Pith. sign in

REVIEW 5 major objections 5 minor 57 references

Enhancing Cross-task Transfer of Large Language Models via Activation Steering

T0 review · 5 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 The paper argues that LLMs can transfer task skills across domains and languages by adding a single precomputed activation-difference vector to the model's hidden state, with no training and no extra input tokens.

desk verdict A plausible cross-task extension of in-context vectors with consistent gains, but the core parallelism claim rests on t-SNE alone and the evaluation lacks error bars. read the letter →

arxiv 2507.13236 v1 pith:PPW67BLW submitted 2025-07-17 cs.CL

Xinyu Tang , Zhihao Lv , Xiaoxue Cheng , Junyi Li , Wayne Xin Zhao , Zujie Wen , Zhiqiang Zhang , Jun Zhou This is my paper · ORCID
classification cs.CL
keywords activationsteeringcross-tasktransferin-contextlearninglow-resourcetaskslatentspacelargelanguagemodelscross-lingualsampleselection
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

Large language models can transfer task-relevant behavior from high-resource to low-resource tasks by having a precomputed vector added to their internal activations, with no parameter updates and no extra input tokens. The paper proposes CAST, which selects influential and diverse examples from a source task, computes the average difference between the model's few-shot and zero-shot activations at the final token, and injects that difference into the forward pass of target queries. Across cross-domain and cross-lingual benchmarks, CAST consistently beats zero-shot, random few-shot, TopK, and DPP selection baselines, and unlike in-context learning, its gains keep increasing as more source examples are used. The practical appeal is that the method runs at the same compute cost as zero-shot prompting while avoiding the context-length and demonstration-sensitivity problems of in-context learning.

What carries the argument

The central object is the contrastive representation-enhanced activation vector $C_L$, defined as the average, over a selected subset of source samples, of the difference between the model's final-token hidden-state activation on a few-shot prompt and on the corresponding zero-shot prompt: $C_L = \frac{1}{n}\sum_i (A_L(f_i) - A_L(z_i))$. The steering update is $\hat{h}_L^{-1} = h_L^{-1} + \lambda C_L$, which adds this precomputed vector to the final token's hidden state at layer $L$ during the forward pass of a low-resource query. The mechanism's justification is the paper's empirical claim that the few-shot-minus-zero-shot difference directions are nearly parallel across tasks, so one source-task vector transfers to another task. The subset selection component, built on a similarity graph with influence and diversity scoring, determines which source samples contribute to $C_L$.

What would settle it

Compute the average cosine similarity between $C_L$ vectors from source and target tasks at the chosen injection layer over the benchmark pairs reported in the paper; if the similarities are near zero or negative on pairs where CAST still improves, then the cross-task parallelism hypothesis is not the working mechanism and the gains would need another explanation.

Watch

Extended reading notes

Core claim

The paper's central discovery is that the internal activation shift caused by in-context examples is largely task-independent: the difference between few-shot and zero-shot final-token activations points in a consistent direction across domains and languages. Building on that, CAST forms a single steering vector $C_L$ on a high-resource task and injects it into target-task forward passes via $\hat{h}_L^{-1} = h_L^{-1} + \lambda C_L$, with $\lambda=1$ in the main experiments. The reported result is consistent improvement over zero-shot and few-shot baselines in every source-target pair tested across five cross-domain targets and six cross-lingual targets, including large gains on distant pairs such as an average of 89.04% for English versus 64.48% for the DPP in-context baseline.

Load-bearing premise

The load-bearing premise is that the few-shot-minus-zero-shot activation difference vectors are nearly parallel across different tasks, so a direction averaged over high-resource samples remains a valid steering direction for a low-resource task; the paper supports this with t-SNE projections rather than a quantitative similarity measure.

