Pith. sign in

REVIEW 4 major objections 5 minor 37 references

A Single Direction of Truth: An Observer Model's Linear Residual Probe Exposes and Steers Contextual Hallucinations

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A single linear direction inside a frozen observer model separates hallucinated from supported text and can steer generation.

desk verdict Useful observer-model probe and a valuable benchmark, but the steering and 27B transfer claims need controls before the causal story holds. read the letter →

arxiv 2507.23221 v1 pith:FHTOVCWB submitted 2025-07-31 cs.LG

Charles O'Neill , Slava Chalnev , Chi Chi Zhao , Max Kirkby , Mudith Jayasekara This is my paper
classification cs.LG
keywords contextualhallucinationlinearprobingresidualstreammechanisticinterpretabilitycausalsteeringdetectionobservermodelContraTales
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

This paper tries to establish that a separate, frozen language model can tell whether a candidate continuation contradicts its source context purely by reading one linear direction in its own internal activations. The observer needs only a single forward pass and a logistic probe on the residual stream at the final full stop, with no access to the generator's weights or logits. Across news summarisation and a new synthetic logical-contradiction benchmark, this probe beats lexical, entity, semantic, and attention-based baselines by 5-27 F1 points, and the direction transfers across Gemma-2 sizes and news domains. The paper also argues the direction is causally real: injecting it into a generator raises or lowers hallucination rates, and the signal localises to a sparse pattern of late-layer MLP activity. If right, this makes cheap, black-box-API hallucination auditing and targeted mitigation possible, and supports the broader claim that contextual consistency is encoded linearly inside transformers.

What carries the argument

The load-bearing object is the residual-stream linear probe: a logistic classifier with weight vector $w$ and bias $b$ applied to the post-layer-norm residual activation $r$ at the final token of the candidate continuation, trained by binary cross-entropy with $\ell^2$ regularisation. Its companion is the gradient-times-activation attribution, which scores how much each head's or MLP's output nudges the probe logit by projecting that output onto the gradient of the probe score with respect to the residual stream; averaged over the final sentence, it isolates where the hallucination signal lives. Together they turn a single forward pass into both a detector and a map of the sub-circuit that carries the signal.

What would settle it

Compile a corpus of naturally occurring, human-verified hallucinated continuations from diverse, independently deployed generators and run the frozen observer probe on it; if F1 falls to near the lexical-overlap baseline on logically subtle cases, the single-direction claim is an artifact of synthetic data.

Watch

Extended reading notes

Core claim

The central claim is that a transformer's residual stream carries a single, approximately linear axis that marks whether a sentence is supported by its preceding context, and that this axis can be read out by a different, frozen model. Concretely, a logistic regression on the post-layer-norm residual activation at the final token of the candidate sentence separates hallucinated from faithful continuations with F1 up to 0.99 on news summarisation and 0.84 on the synthetic ContraTales benchmark, outperforming four baselines by 5-27 points. The direction appears from roughly layers 8-12 onward, with a broad mid-layer plateau shared by Gemma-2 2B, 9B, and 27B observers; it transfers across CNN/DailyMail and XSum with little loss. Gradient-times-activation attribution concentrates the signal in a sparse, dataset-stable late-layer MLP pattern, and patching the normalised probe vector into a generating model bidirectionally modulates hallucination and repetition rates. The paper presents this as evidence for a low-dimensional, internally tracked, causally actionable representation of contextual inconsistency.

Load-bearing premise

The probe's training and evaluation data are synthetic continuations generated by instruction-following models, so the claimed direction may reflect artifacts of that generation pipeline rather than a general internal marker of contextual inconsistency.

Editorial extensions

If this is right

  • A single forward pass and a linear readout suffice to flag unsupported continuations, so hallucination detection becomes cheap enough to run as a post-hoc audit on black-box API outputs.
  • The discovered direction transfers across Gemma-2 model sizes and across news domains, implying contextual inconsistency is encoded along a shared, size-invariant axis.
  • Steering the direction in a generator moves hallucination and repetition rates in opposite directions, offering a concrete control knob for mitigation and a warning that the same knob could amplify fabrication.
  • Unsupervised finetuning on correct-only text improves probe F1 (0.75 to 0.89 for Gemma-2-9B on ContraTales), giving a label-free path to stronger detection on a target domain.
  • The signal localises to sparse, late-layer MLP activity rather than attention patterns, directing future mechanistic work to feed-forward sub-circuits.

