Machine Learning · arXiv:2610.21189 v1 · submitted 15 Oct 2026 · 82 references
Resisting Harmful Instructions through Training-Free Weight Editing
The paper says
“Instruction-following language models can comply with requests that conflict with intended behaviors and safety alignment. We introduce a training-free weight-editing approach for reducing compliance with targeted undesirable instructions.”
more of the abstract
Our method needs only forward passes on a small set of prompt pairs per target behavior, and because the edit is stored in the model's weights, changing or deleting instructions in the prompt cannot undo the weight modification. We evaluate the approach on two target harmful behaviors across five instruction-tuned models: instructed falsehood and instructed sexism. On two instructed-lying benchmarks, it improves mean honest response rates by 6.2 and 4.6 percentage points over the strongest alternative, significantly on three of the five models and within noise on the other two. In a controlled study of seven prompt formats, it maintains higher honest-response rates than single-layer activation steering when orders to lie arrive in single-turn formats not used to construct the edit. For instructed sexism, it reduces the mean sexist-response rate from 57.9% to 24.3%; combining the edit with a fairness prompt reduces it further to 11.9%. These results establish our contrast-based weight editing approach as an effective training-free intervention in two distinct behavioral domains, supporting its use both independently and alongside prompt-based safeguards.
from the abstract
Pith says
Machine review · deepseek-v4-flash · 16 Oct 2026 · what was read was not recorded · not checked by a person
It can be wrong. Challenge this review or reply to it.
Holds up. A clean, well-tested weight-editing method for reducing harmful instruction compliance, with limitations honestly stated. This line is wrong
4 major objections · 6 minor
How often objections from reviews like this one are right: not yet measured
Not checked: figures, code, what the 82 references say.
The paper says, at §3.2, Appendix E
The review did not quote this passage.
Pith says: Objection 1 of 4 · major
The hyperparameter search for the proposed method is extensive: five prompt families, multiple editing strengths α, and for sexism also the contrast ratio r, each selected per model on validation data (Section 3.2, Appendix E). In contrast, the baselines (steering: 3 layers × 5 strengths; set difference: 6 strengths; honest prompt: 3 wordings) have far fewer candidates.
more of objection 1
The paper acknowledges different search budgets in the limitations, but the main comparisons (Table 1, Table 2) do not control for this imbalance. To strengthen the claim that the method is practically effective, the authors should either (a) re-run the comparison with a matched search budget (e.g., by sampling the same number of configurations for each method), or (b) report the performance of the proposed method with a fixed default prompt family and default α,r to show that the gains do not entirely depend on per-model tuning.
The paper says, at §4.4, Table 3; §B, Table 6
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.64):
“We have not run the layer ablation with the set-difference mask, which would separate task from rule directly. 12 A.4 Transfer to further tasks A.5 Transfer to further tasks The edit generalizes to single-turn instructions be- yond those used for selection.”
Pith says: Objection 2 of 4 · major
The format-transfer study compares the weight edit to single-layer contrastive activation addition (CAA). However, the paper's own ablation (Table 6) shows that multi-layer steering can improve over single-layer steering on some models (e.g., Phi-4-mini, Falcon3-3B). The claim that the edit "maintains higher honest-response rates than single-layer activation steering" is true for the comparison made, but it is not yet established …
more of objection 2
… that it is a property of weight space vs. activation space per se, because multi-layer steering is not included as a comparator in the format study. The paper acknowledges this limitation in §4.4, but the conclusion in the abstract and §7 should be qualified to reflect that the comparison is against single-layer CAA, not against all activation-space methods.
The paper says, at §4.1, Table 1; §B, Tables 4–5
The review did not quote this passage.
Pith says: Objection 3 of 4 · major
On two models (OLMo-2-7B on LIARS’ BENCH, Qwen2.5-3B on AI Liar), the lead over the best baseline is within noise (confidence interval covers zero). The paper is transparent about this, but the central claim of effectiveness across five models is weakened by these non-significant results.
more of objection 3
The title and abstract state the method 'improves mean honest response rates by 6.2 and 4.6 percentage points over the strongest alternative,' which is true on average, but readers should be aware that the improvement is not uniform. The authors should consider softening the claim in the abstract (e.g., 'on three of the five models the gain is significant') to match the results more precisely.
The paper says, at §2, Eq. (2)–(3); Appendix L
The review did not quote this passage.
A classifier read this sentence of the paper as acknowledging the concern (probability 0.81):
“This is an optimization of a surrogate quantity: a difference of upper bounds is not a bound on the difference between realized perturbations.”