Editorial extensions

If this is right

  • Cross-task transfer can be served as a lookup: once $C_L$ is extracted from a source task, applying it to a target task costs no parameter updates and no additional input tokens.
  • Inference-time complexity matches zero-shot prompting, not few-shot prompting, because the steering vector adds nothing to the input length.
  • Performance continues to improve as the number of source examples grows, whereas few-shot in-context learning plateaus or declines when demonstrations accumulate.
  • The same steering procedure works across distant language pairs such as German to English and English to Chinese, where cross-task demonstrations are noisy or inconsistent.
  • Injecting into an intermediate layer works better than shallow or final layers, consistent with the paper's entropy analysis that intermediate layers carry the richest representations.

Reading between the lines

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

  • A natural next step is to quantify the parallelism directly: compute the cosine similarity between $C_L$ vectors from different tasks and test whether that similarity predicts the size of the transfer gain, turning the steering direction into a principled source-task selection criterion.
  • If the parallelism claim holds, steering vectors become a form of task arithmetic: adding $C_L$ from one task and subtracting another might compose or suppress behaviors in ways the paper does not explore.
  • The generation-task results in the appendix suggest steering can help where cross-task demonstrations hurt, so one could test whether injecting at intermediate reasoning tokens rather than only the final token extends the method to long chain-of-thought generation.
  • Because the paper notes that CAST requires white-box access to hidden states, a practical follow-up is to approximate the steering vector from a black-box API by probing output distributions, which would preserve the no-training benefit for closed models.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 5 minor

Summary. The manuscript proposes CAST (Cross-task Activation Steering Transfer), a training-free method that improves LLM performance on low-resource target tasks by injecting a precomputed steering vector into the hidden states of the final token (Eq. 9). The steering vector C_L is the average difference between few-shot and zero-shot activations, computed from an influential and diverse subset of high-resource source-task examples (Eq. 8). The paper argues that few-shot and zero-shot activation differences are 'nearly parallel' across tasks (Section 3.2, Figure 2), and that this justifies cross-task transfer in latent space. Experiments cover cross-domain (Table 1) and cross-lingual (Table 2) settings with Llama3.1 and Qwen models, plus scaling, efficiency, ablation, and generation-task analyses. The method consistently outperforms zero-shot, Few-shot Random, TopK, and DPP baselines in the reported tables.

Significance. If the reported results hold, CAST offers a practical and scalable alternative to in-context learning for cross-task transfer, with no parameter updates and no input expansion—this is a meaningful contribution given the context-window and cost limitations of few-shot prompting. The empirical scope is a clear strength: seven source domains, five target domains, six languages, multiple model families, a scalability analysis up to 1000 examples, and an ablation study. The paper also includes a self-contained limitation section (Appendix C). However, the central mechanistic claim—that cross-task transfer works because difference vectors are nearly parallel across tasks—is supported only by t-SNE visualizations, which is insufficient. The lack of quantitative parallelism evidence, missing control conditions, and reliance on single-run numbers leave the conceptual contribution under-supported even though the empirical gains are consistent and sometimes large.