Reading between the lines

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

  • If the direction is read at the final full stop, the same probe could be applied at every sentence boundary to flag the first unsupported sentence in a long generation, something the paper does not test.
  • The inverse hallucination-repetition trade-off suggests the axis may encode a broader faithfulness-versus-novelty continuum; testing the same steering on other failure modes such as sycophancy or style repetition would clarify what the direction actually controls.
  • Because the training data are synthetic continuations, the cleanest test of the paper's claims is application to organic, human-verified hallucinations from many different generators; the paper itself flags this as a limitation.
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 / 5 minor

Summary. The paper proposes an observer-based hallucination detector: a frozen LLM reads a source passage concatenated with a candidate continuation, and a logistic probe on the final-token residual-stream activation classifies the continuation as supported or hallucinated. The authors evaluate on CNN/DailyMail, XSum, and a new synthetic benchmark ContraTales, reporting F1 gains over lexical-overlap, entity-verification, semantic-similarity, and Lookback Lens baselines. They report cross-domain transfer between news datasets, a layer-wise performance plateau across Gemma-2 sizes, gradient-times-activation attributions localizing the signal to late-layer MLP activity, and a steering experiment in which adding or subtracting the probe direction modulates hallucination and repetition rates. They also present an unsupervised fine-tuning result that improves probe accuracy on ContraTales.

Significance. If the core detection claim holds, this is a practically useful contribution: a generator-agnostic, single-forward-pass detector with a transparent linear readout, plus a new benchmark for logical contradictions. The cross-domain transfer result and the sparse MLP attribution pattern are interesting and worth following up. The paper also deserves credit for releasing a benchmark and for explicitly acknowledging limitations around synthetic data and LLM-judge noise. However, the causal steering claim currently lacks essential null controls, the cross-model-size transferability claim is weakened by an unexplained '(sim.)' label on the Gemma-2-27B curve, and Appendix C describes a method that is internally inconsistent with the main text. The detection result is plausible but not fully established against standard entailment baselines or with paired significance testing.

