Miranda Zhang, and Ethan Perez. Towards Understanding Sycophancy in Language Models , May 2025 a...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  36. [36]

    Mrinank Sharma, Meg Tong, Jesse Mu, Jerry Wei, Jorrit Kruthoff, Scott Goodfriend, Euan Ong, Alwin Peng, Raj Agarwal, Cem Anil, Amanda Askell, Nathan Bailey, Joe Benton, Emma Bluemke, Samuel R. Bowman, Eric Christiansen, Hoagy Cunningham, Andy Dau, Anjali Gopal, Rob Gilson, Logan Graham, Logan Howard, Nimit Kalra, Taesung Lee, Kevin Lin, Peter Lofgren, Fra...

    work page internal anchor Pith review arXiv 2025

  37. [37]

    Gemma Team, Aishwarya Kamath, Johan Ferret, Shreya Pathak, Nino Vieillard, Ramona Merhej, Sarah Perrin, Tatiana Matejovicova, Alexandre Ramé, Morgane Rivière, Louis Rouillard, Thomas Mesnard, Geoffrey Cideron, Jean bastien Grill, Sabela Ramos, Edouard Yvinec, Michelle Casbon, Etienne Pot, Ivo Penchev, Gaël Liu, Francesco Visin, Kathleen Kenealy, Lucas Bey...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  38. [38]

    Investigating task-specific prompts and sparse autoencoders for activation monitoring, April 2025

    Henk Tillman and Dan Mossing. Investigating task-specific prompts and sparse autoencoders for activation monitoring, April 2025. URL http://arxiv.org/abs/2504.20271. arXiv:2504.20271 [cs]

    work page arXiv 2025

  39. [39]

    Diffusion earth mover's distance and distribution embeddings, 2021

    Alexander Tong, Guillaume Huguet, Amine Natik, Kincaid MacDonald , Manik Kuchroo, Ronald Coifman, Guy Wolf, and Smita Krishnaswamy. Diffusion earth mover's distance and distribution embeddings, 2021. URL http://arxiv.org/abs/2102.12833

    work page arXiv 2021

  40. [40]

    Jailbroken: How Does LLM Safety Training Fail?

    Alexander Wei, Nika Haghtalab, and Jacob Steinhardt. Jailbroken: How Does LLM Safety Training Fail ?, July 2023. URL http://arxiv.org/abs/2307.02483. arXiv:2307.02483 [cs]

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  41. [41]

    Brown, and Francis Rhys Ward

    Teun van der Weij, Felix Hofstätter, Ollie Jaffe, Samuel F. Brown, and Francis Rhys Ward. AI Sandbagging : Language Models can Strategically Underperform on Evaluations , February 2025. URL http://arxiv.org/abs/2406.07358. arXiv:2406.07358 [cs]

    work page arXiv 2025

  42. [42]

    Bowman, He He, and Shi Feng

    Jiaxin Wen, Ruiqi Zhong, Akbir Khan, Ethan Perez, Jacob Steinhardt, Minlie Huang, Samuel R. Bowman, He He, and Shi Feng. Language Models Learn to Mislead Humans via RLHF , December 2024. URL http://arxiv.org/abs/2409.12822. arXiv:2409.12822 [cs]

    work page arXiv 2024

  43. [43]

    Qwen3 Technical Report

    An Yang, Anfeng Li, Baosong Yang, Beichen Zhang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Gao, Chengen Huang, Chenxu Lv, Chujie Zheng, Dayiheng Liu, Fan Zhou, Fei Huang, Feng Hu, Hao Ge, Haoran Wei, Huan Lin, Jialong Tang, Jian Yang, Jianhong Tu, Jianwei Zhang, Jianxin Yang, Jiaxi Yang, Jing Zhou, Jingren Zhou, Junyang Lin, Kai Dang, Keqin Bao, Kexin Yang, ...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  44. [44]

    Uncovering Latent Chain of Thought Vectors in Language Models , March 2025

    Jason Zhang and Scott Viteri. Uncovering Latent Chain of Thought Vectors in Language Models , March 2025. URL http://arxiv.org/abs/2409.14026. arXiv:2409.14026 [cs]

    work page arXiv 2025

  45. [45]

    Representation Engineering: A Top-Down Approach to AI Transparency

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, J. Zico Kolter, and Dan Hendrycks. Representation Engineering : A Top - Down Approach t...

    work page internal anchor Pith review Pith/arXiv arXiv 2025