major comments (5)
  1. [Sec. 3.2, Fig. 2] The paper's central premise—that the few-shot minus zero-shot activation difference vectors are 'nearly parallel' across tasks—is supported only by the t-SNE visualizations in Figure 2. t-SNE does not preserve global distances or vector directions, so it cannot establish this claim. Please provide a quantitative measure, e.g., the mean pairwise cosine similarity between the C_L vectors (Eq. 8) computed from different source tasks, and/or the distribution of per-sample difference-vector angles across tasks. This is load-bearing because Section 4.3 transfers C_L from source to target tasks; if the vectors are not nearly parallel, the mechanism claimed in Sections 1 and 3.2 is not supported. The same evidence is missing for the cross-lingual setting, where the method is also applied (Table 2).
  2. [Tables 1–2, Sec. 5.1] All accuracy numbers in Tables 1 and 2 are single-run values with no variance estimates or significance tests. Some of the reported gaps are small (e.g., Table 1, MedMCQA with QQP source: CAST 58.40 vs. DPP 56.20; Table 4, Social-i-QA with Qwen2.5-7B-Instruct: CAST 77.66 vs. DPP 77.60). Without multiple seeds and standard deviations, the claim that CAST 'consistently outperforms' baselines is not fully established. Please report mean and std over at least 3–5 runs, or state the evaluation protocol if deterministic.
  3. [Sec. 5.1, Eq. (9)] The injection layer L is selected based on validation-set performance for each target task, while the few-shot baselines have no analogous per-target hyperparameter. This asymmetric selection may inflate CAST's gains. Please (a) report the chosen layer for each target task, (b) show the sensitivity of the results to the layer (Figure 8a already does this for one pair), and (c) either fix L a priori or apply an analogous validation-based selection to the baselines. The 'training-free' claim is not affected, but the comparison fairness is.
  4. [Sec. 4.3, Table 1] There is no control condition that distinguishes task-specific transfer from a generic ICL enhancement effect. If any steering vector with a similar norm improves performance regardless of the source task, then the improvement would not demonstrate cross-task transfer of task-specific knowledge. Please include controls such as injecting a random vector of matched norm, or a C_L computed from a source task deliberately unrelated to the target (e.g., SST2 → MedMCQA), and compare against the task-matched C_L. The variability across source tasks in Table 1 suggests some task-specificity, but a direct control is needed.
  5. [Sec. 5.1 vs. Sec. 4.2, Eq. (5)–(6), Algorithm 1] The hyperparameter names are inconsistent: Section 5.1 says 'β controlling the sparsity of the graph' and 'hop-based decay factor α', but in Section 4.2.1 the sparsity parameter is α (Eq. 5) and in Section 4.2.2 the decay factor is β (Eq. 6); Algorithm 1 uses α for the decay factor in the diversity loop. The stated values (β=20, α=0.2) are also unintuitive given Eq. (5). Please unify the notation and report the exact values used for each hyperparameter to ensure reproducibility.
minor comments (5)
  1. [Sec. 3.1] There is a typo in Section 3.1: 'comoputed' should be 'computed'.
  2. [Fig. 2 caption] The Figure 2 caption should specify which clusters correspond to zero-shot vs. few-shot and which to source vs. target; the current caption only says 'green for source task, purple for target task' and is ambiguous.
  3. [Sec. 5.3.2] In Section 5.3.2, 'Fugre 4b' should be 'Figure 4b'.
  4. [Algorithm 1 and Algorithm 2] Algorithm 1 contains typos: 'neighibor' should be 'neighbor', and 'Influencial' should be 'Influential'.
  5. [Appendix D] The case studies are presented without commentary; a short discussion of when steering helps or hurts (and why) would make the case-study appendix more informative.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the steering vector is computed from source-task activations and evaluated on target tasks, with only standard validation-based layer selection.

full rationale

The derivation chain is self-contained. CAST computes the steering vector C_L (Eq. 8) as the average of few-shot minus zero-shot activations over selected high-resource samples, then applies it at inference to the target hidden state (Eq. 9). The target evaluation does not feed back into C_L or into the subset-selection procedure; the only target-dependent choice is the injection layer, which is selected on a validation set (Sec. 5.1) and is standard hyperparameter tuning rather than an input to the derivation. The paper's load-bearing empirical premise, that difference vectors are nearly parallel across tasks (Sec. 3.2), is supported only by t-SNE plots, which do not preserve vector differences; that is an evidence-quality weakness, not circularity, because the premise is not defined in terms of the predicted outputs and the method remains an honest empirical test of that premise. Citations to the authors' prior work ([30] on reasoning and [9] on prompt transfer) appear only in related work and are not used to justify the CAST mechanism; the injection procedure explicitly follows the external ICML 2024 in-context vectors work [28]. No equation reduces to its own input by construction, and no fitted parameter is renamed as a prediction. Therefore no circular step is exhibited.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The method introduces seven hyperparameters, several selected on the validation set, and one unreported value (k). The core transfer assumption, parallel activation differences across tasks, is an empirical premise evidenced only qualitatively. No new entities are postulated.