Pith says: Objection 4 of 4 · major
The scoring rule is motivated by a local perturbation bound (Appendix L) that holds at fixed layer inputs and assumes all edited weights have positive contrast. The paper acknowledges that the bound is a surrogate and that the padding (nonpositive contrast weights) can weaken the guarantee.
more of objection 4
However, the empirical success of the method is not directly tied to the bound; the bound is used as motivation. The authors should clarify whether the bound actually predicts which edits will work (e.g., does the bound correlate with honest-rate improvement across hyperparameters?), or whether it is purely a post-hoc justification. Without such evidence, the bound adds little to the empirical results and could be de-emphasized.
6 minor objections
The paper says, at §1, Figure 1
Pith says: minor objection 1 of 6
Figure 1 caption says 'cj > 0 ⇔ n̄F_j > 2.7 n̄T_j', but from Eq. (2) with r=0.37, the threshold is 1/r ≈ 2.7 only when v = r/(1+r) ≈ 0.27. A quick derivation would help: c_j = v n̄F_j - (1-v)n̄T_j > 0 ⇔ n̄F_j > ((1-v)/v)n̄T_j = (1/r)n̄T_j. Consider adding this to the caption for clarity.
The paper says, at §3.2
Pith says: minor objection 2 of 6
The usability checks (perplexity within 1.15×, honesty control drop <5 points, polarity check) are defined but the exact thresholds and rationale (why 1.15×, why 5 points) are not justified. A sentence citing prior work (e.g., Li et al., 2024 for the 1.15× threshold) or explaining the choice would improve reproducibility.
show 4 more
The paper says, at §4.4, Table 3
Pith says: minor objection 3 of 6
The format study reports results on validation and test splits. Please state explicitly that the test set results (Table 3) were obtained from the configurations chosen on the validation set, and that no further tuning was done on the test set. This is good practice and should be stated for transparency.
The paper says, at Appendix L, Eq. (9)
Pith says: minor objection 4 of 6
The inequality B_i(T) ≤ r B_i(F) is derived under the assumption that all selected weights have positive contrast. In practice, the mask may include nonpositive-score weights (padding). The paper discusses this in the next paragraph, but consider noting in the main text (§2) that the bound is only approximate for the full mask, to avoid overstating the theoretical guarantee.
The paper says, at §6, Related Work
Pith says: minor objection 5 of 6
The related work section is concise but omits a mention of weight-intervention methods like task vectors (Ilharco et al., 2023) and model patching (e.g., Meng et al., 2022). While the paper focuses on training-free methods, a brief acknowledgment of these alternative weight-edit approaches would provide a more complete context.
The paper says, at §1, Abstract
Pith says: minor objection 6 of 6
The abstract states 'significantly on three of the five models and within noise on the other two.' This is an accurate summary, but consider stating the p-value or confidence-level threshold (e.g., '95% bootstrap interval excludes zero') in the abstract for clarity.
What would settle it
As a further sanity check, one could recompute the main results (Table 1) using a completely independent set of 25 prompt pairs for each model, drawn without validation-based family selection, to quantify the variance introduced by prompt selection; the existing analyses in Appendices A.2 and M already indicate this would not change the direction of the conclusions.
from the skeptic's reading
How the review went
4 readings by deepseek-v4-flash · 16 Oct 2026
- Reader holds up
- Skeptic partly agrees
- Editor wrote the verdict above
- Referee minor revision: 4 major objections, 6 minor
Not run on this paper: the simulated authors' defence.
Pith did not check this review's objections against the paper.
The review in full
The review in brief, machine-written
The paper's central contribution is a graded activation-contrast scoring rule that selects weights for editing without training. The method is simple—forward passes on 25 paired prompts—and the experiments are thorough: five model families, two undesirable behaviors (lying and sexism), multiple baselines including set-difference and activation steering, ablations, and a prompt-format transfer study. The results are consistent and practically meaningful: honest-response rates improve by 6–10 percentage points over the best alternative on two lying benchmarks, and the sexist-response rate drops from 57.9% to 24.3%. The paper also demonstrates that the edit can complement prompting, further reducing sexist responses to 11.9%. A matched ablation shows the graded contrast rule outperforms the set-difference rule by 13.6 percentage points on average, which supports the design choice. The main soft spot is the causal interpretation of the activation contrast. The paper acknowledges that the local perturbation bound in Appendix L is a surrogate, not a guarantee. The empirical evidence supports correlation but not strict causality. That is fine for a practical paper, but readers should not over-interpret the mechanism. The method is also sensitive to the choice of prompt family and hyperparameters (r, alpha, rho), as evidenced by the validation-based selection over five families per model—this limits turnkey applicability without validation data. The paper is transparent about this and tests on separate benchmarks and a format-transfer study, which mitigates the concern. Overall, the claims are appropriately scoped and the limitations are clearly stated. I would accept a peer-review invitation for this paper. It is a solid empirical contribution that advances the state of the art in training-free safety interventions. The promised code and data should help reproducibility. It is not a breakthrough, but it is honest, useful work.