major comments (4)
  1. [§4.4, Figure 5] The steering experiment does not include a null-direction control. Without injecting a random unit vector, a shuffled probe, or an equal-norm vector from a different layer, the observed monotone trade-off between hallucination and repetition is compatible with a generic off-distribution perturbation: large positive scaling can push activations out of distribution and produce text that a judge labels as hallucinated, while large negative scaling can suppress diversity and increase repetition. In addition, the hallucination rate is judged by GPT-4.1, the same model family used to create the training continuations, so a shared style bias cannot be excluded. The paper's own Limitations section acknowledges LLM-judge noise but does not supply the controls needed to separate a learned 'truth direction' from a generic perturbation. Please add at least: (i) injection of a random unit vector and a same-norm vector from another layer; (ii) an independent judge or a human-verified subset; (iii) per-alpha confidence intervals and judge-agreement statistics.
  2. [Figure 1, §4.1, abstract] The Gemma-2-27B curve is labeled '(sim.)' with no explanation in the caption or the text. If the 27B result is simulated, the abstract's claim of 'robust mid-layer performance across Gemma-2 models (2B to 27B)' is not supported by data; if it is real, the label is misleading and the setup should be described. This is load-bearing for the cross-model-size generalization claim and must be corrected before the paper can be accepted.
  3. [Appendix C] Appendix C describes a 35-dimensional manual feature pipeline using attention layers 40, 42, and 44 of Gemma-2-27B, residual layer m=28, SciSpaCy clinical entity extraction, and 'transcript/note chunks.' This is internally inconsistent with §3, which defines a single logistic probe on the final-token residual stream, and with the datasets described in Table 1 (news articles and ContraTales story prefixes, not clinical transcripts). Figure 7 and Table 3 appear to report results from an unrelated system. This makes the method section contradictory and the paper as a whole not reproducible. The appendix must be removed or fully rewritten to match the main method; if it is meant to be part of the method, the main text must be corrected accordingly.
  4. [§4.2, Figure 2] The headline comparison to baselines is not supported by a significance test or by an entailment/NLI-style baseline. On ContraTales, Lookback Lens has a 95% confidence interval of 0.48 ± 0.11 while the probe reports 0.75 ± 0.04; a paired test across folds is needed to establish the claimed 9–27 point advantage. More importantly, a standard NLI/entailment detector (e.g., a DeBERTa NLI model or a strong LLM judge) is an obvious baseline for contextual hallucination detection and is absent. Without it, the claim that the probe 'outperforms baselines by 5–27 points' is not fully established; at minimum the paper should state that NLI baselines are out of scope.
minor comments (5)
  1. [§3.1, Table 1] Section 3.1 says hallucinated continuations were 'produced by prompting gpt-4.1' while Table 1 lists the continuation generator as 'gpt-4.1-mini'; please harmonize the model names throughout.
  2. [Figure 5 caption] The Figure 5 caption is non-descriptive ('We use steering to generate outputs for CNNDM...'); it should state the metric definitions, the sample size per alpha, and whether confidence intervals are shown.
  3. [Limitations] The Limitations section mentions 'medical datasets,' but no medical dataset appears in the paper; either add such a dataset or remove the reference.
  4. [Title and §4.1] The title and abstract emphasize 'a single direction,' but §4.1 states 'The uniqueness of this linear direction is not established here'; please make the wording consistent with the evidence.
  5. [Data availability] The paper states that the ContraTales benchmark is released, but no URL, repository, or download link is provided; please add a data-availability statement.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the central detector is a standard held-out supervised probe evaluation, and the weaknesses identified are validity/control gaps rather than derivations that reduce to their own inputs.

full rationale

The paper's detection claim rests on a logistic probe trained with binary cross-entropy and evaluated by 5-fold cross-validation on held-out pairs from the same synthetic pipeline. That is conventional supervised evaluation, not a fitted parameter renamed as a prediction. Cross-domain transfer trains on one news corpus and tests on the other without re-tuning, so no training label leaks into the test score by construction. Layer sweeps and gradient-times-activation attributions are post-hoc descriptions of the trained probe, not self-validating derivations. The paper explicitly disclaims uniqueness of the direction in Section 4.1 ('The uniqueness of this linear direction is not established here'), so no uniqueness theorem is imported from prior work, and no load-bearing self-citation chain is present; the cited prior work is external. Two weaknesses deserve flagging but are not circularity. First, the causal steering experiment (Section 4.4) patches a probe vector trained on the same Gemma-2-2B observer into that same model and judges hallucination with GPT-4.1, a model from the same family as the synthetic label generator; this lacks a random- or shuffled-direction control and therefore does not show that the learned direction is privileged over generic perturbation. Second, Figure 1 labels the Gemma-2-27B curve as '(sim.)' without explanation, which if simulated would leave the abstract's '2B to 27B' transfer claim unsupported. The paper's own Limitations section concedes both the synthetic-artifact risk and the LLM-judge noise. These are real threats to causal and generalization interpretations, but they are confounds and missing controls, not equivalences by construction: the judge's labels and the RapidFuzz repetition rates are new measurements, not algebraic consequences of the fitted probe. Accordingly, no circular step meets the quote-and-reduction bar, and the appropriate circularity score is 0.

