Pith. sign in

REVIEW 3 major objections 5 minor 17 references

On the Generalizability of "Competition of Mechanisms: Tracing How Language Models Handle Facts and Counterfactuals"

T0 review · 3 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 This reproduction confirms the core 'competition of mechanisms' findings for GPT-2 and Pythia, but shows the recommended attention-head ablation does not generalize across architectures, prompts, or domains.

desk verdict Reproduction is solid and the domain-leakage finding is important, but the abstract overclaims the Llama result that the body itself leaves unresolved. read the letter →

arxiv 2506.22977 v1 pith:FWXZ2UHY submitted 2025-06-28 cs.CL cs.LG

Asen Dotsinski , Udit Thakur , Marko Ivanov , Mohammad Hafeez Khan , Maria Heuss This is my paper
classification cs.CLcs.LG
keywords mechanisticinterpretabilityattentionheadablationcounterfactualreasoninglogitlenscompetitionofmechanismsreproductionstudydatasetdomainbiaslanguagemodelgeneralization
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 asks how far the 'competition of mechanisms' explanation of how language models handle facts versus counterfactuals actually extends. It independently reproduces the original study's three core findings—that factual and counterfactual information occupy different token positions, that attention blocks dominate the competition, and that a few specialized attention heads decide the outcome—on GPT-2 small and Pythia 6.9B. It then extends the setup to Llama 3.1 8B, question-format prompts, different premise words, and a domain-classified version of the datasets, and finds each change matters: Llama shows much weaker attention-head specialization, the question format brings GPT-2 close to a fifty-fifty split, and the recommended head ablation is effective mainly in the automotive and technology categories that dominate the original data. The paper's central claim is that the ablation's effectiveness is conditional on model architecture, prompt structure, domain, and task, so the original mechanism-competition picture does not generalize as stated.

What carries the argument

The paper's working machinery is the logit-lens inspection of residual-stream position-by-position logits, combined with attention-head ablation that scales the attention of the last token to the attribute position (heads L10H7 and L11H10 in GPT-2 small, L17H28/L20H18/L21H8 in Pythia 6.9B) by a factor α. That combination exposes which head writes the counterfactual token into the final position; the paper runs it across models, prompt structures, premise words, and domains to test whether the same heads remain decisive. The central measured quantity is the factual-over-counterfactual margin Δ_cofa, which the ablation is meant to flip.

What would settle it

Take the original and domain-split datasets, remove every prompt whose subject token contains the factual token, and re-run the L10H7/L11H10 ablation on the remaining underrepresented-domain prompts: if factual predictions still rise to roughly 50%, the paper's domain-dependence conclusion is wrong. For the Llama result, replace the logit-lens attribution with activation patching or a tuned lens: if a small set of decisive heads appears, the paper's claim of reduced specialization is an artifact of the measurement tool.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the mechanism-competition story from the original study reproduces cleanly only under the original conditions. On GPT-2 small and Pythia 6.9B with the 'Redefine:' prompt, the paper confirms the positional encoding of factual and counterfactual tokens, the stronger role of attention blocks than MLP blocks, and the existence of a few high-impact heads that promote the counterfactual by suppressing it rather than by boosting the factual token. On Llama 3.1 8B, competition between the two mechanisms still appears, but the head-level specialization largely disappears: only one strong counterfactual head was found, factual contributions are spread widely, and even the best ablation moves factual predictions only from 15% to 22%. Recasting the prompt as a question reduces the copy tendency so much that GPT-2 predicts factual and counterfactual tokens almost equally often, and the choice of premise word (for example 'Review' versus 'Validate') shifts factual-prediction rates from 1.9% to 6.4%, with the gap widening under ablation. Finally, the paper shows that the apparent importance of the subject token for factual recall is inflated by the original dataset's concentration on product domains whose names contain the factual token; in underrepresented domains, the proposed head ablation no longer flips predictions, and on the smaller cross-domain dataset it can even increase counterfactual predictions.

Load-bearing premise

The load-bearing premise is that the logit-lens and attention-attribution tools used in the original pipeline actually reveal the internal causes of the final prediction; if they misattribute contributions, especially for Llama 3.1 8B, both the reproduced findings and the new negative results are compromised.