The strongest objection
The paper provides extensive empirical evidence across two behaviors, five models, multiple baselines, and numerous ablations. The method's effectiveness is demonstrated consistently, with limitations clearly acknowledged. The central claim is supported by the evidence, and the weakest assumption identified by the reader (causal interpretation of activation contrast) is explicitly discussed and does not undermine the empirical results.
The paper is thorough, transparent, and empirically rigorous. The reader's concern about causality is acknowledged by the authors and does not undermine the central empirical claim that the method is effective. Therefore no change to the verdict is needed.
The referee's report
If the empirical results hold, the method provides a practical, training-free intervention for reducing compliance with specific undesirable instructions. Its strengths include (i) the use of only forward passes, requiring no gradients or labeled responses; (ii) the edit is stored in weights, so it is not circumvented by changing prompts; (iii) extensive evaluation across two behaviors, five model families, multiple baselines, and numerous ablations; (iv) the method complements prompting (the sexism edit plus fairness prompt beats either alone). The systematic ablations of the scoring rule (graded contrast vs. set difference, per-prompt vs. pooled norms) and the controlled format-transfer study are valuable contributions. The paper is transparent about limitations: the causal interpretation of activation contrast, variability across models, sensitivity to prompt-family selection, and costs to general capability and legitimate requests.
Pith's reading of the paper (a machine's summary, not the authors')
This paper introduces a method to reduce a language model's compliance with specific harmful instructions—such as orders to lie or produce sexist content—by editing a small fraction of its weights. The edit requires only forward passes on paired prompts contrasting the undesirable and desirable behaviors, and once applied, the change persists in the model's weights regardless of prompt modifications. On two benchmarks for instructed falsehood, the method improves honest response rates by 6.2 and 4.6 percentage points over the strongest alternative. On an instructed sexism task, it cuts sexist responses from 57.9% to 24.3%, and combining the edit with a fairness prompt further reduces the rate to 11.9%. The approach provides a practical, training-free intervention that works standalone or alongside prompt-based safeguards.
What carries the argument. The core machinery is the weighted activation-contrast score $s_{ij} = |W_{ij}| c_j$, where $c_j = v \bar{n}^F_j - (1-v) \bar{n}^T_j$ measures how much more active an input channel is under undesirable instructions ($F$) than under desirable ones ($T$). Weights with the highest scores in each matrix are rescaled by a factor $(1+\alpha)$, typically attenuating or reversing their contribution ($\alpha < 0$). The positive-contrast threshold $1/r$ (default $r=0.37$) ensures that only channels with activity more than $1/r$-fold higher under $F$ than under $T$ are selected.
Core claim. The paper's central claim is that a weighted activation-contrast scoring rule, which ranks each weight by its magnitude times the difference in mean input-channel activity between undesirable and desirable instruction conditions, identifies a small set of causally relevant weights. Scaling down or deleting these weights (the top 4% per matrix) reduces the model's compliance with the targeted undesirable instructions while leaving other capabilities largely intact. The rule is training-free and generalizes across two disparate behavioral domains (instructed falsehood and instructed sexism) and five model families, outperforming set-difference selection and single-layer activation steering.
The premise it rests on. The method assumes that the activation contrast $c_j$ identifies weights causally responsible for compliance with undesirable instructions, rather than merely correlating with superficial differences between the instruction formats.
If the paper is right (Pith's inferences, not the authors')
The edit provides a practical defense when the operator controls the model weights but cannot guarantee the system prompt, since the weight modification outlasts any prompt change.
Combined with a fairness prompt, the sexism edit reduces sexist responses to 11.9%, demonstrating that weight editing and prompt engineering can be stacked for greater effect.
The method transfers to single-turn prompt formats not used during construction, outperforming single-layer activation steering (mean honest rate 95.8% vs 83.4%) across four models.
The edit works on models up to 32B parameters, with gains in honesty over the unedited model increasing with model size (e.g., +75.5 points at Qwen2.5-32B on LIARS' BENCH).
Selectivity remains a challenge: the edit reduces compliance with legitimate falsehood requests (fiction, role-play) by 8–29 points on four models, motivating richer reference prompts.