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

The central technical contribution is a supervised linear probe on residual stream activations. The main assumptions are the contested linear representation hypothesis and the representativeness of synthetic GPT-4.1 continuations. The causal steering claims depend on a hand-chosen scaling range and an uncalibrated LLM judge.

free parameters (4)
  • Probe layer selection = varies by model and dataset
    The layer for the probe is selected via inner-fold validation on the training set, so performance numbers depend on this choice.
  • Steering scale alpha = range -60 to +60
    The steering strength is chosen by hand; the magnitude of the effect depends on this arbitrary scaling.
  • ContraTales generation few-shot examples = 24 seed examples
    The synthetic benchmark is generated with a hand-chosen seed set, so the difficulty of the benchmark depends on this choice.
  • GPT-4.1 judge threshold = not specified
    The steering evaluation relies on a GPT-4.1 binary judge; the balance of hallucination versus repetition follows from the judge's own bias, which is not calibrated against human labels.
assumptions (3)
  • domain assumption Linear representation hypothesis: hallucination is encoded along a single linear direction in residual stream activations
    The entire probe method assumes this hypothesis, which the paper itself notes is contested (e.g., Engels et al. 2024).
  • domain assumption The observer model's representation of contextual inconsistency is shared across model sizes and generalizes from synthetic training data
    Cross-model and cross-domain transfer are assumed in order to claim a single transferable direction.
  • domain assumption Gradient-times-activation contributions are a faithful proxy for causal influence
    The attribution method is a first-order estimate, and the paper uses it to claim localization to sparse MLP activity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of A Single Direction of Truth: An Observer Model's Linear Residual Probe Exposes and Steers Contextual Hallucinations." pith.science (2026). https://pith.science/paper/FHTOVCWB

@misc{pith2026250723221,
  author       = {Pith},
  title        = {Pith review of: A Single Direction of Truth: An Observer Model's Linear Residual Probe Exposes and Steers Contextual Hallucinations},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FHTOVCWB}},
  note         = {Machine review of arXiv:2507.23221}
}
read the original abstract

Contextual hallucinations -- statements unsupported by given context -- remain a significant challenge in AI. We demonstrate a practical interpretability insight: a generator-agnostic observer model detects hallucinations via a single forward pass and a linear probe on its residual stream. This probe isolates a single, transferable linear direction separating hallucinated from faithful text, outperforming baselines by 5-27 points and showing robust mid-layer performance across Gemma-2 models (2B to 27B). Gradient-times-activation localises this signal to sparse, late-layer MLP activity. Critically, manipulating this direction causally steers generator hallucination rates, proving its actionability. Our results offer novel evidence of internal, low-dimensional hallucination tracking linked to specific MLP sub-circuits, exploitable for detection and mitigation. We release the 2000-example ContraTales benchmark for realistic assessment of such solutions.

Figures

Figures reproduced from arXiv: 2507.23221 by the authors.

Figure 1
Figure 1. Layer-wise detection performance of residual-stream linear probes. Each curve shows the F1 score (5-fold CV) of a logistic probe trained on a single transformer layer to classify the final sentence of a document as hallucinated or supported by context. Left: results on CNN/DM summarisation; right: results on the synthetic-contradiction CONTRATALES. The consistent mid-layer plateau across four observer models support… view at source ↗
Figure 2
Figure 2. Comparison of hallucination-detection methods. Bars give mean F1 over five cross-validation folds; whiskers show the 95% bootstrap confidence interval. The residual-stream linear probe (right-most bar in each group) consistently exceeds all baselines – lexical overlap, entity verification, semantic similarity, and Lookback Lens – across the news datasets (XSUM, CNN/DM) and the logically harder CONTRATALES. the lowes… view at source ↗
Figure 3
Figure 3. Cross-domain transfer performance of hallucination detection methods. F1 scores for detectors trained on one news dataset (CNN/DM or XSUM) and evaluated on the other. Features were extracted from layer 20 of a Gemma-2-9B observer. The linear probe demonstrates high transferability compared to baseline methods. CNN/DM XSum ContraTales 10 9 8 7 6 5 4 3 2 1 −0.2 −0.1 0 0.1 0.2 Attribution Value Dataset Layer [PITH_FUL… view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: Aggregated MLP layer attributions for the hallucination probe. Mean MLP attributions (A¯ (ℓ) mlp) per layer for a linear probe trained on layer 10 of Gemma-2- 9B. Attributions are presented for evaluations on CNN/DM, XSUM, and CONTRATALES, revealing a consistent patter…
Figure 5
Figure 5. Figure 5: We use steering to generate outputs for CNNDM [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 7
Figure 7. Figure 7: Comparison of top feature importances for hallucination detection across datasets. [PITH_FULL_IMAGE:figures/full_fig_p015_7.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