Editorial extensions

If this is right

  • In small models, the question-format prompt alone brings GPT-2's factual and counterfactual predictions near a fifty-fifty split, making prompt restructuring as effective as the original head ablation; combining the two is far stronger.
  • The recommended GPT-2 heads L10H7 and L11H10 do not transfer across domains: under the same ablation, underrepresented categories mostly still predict the counterfactual token, and the MQuAKE-based test shows weakened or even reversed effects.
  • Larger and newer models like Llama 3.1 8B exhibit the factual and counterfactual competition but without concentrated head specialization, so the original method does not identify the right levers for attention-head control in those models.
  • Because the original dataset's dominant domains contain prompts where the subject itself contains the factual token, the subject-position encoding that the original study emphasized is partly an artifact of dataset bias.

Reading between the lines

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

  • Editorial inference: if the subject-token leakage explanation is right, re-running the same ablation on a leakage-filtered COUNTERFACT subset should restore or abolish the domain effect, giving a direct causal test of the paper's explanation.
  • Editorial inference: the near-equivalence of prompt restructuring and head ablation suggests that for practical safety interventions, sanitizing or reformulating prompts may be cheaper and more reliable than circuit-level edits, a direction the paper's practical discussion implies but does not state as a recommendation.
  • Editorial inference: for Llama-like models, the failure to find specialized heads could be an artifact of the logit-lens method rather than absent specialization; applying a tuned lens or activation patching would settle which reading is correct, and the paper's negative result on Llama remains its least settled finding.
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

3 major / 5 minor

Summary. The paper is a reproduction and extension study of Ortu et al. (2024), which analyzed competition between factual recall and counterfactual copy mechanisms in language models. The authors re-filter the original COUNTERFACT-derived datasets, reproduce the original positional-encoding, attention-block-dominance, and attention-head-specialization findings on GPT-2 small and Pythia 6.9B, and extend the analysis to Llama 3.1 8B, to question-answer prompt reformulations, to alternative premise words, and to domain-stratified subsets. The paper concludes that the proposed attention-head ablation is less effective for underrepresented domains and that its effectiveness varies with architecture, prompt structure, domain, and task.

Significance. If the results are taken at face value, the paper makes a useful contribution to the reproducibility literature in mechanistic interpretability and to the practical question of how widely head-level circuit claims transfer. Its strengths include a careful re-filtering of the datasets, a public code repository, the reproduction of the core GPT-2 and Pythia results, and the domain-leakage observation that many COUNTERFACT prompts contain the factual token inside the subject, which can spuriously explain subject-position importance. The main weakness is that the central architecture-dependence claim rests on Llama 3.1 8B results that the body itself explicitly labels as ambiguous, and the abstract overstates this result and the domain result.