The observation that the sexism mask (built without any lie prompts) also raises honesty by 37 points on Llama suggests the method removes shared circuitry for following orders to produce content the model would otherwise avoid, beyond the specific targeted behavior.
If the activation contrast truly identifies causal weights, contrast-based editing could be applied to other behavioral targets (e.g., sycophancy, toxicity) with minimal per-target cost, though the paper's limitations remind us that effectiveness on other targets remains to be evaluated.
The method's selectivity gap—reducing compliance with legitimate fiction and role-play—points toward a necessary extension: constructing reference prompts that separate harmful instructions from permissible requests for false statements (e.g., creative writing).
The multi-turn results (mixed transfer) suggest that instruction-following in conversational contexts engages different weight pathways than single-turn orders, and may require dedicated multi-turn construction pairs.
Check this review
Pith stored no record of this review's passes and took no fingerprint of it. Reviews made by code from 13 Nov 2026 on carry both.
How often reviews like this one are right
Reviews like this one: written by deepseek-v4-flash, where what was read was not recorded.
How often objections from reviews like this one are right: not yet measured
Reviews by people
This review was written by a machine. No person has reviewed this paper on Pith yet. A scientist listed on Pith can review it for a fee, and their review is published here whatever it concludes.
The paper
Abstract
Instruction-following language models can comply with requests that conflict with intended behaviors and safety alignment. We introduce a training-free weight-editing approach for reducing compliance with targeted undesirable instructions. Our method needs only forward passes on a small set of prompt pairs per target behavior, and because the edit is stored in the model's weights, changing or deleting instructions in the prompt cannot undo the weight modification. We evaluate the approach on two target harmful behaviors across five instruction-tuned models: instructed falsehood and instructed sexism. On two instructed-lying benchmarks, it improves mean honest response rates by 6.2 and 4.6 percentage points over the strongest alternative, significantly on three of the five models and within noise on the other two. In a controlled study of seven prompt formats, it maintains higher honest-response rates than single-layer activation steering when orders to lie arrive in single-turn formats not used to construct the edit. For instructed sexism, it reduces the mean sexist-response rate from 57.9% to 24.3%; combining the edit with a fairness prompt reduces it further to 11.9%. These results establish our contrast-based weight editing approach as an effective training-free intervention in two distinct behavioral domains, supporting its use both independently and alongside prompt-based safeguards.
The paper's own sentences, arranged, kept beside the review and separate from it. Each label is drawn from a closed vocabulary; the number beside it is the probability the classifier gave that label. No sentence in this record was written by a model.
abstract compiled · 2026-10-16
The authors' main claim, as the classifier picked it out 1.00
“These results establish our contrast-based weight editing approach as an effective training-free intervention in two distinct behavioral domains, supporting its use both independently and alongside prompt-based safeguards.”
How the main claim is stated: asserted 1.00
What the abstract itself states, as probabilities: claims first 0.05 · states limitation 0.40 · human participants 0.17 · controlled comparison 0.87 · randomized assignment 0.02 · states quantitative result 0.99 · compares with prior results 0.97 · mentions code or data availability 0.03
The abstract, sentence by sentence
This table labels the abstract only (arXiv listing abstract v1). The number is the probability the classifier gave the label.
| Sentence | Role | The authors' own work | States the main conclusion | Strength |
|---|---|---|---|---|
| Instruction-following language models can comply with requests that conflict with intended behaviors and safety alignment. | background 1.00 | 0.07 | 0.02 | asserted 0.57 |
| We introduce a training-free weight-editing approach for reducing compliance with targeted undesirable instructions. | contribution 0.96 | 0.88 | 0.10 | asserted 0.99 |
| Our method needs only forward passes on a small set of prompt pairs per target behavior, and because the edit is stored in the model's weights, changing or deleting instructions in the prompt cannot undo the weight modification. | method 0.98 | 0.73 | 0.07 | asserted 1.00 |
| We evaluate the approach on two target harmful behaviors across five instruction-tuned models: instructed falsehood and instructed sexism. | method 0.80 | 0.95 | 0.03 | asserted 0.90 |
| On two instructed-lying benchmarks, it improves mean honest response rates by 6.2 and 4.6 percentage points over the strongest alternative, significantly on three of the five models and within noise on the other two. | result 1.00 | 0.96 | 0.08 | asserted 1.00 |
| In a controlled study of seven prompt formats, it maintains higher honest-response rates than single-layer activation steering when orders to lie arrive in single-turn formats not used to construct the edit. | result 0.99 | 0.94 | 0.05 | asserted 1.00 |
| For instructed sexism, it reduces the mean sexist-response rate from 57.9% to 24.3%; combining the edit with a fairness prompt reduces it further to 11.9%. | result 1.00 | 0.96 | 0.08 | asserted 1.00 |
| These results establish our contrast-based weight editing approach as an effective training-free intervention in two distinct behavioral domains, supporting its use both independently and alongside prompt-based safeguards. | interpretation 1.00 | 0.88 | 0.94 | asserted 1.00 |
What kind of paper the classifier reports this is: method or tool 1.00 · evidence from observation or experiment 0.96
What was examined, and what was not
Sources read
- arXiv listing abstract v1
Not examined
- full text
- figures and tables
- references and citing sentences
- relations to other papers
Levels run
- P abstract
Not run
- S sections
- T body sentences
- C citations
- R cross-paper relations
How the labels were made
Compiled by classifier-1.13.0 via TypeSafe API, question bank abstract-p1-20260923, 2026-10-16: 1 request, 6,301 input tokens.