37 extracted references · 6 canonical work pages

  1. [1]

    and Bengio, Y

    Alain, G. and Bengio, Y. Understanding intermediate layers using linear classifier probes. arXiv preprint arXiv:1610.01644, 2016

    arXiv 2016

  2. [2]

    and Mitchell, T

    Azaria, A. and Mitchell, T. The Internal State of an LLM Knows When It’s Lying . In Findings of the Association for Computational Linguistics: EMNLP 2023, pp.\ 967--976, Singapore, Dec 2023. Association for Computational Linguistics. doi:10.18653/v1/2023.findings-emnlp.68. URL https://aclanthology.org/2023.findings-emnlp.68/

    doi:10.18653/v1/2023.findings-emnlp.68 2023

  3. [3]

    Discovering Latent Knowledge in Language Models Without Supervision

    Burns, C., Ye, H., Klein, D., and Steinhardt, J. Discovering Latent Knowledge in Language Models Without Supervision . In International Conference on Learning Representations, 2023. URL https://openreview.net/forum?id=ETKGuby0hcs

    work page 2023

  4. [4]

    Cammarata, N., Carter, S., Goh, G., Olah, C., Petrov, M., Schubert, L., Voss, C., Egan, B., and Lim, S. K. Thread: Circuits. Distill, 2020. doi:10.23915/distill.00024. URL https://distill.pub/2020/circuits

    doi:10.23915/distill.00024 2020

  5. [5]

    Chuang, Y.-S., Qiu, L., Hsieh, C.-Y., Krishna, R., Kim, Y., and Glass, J. R. Lookback Lens: Detecting and Mitigating Contextual Hallucinations in Large Language Models Using Only Attention Maps , 2024. URL https://arxiv.org/abs/2407.07071

    arXiv 2024

  6. [6]

    Sparse autoencoders find highly interpretable features in language models

    Cunningham, H., Ewart, A., Riggs, L., Huben, R., and Sharkey, L. Sparse autoencoders find highly interpretable features in language models. arXiv preprint arXiv:2309.08600, 2023

    arXiv 2023

  7. [7]

    H., Lasenby, R., Drain, D., Chen, C., Grosse, R., McCandlish, S., Kaplan, J., Amodei, D., Wattenberg, M., and Olah, C

    Elhage, N., Hume, T., Olsson, C., Schiefer, N., Henighan, T., Kravec, S., Dodds, Z. H., Lasenby, R., Drain, D., Chen, C., Grosse, R., McCandlish, S., Kaplan, J., Amodei, D., Wattenberg, M., and Olah, C. Toy models of superposition. arXiv preprint arXiv:2209.10652, 2022. URL https://arxiv.org/abs/2209.10652

    arXiv 2022

  8. [8]

    J., Gurnee, W., and Tegmark, M

    Engels, J., Liao, I., Michaud, E. J., Gurnee, W., and Tegmark, M. Not all language model features are linear. arXiv e-prints, pp.\ arXiv--2405, 2024

    work page 2024

Show all 37 references
  1. [9]

    Detecting hallucinations in large language models using semantic entropy

    Farquhar, S., Kossen, J., Kuhn, L., and Gal, Y. Detecting hallucinations in large language models using semantic entropy . Nature, 630: 0 625--630, 2024. doi:10.1038/s41586-024-07421-0

    2024 doi

  2. [10]

    Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models

    Ferrando, J., Obeso, O., Rajamanoharan, S., and Nanda, N. Do I Know This Entity? Knowledge Awareness and Hallucinations in Language Models . arXiv preprint arXiv:2411.14257, 2024

    2024 arXiv

  3. [11]