free parameters (8)
  • Injection strength lambda = 1
    Set based on Figure 8b hyperparameter analysis; controls magnitude of added activation vector.
  • Injection layer L = determined per target via validation set
    Section 5.1: 'The injection layer is determined based on performance on the validation set', a selection on target data.
  • Subset size n = 20
    Section 5.1 implementation details; number of selected source examples.
  • Graph sparsity (called beta in implementation, alpha in Eq. 5) = 20
    Section 5.1 lists beta=20 controlling graph sparsity, while Eq. (5) writes k_i = ceil(alpha * s_i * (|V|-1)). Naming inconsistent.
  • Hop decay factor (called alpha in implementation, beta in Eq. 6) = 0.2
    Section 5.1 lists alpha=0.2, while Eq. (6) uses beta^i. Algorithm 1 also uses alpha^i.
  • Trade-off parameter gamma = 0.2
    Section 5.1; balances influence and diversity in Eq. (7).
  • Neighborhood depth k = not reported
    Used in Eq. (6) and Algorithm 1 for diversity penalty; no value given in implementation details.
  • Diffusion simulations = 20
    Section 4.2.2: 20 repeats for influence score averaging.
assumptions (5)
  • domain assumption The final token hidden state aggregates the most contextually salient information in LLMs.
    Used to extract activations in Sections 3.1 and 4.3; standard in activation steering but an assumption about representation.
  • domain assumption Few-shot minus zero-shot activation difference vectors are nearly parallel across tasks.
    Central premise for cross-task transfer; evidence is t-SNE visualization in Section 3.2, no quantitative validation.
  • standard math Matrix-based entropy with alpha to 1 quantifies feature diversity and redundancy as claimed.
    Defined in Eq. (1) following prior work [33,34,35]; accepted in prior literature.
  • domain assumption Influence scores stay fixed during the iterative greedy selection.
    Algorithm 2 reuses precomputed I(v); the paper asserts this is efficient but does not justify that influence is selection-independent.
  • domain assumption BGE embeddings provide adequate similarity for graph construction.
    Graph construction in Section 4.2.1 uses BGE [37]; quality of transfer depends on this embedding space.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Enhancing Cross-task Transfer of Large Language Models via Activation Steering." pith.science (2026). https://pith.science/paper/PPW67BLW

@misc{pith2026250713236,
  author       = {Pith},
  title        = {Pith review of: Enhancing Cross-task Transfer of Large Language Models via Activation Steering},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/PPW67BLW}},
  note         = {Machine review of arXiv:2507.13236}
}
read the original abstract

Large language models (LLMs) have shown impressive abilities in leveraging pretrained knowledge through prompting, but they often struggle with unseen tasks, particularly in data-scarce scenarios. While cross-task in-context learning offers a direct solution for transferring knowledge across tasks, it still faces critical challenges in terms of robustness, scalability, and efficiency. In this paper, we investigate whether cross-task transfer can be achieved via latent space steering without parameter updates or input expansion. Through an analysis of activation patterns in the latent space of LLMs, we observe that the enhanced activations induced by in-context examples have consistent patterns across different tasks. Inspired by these findings, we propose CAST, a novel Cross-task Activation Steering Transfer framework that enables effective transfer by manipulating the model's internal activation states. Our approach first selects influential and diverse samples from high-resource tasks, then utilizes their contrastive representation-enhanced activations to adapt LLMs to low-resource tasks. Extensive experiments across both cross-domain and cross-lingual transfer settings show that our method outperforms competitive baselines and demonstrates superior scalability and lower computational costs.

Figures

Figures reproduced from arXiv: 2507.13236 by the authors.