major comments (3)
  1. [Section 4.2 / Abstract] The abstract states that Llama 3.1 8B shows 'greatly reduced attention head specialization' as an established finding, but the body of Section 4.2 explicitly leaves open two possibilities: 'Either the logit lens method fails to discover the highly specialized attention heads, or they are simply not as concentrated for Llama 3.1 8B.' The ablation experiments do not resolve this ambiguity because the ablated heads were selected using the same logit-lens attribution that the paper acknowledges may be unreliable for larger models (citing Belrose et al., 2023), and the resulting increase in factual predictions was only from 15% to 22%. This is load-bearing for the architecture-dependence component of the central claim; the abstract should be reworded to present the finding as provisional, or the authors should provide disambiguating evidence (e.g., tuned-lens attribution, a random-head ablation baseline, or a broader head search).
  2. [Section 4.1 vs. Abstract and Section 2] The abstract and the introduction claim successful reproduction of the original paper's 'localization of factual and counterfactual information,' but Section 4.1 says the opposite for a key part of that claim: 'the claim that factual information is stored in the subject position does not seem to be backed up by our results or those of the original paper.' Since Section 2 lists 'Positional Information Encoding: factual attributes appearing in subject positions' as one of the core reproduced claims, the paper is internally inconsistent about what was reproduced. Please re-scope the reproduction claim in the abstract and Section 2, and clearly separate the confirmed parts (counterfactual in attribute position, monotonic increase with layers) from the unconfirmed part (factual in subject position).
  3. [Section 4.5 / Abstract] The abstract's claim that the Ortu et al. (2024) ablation is 'ineffective for domains that are underrepresented in their dataset' is stronger than the evidence supports. Figure 4 reports only post-ablation counts per domain, with no per-domain baseline, so 'ineffective' (no effect) cannot be distinguished from 'effective but insufficient to flip the majority prediction.' The body itself uses the weaker phrase 'making ablation less effective.' Additionally, the MQuAKE results (Figure 5) rest on only 228 filtered prompts, and the paper acknowledges they are tentative. Please either add per-domain baseline-vs-ablation comparisons and confidence intervals, or revise the abstract and Section 5.2 to state the finding more cautiously, e.g., 'the ablation does not transfer to underrepresented domains in these datasets.'
minor comments (5)
  1. [Section 3.3.2] The phrase 'somewhat arbitrarily decided' is informal for a journal article; consider describing more precisely how the original study selected the heads and scaling factor, and note that this reproduction follows that selection by design.
  2. [Figures 4 and 5] The domain-level win counts are difficult to read because the category names are crowded and the baseline counts are not shown; a table with per-domain baseline and post-ablation factual/counterfactual counts would make the domain claim easier to verify.
  3. [Section 4.3] The statement that the QnA prompt structure is 'significantly more important' than the premise word is based only on GPT-2 small experiments; please state this scope explicitly in the same paragraph and in the conclusion to avoid overgeneralization.
  4. [Section 4.4 / Table 1] The differences across premise words are reported as counts and percentages without any uncertainty quantification; although the counts are deterministic on the fixed dataset, reporting bootstrap intervals or a simple chi-square test would strengthen the comparison.
  5. [Section 3.2.1] When reporting that only 5,179 of 10,000 GPT-2 small prompts and 9,137 of 10,000 Pythia prompts have the factual token as the top prediction, please specify the decoding criterion used (e.g., greedy top-1 vs. highest logit) and how the original dataset was obtained, so that this verification is fully reproducible.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the reproduction and extensions test external claims against fresh data and independent benchmarks, so no claimed result reduces to its inputs by construction.

full rationale

This is a reproduction and extension study, not a derivation. The paper independently verifies claims from Ortu et al. (2024) on GPT-2 and Pythia, then tests those claims on Llama 3.1 8B, new prompt structures, premise words, and new domains. The central negative results, such as the reduced effectiveness of the proposed attention-head ablation on underrepresented domains, are evaluated on data that was not used to select the heads; the ablation is inherited from the original study for consistency rather than fitted to the new experiments. The only inherited elements are the same head identities and scaling factors, explicitly chosen to maintain consistency with Ortu et al. (2024), which is a benchmark choice rather than a circular step. The ambiguity acknowledged in Section 4.2, namely that the logit lens may fail on Llama 3.1 8B, is a methodological validity risk and is explicitly hedged in the body and limitations; it does not make any result equivalent to its inputs by definition. No fitted parameter is renamed as a prediction, and no load-bearing claim depends on a self-citation chain. The paper is honest about its limitations, including the small MQuAKE sample and the reliance on GPT-2 for several extensions. Overall, the derivation chain is self-contained with respect to the claims it advances, so no circularity is present.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The study introduces no new entities or theoretical constructs. Its analysis depends on several assumptions from the interpretability toolkit and on dataset filtering choices. Free parameters are the arbitrary ablation settings and premise word set.

free parameters (3)
  • Ablation scaling factor alpha = 5 (also 50 and 0 in Llama experiments)
    The magnitude of attention scaling in ablation experiments is arbitrary and inherited from Ortu et al. (2024); the paper notes the choice is somewhat arbitrary (Section 3.3.2).
  • Number and identity of ablated attention heads = L10H7, L11H10 for GPT-2; L17H28, L20H18, L21H8 for Pythia; L27H20, L28H15, L31H14 for Llama
    Heads are selected from qualitative attribution scores, not from a formal criterion; ablation effectiveness is the main dependent variable, so the head choice is a free parameter.
  • Premise words set = Redefine, Assess, Fact Check, Review, Validate, Verify
    The set of premise words is chosen by the authors; results vary across them, making the set itself a parameter of the study.
assumptions (5)
  • domain assumption The residual stream decomposition and logit lens projection (Eq. 1-3) faithfully reflect the model's internal computations
    Used throughout to attribute factual and counterfactual logits to positions, layers, and components (Section 3.3.1). The paper notes Belrose et al. (2023) suggest this may be unreliable for large models.
  • domain assumption Scaling attention weights by alpha is a causally valid intervention that isolates the selected head's role
    Attention modification (Eq. 4) is used to test head importance; the paper acknowledges the choice of heads and scaling is somewhat arbitrary (Section 3.3.2).
  • ad hoc to paper The prompt template 'Redefine: {s} {r} {tcofa}. {s} {r}' triggers a well-defined competition between factual recall and counterfactual copy mechanisms
    All reproduction experiments use this template inherited from Ortu et al. (2024); the paper's own QnA results show the template is not neutral.
  • domain assumption Domain labels from NeMo Curator are accurate and meaningful for grouping prompts
    The domain analysis relies on automatic classification into 26 categories (Section 3.2.4).
  • ad hoc to paper The filtered datasets (only prompts where the model predicts factual or counterfactual token) are representative of the underlying phenomena
    Filtering reduces the original 10,000-prompt datasets substantially, especially MQuAKE (228 prompts), potentially introducing selection bias (Section 3.2.4).