Every quoted sentence on this record is the authors'. Every label is from Pith's closed vocabulary, with the probability shown beside it. This record as JSON
Figures (12)
Showing 11 of the paper's 12 figures. Figures are the authors', reproduced from the paper.
References (82)
- [1] A Simple and Effective Pruning Approach for Large Language Models , author =. International Conference on Learning Representations (ICLR 2024) , year =. 2306.11695 , archivePrefix = on PithPith's reviewarXiv
- [2] Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications , author =. Proceedings of the 41st International Conference on Machine Learning (ICML 2024) , series =. 2024 , eprint = on Pith
- [3] Samuel Schapiro and Core Francisco Park and Felix Sosa and Lav R. Varshney , booktitle =. 2026 , eprint = on Pith
- [5] Improving Instruction-Following in Language Models through Activation Steering , author =. International Conference on Learning Representations (ICLR 2025) , year =. 2410.12877 , archivePrefix = on PithPith's reviewarXiv
- [6] Refusal in Language Models Is Mediated by a Single Direction , author =. Advances in Neural Information Processing Systems (NeurIPS 2024) , year =. 2406.11717 , archivePrefix = on PithPith's reviewarXiv
- [7] Steering Language Models With Activation Engineering , author =. arXiv preprint arXiv:2308.10248 , year =. 2308.10248 , archivePrefix = on PithPith's reviewarXiv
- [8] Andy Zou and Long Phan and Sarah Chen and James Campbell and Phillip Guo and Richard Ren and Alexander Pan and Xuwang Yin and Mantas Mazeika and Ann-Kathrin Dombrowski and Shashwat Goel and Nathaniel Li and Michael J. Byun and Zifan Wang and Alex Mallen and Steven Basart and S... on Pith
- [9] Editing Models with Task Arithmetic , author =. Proceedings of the 11th International Conference on Learning Representations (ICLR 2023) , year =. 2212.04089 , archivePrefix = on PithPith's reviewarXiv
- [10] Steering Language Models with Weight Arithmetic , author =. International Conference on Learning Representations (ICLR 2026) , year =. 2511.05408 , archivePrefix = on PitharXiv
- [11] Kevin Meng and David Bau and Alex Andonian and Yonatan Belinkov , booktitle =. Locating and Editing Factual Associations in. 2022 , eprint = on Pith
- [12] Namhoon Lee and Thalaiyasingam Ajanthan and Philip H. S. Torr , booktitle =. 2019 , eprint = on Pith
- [13] Andrew Lee and Xiaoyan Bai and Itamar Pres and Martin Wattenberg and Jonathan K. Kummerfeld and Rada Mihalcea , booktitle =. A Mechanistic Understanding of Alignment Algorithms: A Case Study on. 2024 , eprint = on Pith
- [14] The Lottery Ticket Hypothesis: Finding Sparse, Trainable Neural Networks , author =. International Conference on Learning Representations (ICLR 2019) , year =. 1803.03635 , archivePrefix = on PithPith's reviewarXiv
- [15] Pointer Sentinel Mixture Models , author =. International Conference on Learning Representations (ICLR 2017) , year =. 1609.07843 , archivePrefix = on PithPith's reviewarXiv
- [16] The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets , author =. Conference on Language Modeling (COLM) , year =. 2310.06824 , archivePrefix = on PithPith's reviewarXiv
- [20] Ren, Richard and Agarwal, Arunim and Mazeika, Mantas and Menghini, Cristina and Vacareanu, Robert and Kenstler, Brad and Yang, Mick and Barrass, Isabelle and Gatti, Alice and Yin, Xuwang and Trevino, Eduardo and Geralnik, Matias and Khoja, Adam and Lee, Dean and Yue, Summer an... on Pith
- [21] Safety Alignment Should Be Made More Than Just a Few Tokens Deep , author =. International Conference on Learning Representations (ICLR) , year =. 2406.05946 , archivePrefix = on PithPith's reviewarXiv
- [23] Understanding (Un)Reliability of Steering Vectors in Language Models , author =. ICLR 2025 Workshop on Foundation Models in the Wild , year =. 2505.22637 , archivePrefix = on PithPith's reviewarXiv