    The pile: An 800gb dataset of diverse text for language modeling

    Gao, L., Biderman, S., Black, S., Golding, L., Hoppe, T., Foster, C., Phang, J., He, H., Thite, A., Nabeshima, N., et al. The pile: An 800gb dataset of diverse text for language modeling. arXiv preprint arXiv:2101.00027, 2020

    2020 arXiv

  4. [12]

    spacy: Industrial-strength natural language processing in python

    Honnibal, M., Montani, I., Van Landeghem, S., Boyd, A., et al. spacy: Industrial-strength natural language processing in python. 2020

    2020

  5. [13]

    A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions

    Huang, L., Yu, W., Ma, W., Zhong, W., Feng, Z., Wang, H., Chen, Q., Peng, W., Feng, X., Qin, B., and Liu, T. A Survey on Hallucination in Large Language Models: Principles, Taxonomy, Challenges, and Open Questions . ACM Transactions on Information Systems, 1 0 (1): 0 1--58, 20...

    2024 doi

  6. [14]

    M., Maxwell, T., Cheng, N., et al

    Hubinger, E., Denison, C., Mu, J., Lambert, M., Tong, M., MacDiarmid, M., Lanham, T., Ziegler, D. M., Maxwell, T., Cheng, N., et al. Sleeper agents: Training deceptive llms that persist through safety training. arXiv preprint arXiv:2401.05566, 2024

    2024 arXiv

  7. [15]

    Survey of Hallucination in Natural Language Generation

    Ji, Z., Lee, N., Frieske, R., Yu, T., Su, D., Xu, Y., and Fung, P. Survey of Hallucination in Natural Language Generation . ACM Computing Surveys, 55 0 (12): 0 1--38, 2023. doi:10.1145/3571730

    2023 doi

  8. [16]

    Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs , 2024

    Kossen, J., Han, J., Razzak, M., Schut, L., Malik, S., and Gal, Y. Semantic Entropy Probes: Robust and Cheap Hallucination Detection in LLMs , 2024. URL https://arxiv.org/abs/2406.15927

    2024 arXiv

  9. [17]

    and Frey, B

    Makhzani, A. and Frey, B. K-sparse autoencoders. arXiv preprint arXiv:1312.5663, 2013

    2013 arXiv

  10. [18]

    Manakul, P., Liusie, A., and Gales, M. J. F. SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language Models . In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp.\ 9004--9017, Singapore, Dec 2023. As...

    2023 doi

  11. [19]

    J., Belinkov, Y., Bau, D., and Mueller, A

    Marks, S., Rager, C., Michaud, E. J., Belinkov, Y., Bau, D., and Mueller, A. Sparse feature circuits: Discovering and editing interpretable causal graphs in language models. arXiv preprint arXiv:2403.19647, 2024

    2024 arXiv

  12. [20]

    On faithfulness and factuality in abstractive summarization

    Maynez, J., Narayan, S., Bohnet, B., and McDonald, R. On faithfulness and factuality in abstractive summarization. arXiv preprint arXiv:2005.00661, 2020

    2005 arXiv

  13. [21]

    Controllable Context Sensitivity and the Knob Behind It

    Minder, J., Du, K., Stoehr, N., Monea, G., Wendler, C., West, R., and Cotterell, R. Controllable Context Sensitivity and the Knob Behind It . arXiv preprint arXiv:2411.07404, 2024

    2024 arXiv

  14. [22]

    A., and Kriegeskorte, N

    Mur, M., Bandettini, P. A., and Kriegeskorte, N. Revealing representational content with pattern-information fmri—an introductory guide. Social cognitive and affective neuroscience, 4 0 (1): 0 101--109, 2009

    2009

  15. [23]

    Nan, F., Nallapati, R., Wang, Z., Santos, C. N. d., Zhu, H., Zhang, D., McKeown, K., and Xiang, B. Entity-level factual consistency of abstractive text summarization. arXiv preprint arXiv:2102.09130, 2021

    2021 arXiv

  16. [24]

    Emergent linear representations in world models of self-supervised sequence models

    Nanda, N., Lee, A., and Wattenberg, M. Emergent linear representations in world models of self-supervised sequence models. arXiv preprint arXiv:2309.00941, 2023