how reviews work

0 comments
Cite this review

Pith. "Pith review of On the Generalizability of "Competition of Mechanisms: Tracing How Language Models Handle Facts and Counterfactuals"." pith.science (2026). https://pith.science/paper/FWXZ2UHY

@misc{pith2026250622977,
  author       = {Pith},
  title        = {Pith review of: On the Generalizability of "Competition of Mechanisms: Tracing How Language Models Handle Facts and Counterfactuals"},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FWXZ2UHY}},
  note         = {Machine review of arXiv:2506.22977}
}
read the original abstract

We present a reproduction study of "Competition of Mechanisms: Tracing How Language Models Handle Facts and Counterfactuals" (Ortu et al., 2024), which investigates competition of mechanisms in language models between factual recall and counterfactual in-context repetition. Our study successfully reproduces their primary findings regarding the localization of factual and counterfactual information, the dominance of attention blocks in mechanism competition, and the specialization of attention heads in handling competing information. We reproduce their results on both GPT-2 (Radford et al., 2019) and Pythia 6.9B (Biderman et al., 2023). We extend their work in three significant directions. First, we explore the generalizability of these findings to even larger models by replicating the experiments on Llama 3.1 8B (Grattafiori et al., 2024), discovering greatly reduced attention head specialization. Second, we investigate the impact of prompt structure by introducing variations where we avoid repeating the counterfactual statement verbatim or we change the premise word, observing a marked decrease in the logit for the counterfactual token. Finally, we test the validity of the authors' claims for prompts of specific domains, discovering that certain categories of prompts skew the results by providing the factual prediction token as part of the subject of the sentence. Overall, we find that the attention head ablation proposed in Ortu et al. (2024) is ineffective for domains that are underrepresented in their dataset, and that the effectiveness varies based on model architecture, prompt structure, domain and task.

Figures

Figures reproduced from arXiv: 2506.22977 by the authors.