- [26] Measuring Massive Multitask Language Understanding , author =. International Conference on Learning Representations (ICLR) , year =. 2009.03300 , archivePrefix = on PithPith's reviewarXiv
- [27] Li, Nathaniel and Pan, Alexander and Gopal, Anjali and Yue, Summer and Berrios, Daniel and Gatti, Alice and Li, Justin D. and Dombrowski, Ann-Kathrin and Goel, Shashwat and Mukobi, Gabriel and Helm-Burger, Nathan and Lababidi, Rassin and Justen, Lennart and Liu, Andrew Bo and ... on Pith
show 62 more
- [32] Grattafiori, Aaron and Dubey, Abhimanyu and Jauhri, Abhinav and Pandey, Abhinav and Kadian, Abhishek and others , journal =. The. 2024 , url = on Pith
- [37] Discovering Latent Knowledge in Language Models Without Supervision , author =. International Conference on Learning Representations (ICLR) , year =. 2212.03827 , archivePrefix = on PithPith's reviewarXiv
- [38] Probabilities of Causation: Three Counterfactual Interpretations and Their Identification , author =. Synthese , volume =. 1999 , doi = on Pith
- [39] Probabilities of Causation: Bounds and Identification , author =. Annals of Mathematics and Artificial Intelligence , volume =. 2000 , doi = on Pith
- [40] Mingjie Sun and Zhuang Liu and Anna Bair and J. Zico Kolter , booktitle =. A. 2024 , eprint = on Pith
- [41] Boyi Wei and Kaixuan Huang and Yangsibo Huang and Tinghao Xie and Xiangyu Qi and Mengzhou Xia and Prateek Mittal and Mengdi Wang and Peter Henderson , booktitle =. Assessing the. 2024 , eprint = on Pith
- [42] Alessandro Stolfo and Vidhisha Balachandran and Safoora Yousefi and Eric Horvitz and Besmira Nushi , booktitle =. Improving. 2025 , eprint = on Pith
- [43] Andy Arditi and Oscar Obeso and Aaquib Syed and Daniel Paleka and Nina Panickssery and Wes Gurnee and Neel Nanda , booktitle =. Refusal in. 2024 , eprint = on Pith
- [44] Alexander Matt Turner and Lisa Thiergart and Gavin Leech and David Udell and Juan J. Vazquez and Ulisse Mini and Monte MacDiarmid , journal =. Steering. 2023 , eprint = on Pith
- [45] Andy Zou and Long Phan and Sarah Chen and James Campbell and Phillip Guo and Richard Ren and Alexander Pan and Xuwang Yin and Mantas Mazeika and Ann-Kathrin Dombrowski and Shashwat Goel and Nathaniel Li and Michael J. Byun and Zifan Wang and Alex Mallen and Steven Basart and S... on Pith
- [46] Gabriel Ilharco and Marco Tulio Ribeiro and Mitchell Wortsman and Suchin Gururangan and Ludwig Schmidt and Hannaneh Hajishirzi and Ali Farhadi , booktitle =. Editing. 2023 , eprint = on Pith
- [47] Constanza Fierro and Fabien Roger , booktitle =. Steering. 2026 , eprint = on Pith
- [48] Kevin Meng and David Bau and Alex Andonian and Yonatan Belinkov , booktitle =. Locating and. 2022 , eprint = on Pith
- [49] Andrew Lee and Xiaoyan Bai and Itamar Pres and Martin Wattenberg and Jonathan K. Kummerfeld and Rada Mihalcea , booktitle =. A. 2024 , eprint = on Pith
- [50] Jonathan Frankle and Michael Carbin , booktitle =. The. 2019 , eprint = on Pith
- [51] Stephen Merity and Caiming Xiong and James Bradbury and Richard Socher , booktitle =. Pointer. 2017 , eprint = on Pith
- [52] Samuel Marks and Max Tegmark , booktitle =. The. 2024 , eprint = on Pith
- [54] Lorenzo Pacchiardi and Alex J. Chan and Sören Mindermann and Ilan Moscovitz and Alexa Y. Pan and Yarin Gal and Owain Evans and Jan Brauner , booktitle =. How to. 2024 , eprint = on Pith
- [55] Kieron Kretschmar and Walter Laurito and Sharan Maiya and Samuel Marks , journal =. Liars'. 2025 , url =. 2511.16035 , archivePrefix = on PitharXiv
- [56] Richard Ren and Arunim Agarwal and Mantas Mazeika and Cristina Menghini and Robert Vacareanu and Brad Kenstler and Mick Yang and Isabelle Barrass and Alice Gatti and Xuwang Yin and Eduardo Trevino and Matias Geralnik and Adam Khoja and Dean Lee and Summer Yue and Dan Hendrycks... on PitharXiv