    2023 arXiv

  17. [25]

    B., and Lapata, M

    Narayan, S., Cohen, S. B., and Lapata, M. Don't give me the details, just the summary! topic-aware convolutional neural networks for extreme summarization. arXiv preprint arXiv:1808.08745, 2018

    2018 arXiv

  18. [26]

    J., and Veitch, V

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

    2023 arXiv

  19. [27]

    A practical review of mechanistic interpretability for transformer-based language models

    Rai, D., Zhou, Y., Feng, S., Saparov, A., and Yao, Z. A practical review of mechanistic interpretability for transformer-based language models. arXiv preprint arXiv:2407.02646, 2024. URL https://arxiv.org/abs/2407.02646

    2024

  20. [28]

    Hallushield: A mechanistic approach to hallucination resistant models

    Research, A. Hallushield: A mechanistic approach to hallucination resistant models. https://apartresearch.com/project/hallushield-a-mechanistic-approach-to-hallucination-resistant-models, 2025. White paper

    2025

  21. [29]

    J., and Manning, C

    See, A., Liu, P. J., and Manning, C. D. Get to the point: Summarization with pointer-generator networks. arXiv preprint arXiv:1704.04368, 2017

    2017 arXiv

  22. [30]

    Constructing benchmarks and interventions for combating hallucinations in llms

    Simhi, A., Herzig, J., Szpektor, I., and Belinkov, Y. Constructing benchmarks and interventions for combating hallucinations in llms. arXiv preprint arXiv:2404.09971, 2024

    2024 arXiv

  23. [31]

    Trust Me, I'm Wrong: High-Certainty Hallucinations in LLMs

    Simhi, A., Itzhak, I., Barez, F., Stanovsky, G., and Belinkov, Y. Trust Me, I'm Wrong: High-Certainty Hallucinations in LLMs . arXiv preprint arXiv:2502.12964, 2025. URL https://arxiv.org/abs/2502.12964

    2025 arXiv

  24. [32]

    The curious case of hallucinatory (un) answerability: Finding truths in the hidden states of over-confident large language models

    Slobodkin, A., Goldman, O., Caciularu, A., Dagan, I., and Ravfogel, S. The curious case of hallucinatory (un) answerability: Finding truths in the hidden states of over-confident large language models. arXiv preprint arXiv:2310.11877, 2023

    arXiv

  25. [33]

    Redeep: Detecting hallucination in retrieval augmented generation via mechanistic interpretability

    Sun, Z., Zang, X., Zheng, K., Song, Y., Xu, J., Zhang, X., Yu, W., and Li, H. Redeep: Detecting hallucination in retrieval augmented generation via mechanistic interpretability. arXiv preprint arXiv:2410.11414, 2024. URL https://arxiv.org/abs/2410.11414

    2024

  26. [34]

    Cost-Effective Hallucination Detection for LLMs , 2024

    Valentin, S., Fu, J., Detommaso, G., Xu, S., Zappella, G., and Wang, B. Cost-Effective Hallucination Detection for LLMs , 2024. URL https://arxiv.org/abs/2407.21424

    2024 arXiv

  27. [35]

    Yu, L., Cao, M., Cheung, J. C. K., and Dong, Y. Mechanistic understanding and mitigation of language model non factual hallucinations. In Findings of the Association for Computational Linguistics: EMNLP 2024, pp.\ 7943--7956, Miami, USA, 2024. doi:10.18653/v1/2024.findings-emn...

    2024 doi

  28. [36]

    Attention satisfies: A constraint-satisfaction lens on factual errors of language models, 2024

    Yuksekgonul, M., Chandrasekaran, V., Jones, E., Gunasekar, S., Naik, R., Palangi, H., Kamar, E., and Nushi, B. Attention satisfies: A constraint-satisfaction lens on factual errors of language models, 2024. URL https://arxiv.org/abs/2309.15098

    2024 arXiv

  29. [37]

    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 6, 2026 · model on record in the stance chip above.