Figure 1
Figure 1. The logit values for Pythia 6.9B across different positions and layers on the filtered "Redefine" [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Logit inspection for Llama 3.1 8B on the GPT-2 small "Redefine" dataset. [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Logit inspection for GPT-2 small on different domains. The results for "Autos and Vehicles" closely [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (22 more)
Figure 4
Figure 4. Figure 4: GPT-2 domain-level wins at attribute position after ablation of L10H7 and L11H10 [PITH_FULL_IMAGE:figures/full_fig_p011_4.png]
Figure 5
Figure 5. Figure 5: Per domain ablation results for GPT2 on the filtered MQuAKE dataset (GPT-2 small, [PITH_FULL_IMAGE:figures/full_fig_p011_5.png]
Figure 6
Figure 6. Figure 6: Logit inspection for GPT-2 small on the "Redefine" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p016_6.png]
Figure 7
Figure 7. Figure 7: Logit inspection for GPT-2 small, per attention head, on the "Redefine" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p016_7.png]
Figure 8
Figure 8. Figure 8: Logit inspection of the aggregate impact of attention and MLP blocks on [PITH_FULL_IMAGE:figures/full_fig_p016_8.png]
Figure 9
Figure 9. Figure 9: Logit inspection for Pythia 6.9B on the "Redefine" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_9.png]
Figure 10
Figure 10. Figure 10: Logit inspection for Pythia 6.9B, per attention head, on the "Redefine" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p017_10.png]
Figure 11
Figure 11. Figure 11: Logit inspection of the aggregate impact of attention and MLP blocks on [PITH_FULL_IMAGE:figures/full_fig_p017_11.png]
Figure 12
Figure 12. Figure 12: Logit inspection for Llama 3.1 8B on the "Redefine" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_12.png]
Figure 13
Figure 13. Figure 13: Logit inspection for Llama 3.1 8B, per attention head, on the "Redefine" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]
Figure 14
Figure 14. Figure 14: Logit inspection of the aggregate impact of attention and MLP blocks on [PITH_FULL_IMAGE:figures/full_fig_p018_14.png]
Figure 15
Figure 15. Figure 15: Comparison of ablation results (L28H15 and L31H14) for Llama 3.1 8B on the "Redefine" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p019_15.png]
Figure 16
Figure 16. Figure 16: Ablation results (L27H20, α = 0) for Llama 3.1 8B on the "Redefine" Dataset. 19 [PITH_FULL_IMAGE:figures/full_fig_p019_16.png]
Figure 17
Figure 17. Figure 17: Logit inspection for Gpt2 on the QnA Dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_17.png]
Figure 18
Figure 18. Figure 18: Logit inspection for GPT-2 small, per attention head, on the "QnA" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p020_18.png]
Figure 19
Figure 19. Figure 19: Logit inspection of the aggregate impact of attention and MLP blocks on [PITH_FULL_IMAGE:figures/full_fig_p020_19.png]
Figure 20
Figure 20. Figure 20: Logit inspection for Pythia 6.9B on the "QnA" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_20.png]
Figure 21
Figure 21. Figure 21: Logit inspection for Pythia 6.9B, per attention head, on the "QnA" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p021_21.png]
Figure 22
Figure 22. Figure 22: Logit inspection of the aggregate impact of attention and MLP blocks on [PITH_FULL_IMAGE:figures/full_fig_p021_22.png]
Figure 23
Figure 23. Figure 23: Logit inspection for Llama 3.1 8B on the "QnA" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p022_23.png]
Figure 24
Figure 24. Figure 24: Logit inspection for Llama 3.1 8B, per attention head, on the "QnA" Dataset. [PITH_FULL_IMAGE:figures/full_fig_p022_24.png]
Figure 25
Figure 25. Figure 25: Logit inspection of the aggregate impact of attention and MLP blocks on [PITH_FULL_IMAGE:figures/full_fig_p022_25.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

17 extracted references · 9 canonical work pages

  1. [1]

    Eliciting latent predictions from transformers with the tuned lens.arXiv preprint arXiv:2303.08112,

    Nora Belrose, Zach Furman, Logan Smith, Danny Halawi, Igor Ostrovsky, Lev McKinney, Stella Biderman, and Jacob Steinhardt. Eliciting latent predictions from transformers with the tuned lens.arXiv preprint arXiv:2303.08112,

    arXiv

  2. [6]

    Dissecting recall of factual associations inauto-regressivelanguagemodels

    Mor Geva, Jasmijn Bastings, Katja Filippova, and Amir Globerson. Dissecting recall of factual associations inauto-regressivelanguagemodels. In Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, pp. 12216–12235,

    work page 2023

  3. [7]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783 ,

    arXiv

  4. [8]

    Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437,

    Aixin Liu, Bei Feng, Bing Xue, Bingxuan Wang, Bochao Wu, Chengda Lu, Chenggang Zhao, Chengqi Deng, Chenyu Zhang, Chong Ruan, et al. Deepseek-v3 technical report.arXiv preprint arXiv:2412.19437,

    arXiv

  5. [10]

    Accessed: 2025-02-21. Nvidia. Nemocurator domain classifier.https://huggingface.co/nvidia/domain-classifier. Accessed: 2025-01-30. Francesco Ortu, Zhijing Jin, Diego Doimo, Mrinmaya Sachan, Alberto Cazzaniga, and Bernhard Schölkopf. Competition of mechanisms: Tracing how language models handle facts and counterfactuals. In Lun- Wei Ku, Andre Martins, and ...