- [57] Xiangyu Qi and Ashwinee Panda and Kaifeng Lyu and Xiao Ma and Subhrajit Roy and Ahmad Beirami and Prateek Mittal and Peter Henderson , booktitle =. Safety. 2025 , eprint = on Pith
- [59] Joschka Braun and Carsten Eickhoff and David Krueger and Seyed Ali Bahrainian and Dmitrii Krasheninnikov , booktitle =. Understanding (. 2025 , eprint = on Pith
- [60] Matthew Nguyen and Kyle Cox and Austin Meek and Iván Arcuschin , journal =. On the. 2026 , url =. 2607.29062 , archivePrefix = on PithPith's reviewarXiv
- [61] Lucas Pinto , journal =. Activation. 2026 , url =. 2607.09156 , archivePrefix = on PithPith's reviewarXiv
- [62] Dan Hendrycks and Collin Burns and Steven Basart and Andy Zou and Mantas Mazeika and Dawn Song and Jacob Steinhardt , booktitle =. Measuring. 2021 , eprint = on Pith
- [63] Nathaniel Li and Alexander Pan and Anjali Gopal and Summer Yue and Daniel Berrios and Alice Gatti and Justin D. Li and Ann-Kathrin Dombrowski and Shashwat Goel and Gabriel Mukobi and Nathan Helm-Burger and Rassin Lababidi and Lennart Justen and Andrew Bo Liu and Michael Chen a... on Pith
- [65] Teun van der Weij and Felix Hofstätter and Ollie Jaffe and Samuel F. Brown and Francis Rhys Ward , booktitle =. 2025 , eprint = on Pith
- [67] Jérémy Scheurer and Mikita Balesni and Marius Hobbhahn , journal =. Large. 2023 , url =. 2311.07590 , archivePrefix = on PithPith's reviewarXiv
- [68] Aaron Grattafiori and Abhimanyu Dubey and Abhinav Jauhri and Abhinav Pandey and Abhishek Kadian and Ahmad Al-Dahle and Aiesha Letman and Akhil Mathur and Alan Schelten and Alex Vaughan and others , journal =. The. 2024 , url =. 2407.21783 , archivePrefix = on PithPith's reviewarXiv
- [69] Qwen2.5. arXiv preprint arXiv:2412.15115 , year =. 2412.15115 , archivePrefix = on PithPith's reviewarXiv
- [70] arXiv preprint arXiv:2503.01743 , year =. 2503.01743 , archivePrefix = on PithPith's reviewarXiv
- [72] Gemma 3. arXiv preprint arXiv:2503.19786 , year =. 2503.19786 , archivePrefix = on PithPith's reviewarXiv
- [73] Collin Burns and Haotian Ye and Dan Klein and Jacob Steinhardt , booktitle =. Discovering. 2023 , eprint = on Pith
- [74] Bryan R. Christ and Zack Gottesman and Jonathan Kropko and Thomas Hartvigsen , booktitle =. 2025 , eprint = on Pith
- [75] Boerner, Timothy J. and Deems, Stephen and Furlani, Thomas R. and Knuth, Shelley L. and Towns, John , title =. Practice and Experience in Advanced Research Computing (. 2023 , doi = on Pith
- [76] Generative Interpretability via Scalable Neuro-Symbolic Models , author =. 2026 , eprint = on Pith
- [77] Instruction-Following Evaluation for Large Language Models , author =. arXiv preprint arXiv:2311.07911 , year =. 2311.07911 , archivePrefix = on PithPith's reviewarXiv
- [78] Assessing the Brittleness of Safety Alignment via Pruning and Low-Rank Modifications , author =. Proceedings of the 41st International Conference on Machine Learning (ICML 2024) , year =. 2402.05162 , archivePrefix = on PithPith's reviewarXiv
- [79] CreativityNeuro: Steering Language Model Weights to Improve Divergent Thinking and Reduce Mode Collapse , author =. ICML 2026 Workshop on Creativity & Generative AI , year =. 2607.01433 , archivePrefix = on PithPith's reviewarXiv
- [80] Steering Llama 2 via Contrastive Activation Addition , author =. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) , year =. 2312.06681 , archivePrefix = on PithPith's reviewarXiv
- [81] Steering Language Models With Activation Engineering , author =. 2024 , eprint = on Pith
- [82] Representation Engineering: A Top-Down Approach to AI Transparency , author =. 2025 , eprint = on Pith