Figure 2
Figure 2. t-SNE cluster across domains (green for source task, purple for target task) ality modeling [24], instruction following [25], hallucination mitigation [26, 27], safety enhance￾ment [28], and reasoning improvement [29, 30]. In this work, we leverage the activation steering method to transfer knowledge from the data-sufficient tasks to address data-scarce tasks. 3 Empirical Study The Hopfieldian view of cognition [31]… view at source ↗
Figure 3
Figure 3. The overview of CAST. (a) An illustration of three phases in our influential and diverse subset selection strategy. (1) We first construct a similarity-based sample graph of the high-resource task. (2) Next, we quantify each sample’s influence and diversity. (3) Then, we use an iterative greedy graph search algorithm to choose the highest-scoring sample at each step, and finally form the subset of the high-resource … view at source ↗
Figure 4
Figure 4. Detailed analysis of our proposed method. [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figures from the paper (4 more)
Figure 5
Figure 5. Figure 5: Different Model Sizes 1 2 5 10 20 50 100 200 500 1000 Number of Examples 50 52 54 56 58 Accuracy Zero-shot Few-shot Ours (a) Source Task: MNLI, Target Task: MedMCQA 1 2 5 10 20 50 100 200 500 1000 Number of Examples 55 58 61 64 67 Accuracy Zero-shot Few-shot Ours (b) S…
Figure 6
Figure 6. Figure 6: Scalability For the influential and diverse subset selection method, we explore the parameter β controlling the graph sparsity, the hop-based decay factor α, and the trade-off parameter γ between diversity and influence. The results are shown in Figure 7a, Figure 7b, F…
Figure 7
Figure 7. Figure 7: Hyperparameter analysis of the subset selection method. [PITH_FULL_IMAGE:figures/full_fig_p018_7.png]
Figure 8
Figure 8. Figure 8: Hyperparameter analysis of the activation selection approach. [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

57 extracted references · 42 canonical work pages

  1. [1]

    A survey of large language models

    Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Beichen Zhang, Junjie Zhang, Zican Dong, Yifan Du, Chen Yang, Yushuo Chen, Zhipeng Chen, Jinhao Jiang, Ruiyang Ren, Yifan Li, Xinyu Tang, Zikang Liu, Peiyu Liu, Jian-Yun Nie, and Ji-Rong Wen. A survey of large language models. CoRR, abs/2303.18223, 2023

    arXiv 2023

  2. [2]

    Learning to reason with llms, 2024

    OpenAI. Learning to reason with llms, 2024

    2024

  3. [3]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning. arXiv preprint arXiv:2501.12948, 2025

    arXiv 2025

  4. [4]

    The Inherent Limits of Pretrained LLMs: The Unexpected Convergence of Instruction Tuning and In-Context Learning Capabilities

    Irina Bigoulaeva, Harish Tayyar Madabushi, and Iryna Gurevych. The inherent limits of pretrained llms: The unexpected convergence of instruction tuning and in-context learning capabilities. CoRR, abs/2501.08716, 2025

    work page Pith review arXiv 2025

  5. [5]

    A comprehensive survey on transfer learning

    Fuzhen Zhuang, Zhiyuan Qi, Keyu Duan, Dongbo Xi, Yongchun Zhu, Hengshu Zhu, Hui Xiong, and Qing He. A comprehensive survey on transfer learning. Proc. IEEE, 109(1):43–76, 2021

    work page 2021

  6. [6]

    Franke, Ivo Rapant, Fabio Ferreira, and Frank Hutter

    Tobias Strangmann, Lennart Purucker, Jörg K.H. Franke, Ivo Rapant, Fabio Ferreira, and Frank Hutter. Transfer learning for finetuning large language models. In Adaptive Foundation Models: Evolving AI for Personalized and Efficient Learning, 2024

    work page 2024

  7. [7]

    A transfer learning framework for weak to strong generalization

    Seamus Somerstep, Felipe Maia Polo, Moulinath Banerjee, Yaacov Ritov, Mikhail Yurochkin, and Yuekai Sun. A transfer learning framework for weak to strong generalization. In The Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  8. [8]

    Spot: Better frozen model adaptation through soft prompt transfer

    Tu Vu, Brian Lester, Noah Constant, Rami Al-Rfou’, and Daniel Cer. Spot: Better frozen model adaptation through soft prompt transfer. In ACL (1), pages 5039–5059. Association for Computational Linguistics, 2022

    work page 2022

Show all 57 references
  1. [9]

    Learning to transfer prompts for text generation

    Junyi Li, Tianyi Tang, Jian-Yun Nie, Ji-Rong Wen, and Xin Zhao. Learning to transfer prompts for text generation. In NAACL-HLT, pages 3506–3518. Association for Computational Linguistics, 2022

    2022

  2. [10]