    2025

  6. [11]

    Model evaluation for extreme risks

    Toby Shevlane, Sebastian Farquhar, Ben Garfinkel, Mary Phuong, Jess Whittlestone, Jade Leung, Daniel Kokotajlo, Nahema Marchal, Markus Anderljung, Noam Kolt, et al. Model evaluation for extreme risks. arXiv preprint arXiv:2305.15324 ,

    arXiv

  7. [12]

    Snellius Energy Aware Runtime

    SURF User Knowledge Base. Snellius Energy Aware Runtime. https://servicedesk.surf.nl/wiki/ pages/viewpage.action?pageId=62226671. Accessed: 2025-01-30. Miles Turpin, Julian Michael, Ethan Perez, and Samuel Bowman. Language models don’t always say what they think: Unfaithful explanations in chain-of-thought prompting. Advances in Neural Information Process...

    work page 2025

  8. [13]

    Ai sandbagging: Language models can strategically underperform on evaluations.arXiv preprint arXiv:2406.07358 ,

    Teun van der Weij, Felix Hofstätter, Ollie Jaffe, Samuel F Brown, and Francis Rhys Ward. Ai sandbagging: Language models can strategically underperform on evaluations.arXiv preprint arXiv:2406.07358 ,

    arXiv

Show all 17 references
  1. [14]

    Transformers: State-of-the-art natural language processing

    Thomas Wolf, Lysandre Debut, Victor Sanh, Julien Chaumond, Clement Delangue, Anthony Moi, Pierric Cistac, Tim Rault, Remi Louf, Morgan Funtowicz, Joe Davison, Sam Shleifer, Patrick von Platen, Clara Ma, Yacine Jernite, Julien Plu, Canwen Xu, Teven Le Scao, Sylvain Gugger, Mari...

    2020

  2. [15]

    doi: 10.18653/v1/2020.emnlp-demos.6

    Association for Computational Lin- guistics. doi: 10.18653/v1/2020.emnlp-demos.6. URL https://aclanthology.org/2020.emnlp-demos. 6/. Zexuan Zhong, Zhengxuan Wu, Christopher D Manning, Christopher Potts, and Danqi Chen. Mquake: Assessing knowledge editing in language models via...

    2020 doi

  3. [16]

    Redefine

    15 Published in Transactions on Machine Learning Research (06/2025) A "Redefine" Dataset Plots A.1 GPT-2 small Plots 0 1 2 3 4 5 6 7 8 9 10 11 Layer Subject Relation Relation Last Attribute* Subject repeat Relation repeat Last 0 1 2 3 4 5 6 7 8 9 10 11 Layer Subject Relation R...

    2025

  4. [17]

    Redefine

    for Llama 3.1 8B on the "Redefine" Dataset. 19 Published in Transactions on Machine Learning Research (06/2025) B QnA Dataset Plots B.1 GPT-2 small Plots 0 1 2 3 4 5 6 7 8 9 10 11 Layer Subject Relation Relation Last Attribute* Interrogative Relation repeat Subject repeat Last...

    2025

  5. [2020]

    Jail- breaking black box large language models in twenty queries.arXiv preprint arXiv:2310.08419 ,

    Patrick Chao, Alexander Robey, Edgar Dobriban, Hamed Hassani, George J Pappas, and Eric Wong. Jail- breaking black box large language models in twenty queries.arXiv preprint arXiv:2310.08419 ,

    arXiv

  6. [2021]

    Transformer feed-forward layers are key-value memories

    Mor Geva, Roei Schuster, Jonathan Berant, and Omer Levy. Transformer feed-forward layers are key-value memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing , pp. 5484–5495,

    2021

  7. [2022]

    Modified counterfact dataset

    Neel Nanda. Modified counterfact dataset. https://huggingface.co/datasets/NeelNanda/ counterfact-tracing. Accessed: 2025-01-29. Neel Nanda and Joseph Bloom. Transformerlens. https://github.com/TransformerLensOrg/ TransformerLens,

    2025

  8. [2023]

    Language models are few-shot learners

    13 Published in Transactions on Machine Learning Research (06/2025) Tom Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, et al. Language models are few-shot learners. Advances ...

    2025

  9. [2024]

    Accessed: 2025-01-30

    https:// www.dutchdatacenters.nl/en/publications/state-of-the-dutch-data-centers-2024/ . Accessed: 2025-01-30. Nelson Elhage, Neel Nanda, Catherine Olsson, et al. A mathematical framework for transformer circuits. Transformer Circuits Thread,

    2024

Pith tools

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