- [83] Locating and Editing Factual Associations in GPT , author =. Advances in Neural Information Processing Systems (NeurIPS 2022) , year =. 2202.05262 , archivePrefix = on PithPith's reviewarXiv
- [84] SNIP: Single-shot Network Pruning based on Connection Sensitivity , author =. International Conference on Learning Representations (ICLR 2019) , year =. 1810.02340 , archivePrefix = on PithPith's reviewarXiv
- [85] A Mechanistic Understanding of Alignment Algorithms: A Case Study on DPO and Toxicity , author =. Proceedings of the 41st International Conference on Machine Learning (ICML 2024) , year =. 2401.01967 , archivePrefix = on PithPith's reviewarXiv
- [86] The Internal State of an LLM Knows When It's Lying , author =. Findings of the Association for Computational Linguistics: EMNLP 2023 , year =. 2304.13734 , archivePrefix = on PithPith's reviewarXiv
- [87] How to Catch an AI Liar: Lie Detection in Black-Box LLMs by Asking Unrelated Questions , author =. International Conference on Learning Representations (ICLR) , year =. 2309.15840 , archivePrefix = on PithPith's reviewarXiv
- [88] Liars' Bench: Evaluating Lie Detectors for Language Models , author =. 2026 , eprint = on Pith
- [89] The MASK Benchmark: Disentangling Honesty From Accuracy in AI Systems , author =. 2026 , eprint = on Pith
- [90] Analyzing the Generalization and Reliability of Steering Vectors , author =. Advances in Neural Information Processing Systems 37 (NeurIPS 2024) , year =. 2407.12404 , archivePrefix = on PithPith's reviewarXiv
- [91] On the Generalization of Steering Vectors for Chain-of-Thought Faithfulness , author =. 2026 , eprint = on Pith
- [92] Activation Steering Transfer to Agents: One Gain Ratio Does Not Identify Potency and Efficacy , author =. 2026 , eprint = on Pith
- [93] The WMDP Benchmark: Measuring and Reducing Malicious Use With Unlearning , author =. Proceedings of the 41st International Conference on Machine Learning (ICML) , year =. 2403.03218 , archivePrefix = on PithPith's reviewarXiv
- [94] Not what you've signed up for: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection , author =. Proceedings of the 16th ACM Workshop on Artificial Intelligence and Security (AISec '23) , year =. 2302.12173 , archivePrefix = on PithPith's reviewarXiv
- [95] AI Sandbagging: Language Models can Strategically Underperform on Evaluations , author =. International Conference on Learning Representations (ICLR) , year =. 2406.07358 , archivePrefix = on PithPith's reviewarXiv
- [96] AI Deception: A Survey of Examples, Risks, and Potential Solutions , author =. Patterns , year =. 2308.14752 , archivePrefix = on PithPith's reviewarXiv
- [97] Large Language Models can Strategically Deceive their Users when Put Under Pressure , author =. 2024 , eprint = on Pith
- [98] The Llama 3 Herd of Models , author =. 2024 , eprint = on Pith
- [99] Qwen2.5 Technical Report , author =. 2025 , eprint = on Pith
- [100] Phi-4-Mini Technical Report: Compact yet Powerful Multimodal Language Models via Mixture-of-LoRAs , author =. 2025 , eprint = on Pith
- [101] 2 OLMo 2 Furious , author =. Conference on Language Modeling (COLM) , year =. 2501.00656 , archivePrefix = on PithPith's reviewarXiv
- [102] Gemma 3 Technical Report , author =. 2025 , eprint = on Pith
Cited by 1 Pith paper
Reviewed papers on Pith that cite this one. The full citation record
Challenges and discussion
Disagree with the review, or know something it missed? Leave a comment, or challenge this review formally: a challenge becomes a permanent entry in this paper's record, and its decision is published here.
Come back by email
A short list of the week's newly reviewed papers. Nothing is sent until you confirm from your inbox, and every email has a link that stops it. No account.
No comments yet.
Sign in to comment