    Hyperlora: Efficient cross-task generalization via constrained low-rank adapters genera- tion

    Chuancheng Lv, Lei Li, Shitou Zhang, Gang Chen, Fanchao Qi, Ningyu Zhang, and Hai-Tao Zheng. Hyperlora: Efficient cross-task generalization via constrained low-rank adapters genera- tion. In EMNLP (Findings), pages 16376–16393. Association for Computational Linguistics, 2024

    2024

  3. [11]

    Soft language prompts for language transfer

    Ivan Vykopal, Simon Ostermann, and Marián Simko. Soft language prompts for language transfer. CoRR, abs/2407.02317, 2024

    2024 arXiv

  4. [12]

    Multilingual llms are better cross-lingual in-context learners with alignment

    Eshaan Tanwar, Subhabrata Dutta, Manish Borthakur, and Tanmoy Chakraborty. Multilingual llms are better cross-lingual in-context learners with alignment. In ACL (1), pages 6292–6307. Association for Computational Linguistics, 2023

    2023

  5. [13]

    Llms are few-shot in-context low- resource language learners

    Samuel Cahyawijaya, Holy Lovenia, and Pascale Fung. Llms are few-shot in-context low- resource language learners. In NAACL-HLT, pages 405–433. Association for Computational Linguistics, 2024

    2024

  6. [14]

    From classification to generation: Insights into crosslingual retrieval augmented ICL

    Xiaoqian Li, Ercong Nie, and Sheng Liang. From classification to generation: Insights into crosslingual retrieval augmented ICL. CoRR, abs/2311.06595, 2023

    2023 arXiv

  7. [15]

    Language models can exploit cross-task in-context learning for data-scarce novel tasks

    Anwoy Chatterjee, Eshaan Tanwar, Subhabrata Dutta, and Tanmoy Chakraborty. Language models can exploit cross-task in-context learning for data-scarce novel tasks. In ACL (1), pages 11568–11587. Association for Computational Linguistics, 2024. 10

    2024

  8. [17]

    Diverse demonstrations improve in-context compositional generalization

    Itay Levy, Ben Bogin, and Jonathan Berant. Diverse demonstrations improve in-context compositional generalization. In ACL (1), pages 1401–1422. Association for Computational Linguistics, 2023

    2023

  9. [18]

    Complementary explanations for effective in-context learning

    Xi Ye, Srinivasan Iyer, Asli Celikyilmaz, Veselin Stoyanov, Greg Durrett, and Ramakanth Pasunuru. Complementary explanations for effective in-context learning. In ACL (Findings), pages 4469–4484. Association for Computational Linguistics, 2023

    2023

  10. [19]

    Smith, Daniel Khashabi, and Hannaneh Hajishirzi

    Yizhong Wang, Yeganeh Kordi, Swaroop Mishra, Alisa Liu, Noah A. Smith, Daniel Khashabi, and Hannaneh Hajishirzi. Self-instruct: Aligning language models with self-generated instruc- tions. In ACL (1), pages 13484–13508. Association for Computational Linguistics, 2023

    2023

  11. [20]

    Cross-task gen- eralization via natural language crowdsourcing instructions

    Swaroop Mishra, Daniel Khashabi, Chitta Baral, and Hannaneh Hajishirzi. Cross-task gen- eralization via natural language crowdsourcing instructions. In ACL (1), pages 3470–3487. Association for Computational Linguistics, 2022

    2022

  12. [21]

    Continuous transfer learning with label-informed distribution alignment

    Jun Wu and Jingrui He. Continuous transfer learning with label-informed distribution alignment. CoRR, abs/2006.03230, 2020

    2006 arXiv

  13. [22]

    Continual transfer learning for cross-domain click-through rate prediction at taobao

    Lixin Liu, Yanling Wang, Tianming Wang, Dong Guan, Jiawei Wu, Jingxu Chen, Rong Xiao, Wenxiang Zhu, and Fei Fang. Continual transfer learning for cross-domain click-through rate prediction at taobao. In WWW (Companion Volume), pages 346–350. ACM, 2023

    2023

  14. [23]

    Aligning large language models with human preferences through representation engineering

    Wenhao Liu, Xiaohua Wang, Muling Wu, Tianlong Li, Changze Lv, Zixuan Ling, Jianhao Zhu, Cenyuan Zhang, Xiaoqing Zheng, and Xuanjing Huang. Aligning large language models with human preferences through representation engineering. In ACL (1), pages 10619–10638. Association for C...

    2024

  15. [24]

    Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization

    Yuanpu Cao, Tianrong Zhang, Bochuan Cao, Ziyi Yin, Lu Lin, Fenglong Ma, and Jinghui Chen. Personalized steering of large language models: Versatile steering vectors through bi-directional preference optimization. In NeurIPS, 2024

    2024

  16. [25]

    Improving instruction-following in language models through activation steering

    Alessandro Stolfo, Vidhisha Balachandran, Safoora Yousefi, Eric Horvitz, and Besmira Nushi. Improving instruction-following in language models through activation steering. In The Thir- teenth International Conference on Learning Representations, 2025

    2025

  17. [26]

    Viégas, Hanspeter Pfister, and Martin Wattenberg

    Kenneth Li, Oam Patel, Fernanda B. Viégas, Hanspeter Pfister, and Martin Wattenberg. Inference-time intervention: Eliciting truthful answers from a language model. In NeurIPS, 2023

    2023

  18. [27]

    Refusal in language models is mediated by a single direction

    Andy Arditi, Oscar Obeso, Aaquib Syed, Daniel Paleka, Nina Panickssery, Wes Gurnee, and Neel Nanda. Refusal in language models is mediated by a single direction. In NeurIPS, 2024

    2024

  19. [28]

    Sheng Liu, Haotian Ye, Lei Xing, and James Y . Zou. In-context vectors: Making in context learning more effective and controllable through latent space steering. InICML. OpenReview.net, 2024

    2024

  20. [29]

    Improving reasoning performance in large language models via representation engineering

    Bertram Højer, Oliver Simon Jarvis, and Stefan Heinrich. Improving reasoning performance in large language models via representation engineering. In The Thirteenth International Conference on Learning Representations, 2025

    2025

  21. [30]

    Unlocking general long chain-of-thought reasoning capabilities of large language models via representation engineering

    Xinyu Tang, Xiaolei Wang, Zhihao Lv, Yingqian Min, Wayne Xin Zhao, Binbin Hu, Ziqi Liu, and Zhiqiang Zhang. Unlocking general long chain-of-thought reasoning capabilities of large language models via representation engineering. CoRR, abs/2503.11314, 2025

    2025 arXiv

  22. [31]

    Neural networks and physical systems with emergent collective computational abilities

    John J Hopfield. Neural networks and physical systems with emergent collective computational abilities. Proceedings of the national academy of sciences, 79(8):2554–2558, 1982. 11

    1982

  23. [32]

    Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, J

    Andy Zou, Long Phan, Sarah Chen, James Campbell, Phillip Guo, Richard Ren, Alexander Pan, Xuwang Yin, Mantas Mazeika, Ann-Kathrin Dombrowski, Shashwat Goel, Nathaniel Li, Michael J. Byun, Zifan Wang, Alex Mallen, Steven Basart, Sanmi Koyejo, Dawn Song, Matt Fredrikson, J. Zico...

    arXiv

  24. [33]

    Brockmeier, and Luis Gonzalo Sánchez Giraldo

    Oscar Skean, Jhoan Keider Hoyos-Osorio, Austin J. Brockmeier, and Luis Gonzalo Sánchez Giraldo. Dime: Maximizing mutual information by a difference of matrix-based entropies. CoRR, abs/2301.08164, 2023

    2023 arXiv

  25. [34]

    Diff-erank: A novel rank-based metric for evaluating large language models

    Lai Wei, Zhiquan Tan, Chenghai Li, Jindong Wang, and Weiran Huang. Diff-erank: A novel rank-based metric for evaluating large language models. In NeurIPS, 2024

    2024

  26. [35]

    Layer by layer: Uncovering hidden representations in language models

    Oscar Skean, Md Rifat Arefin, Dan Zhao, Niket Patel, Jalal Naghiyev, Yann LeCun, and Ravid Shwartz-Ziv. Layer by layer: Uncovering hidden representations in language models. CoRR, abs/2502.02013, 2025

    2025 arXiv

  27. [36]

    Visualizing data using t-sne

    Laurens van der Maaten and Geoffrey Hinton. Visualizing data using t-sne. Journal of Machine Learning Research, 9(86):2579–2605, 2008

    2008

  28. [37]

    BGE m3- embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation

    Jianlv Chen, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. BGE m3- embedding: Multi-lingual, multi-functionality, multi-granularity text embeddings through self-knowledge distillation. CoRR, abs/2402.03216, 2024

    2024 arXiv

  29. [38]

    Think you have solved question answering? try arc, the AI2 reasoning challenge

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the AI2 reasoning challenge. CoRR, abs/1803.05457, 2018

    2018 arXiv

  30. [39]

    Character-level convolutional networks for text classification

    Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. Character-level convolutional networks for text classification. In NIPS, pages 649–657, 2015

    2015

  31. [40]

    Boolq: Exploring the surprising difficulty of natural yes/no questions

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. In NAACL-HLT (1), pages 2924–2936. Association for Computational Linguistics, 2019

    2019

  32. [41]

    Commonsenseqa: A question answering challenge targeting commonsense knowledge

    Alon Talmor, Jonathan Herzig, Nicholas Lourie, and Jonathan Berant. Commonsenseqa: A question answering challenge targeting commonsense knowledge. In NAACL-HLT (1), pages 4149–4158. Association for Computational Linguistics, 2019

    2019

  33. [42]

    Adina Williams, Nikita Nangia, and Samuel R. Bowman. A broad-coverage challenge corpus for sentence understanding through inference. In NAACL-HLT, pages 1112–1122. Association for Computational Linguistics, 2018

    2018

  34. [43]

    Natural language understand- ing with the quora question pairs dataset

    Lakshay Sharma, Laura Graesser, Nikita Nangia, and Utku Evci. Natural language understand- ing with the quora question pairs dataset. CoRR, abs/1907.01041, 2019

    1907 arXiv

  35. [44]

    Manning, Andrew Y

    Richard Socher, Alex Perelygin, Jean Wu, Jason Chuang, Christopher D. Manning, Andrew Y . Ng, and Christopher Potts. Recursive deep models for semantic compositionality over a sentiment treebank. In EMNLP, pages 1631–1642. ACL, 2013

    2013

  36. [45]

    Korhonen, Jyrki Wallenius, and Pyry Takala

    Pekka Malo, Ankur Sinha, Pekka J. Korhonen, Jyrki Wallenius, and Pyry Takala. Good debt or bad debt: Detecting semantic orientations in economic texts. J. Assoc. Inf. Sci. Technol., 65(4):782–796, 2014

    2014

  37. [46]

    Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering

    Ankit Pal, Logesh Kumar Umapathi, and Malaikannan Sankarasubbu. Medmcqa: A large-scale multi-subject multi-choice dataset for medical domain question answering. In Gerardo Flores, George H Chen, Tom Pollard, Joyce C Ho, and Tristan Naumann, editors, Proceedings of the Conferen...

    2022

  38. [47]

    The sciqa scientific question answering benchmark for scholarly knowledge

    Sören Auer, Dante AC Barone, Cassiano Bartz, Eduardo G Cortes, Mohamad Yaser Jaradeh, Oliver Karras, Manolis Koubarakis, Dmitry Mouromtsev, Dmitrii Pliukhin, Daniil Radyush, et al. The sciqa scientific question answering benchmark for scholarly knowledge. Scientific Reports, 1...

    2023

  39. [48]

    Socialiqa: Commonsense reasoning about social interactions

    Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. Socialiqa: Commonsense reasoning about social interactions. CoRR, abs/1904.09728, 2019

    1904 arXiv

  40. [49]

    Phillip Keung, Yichao Lu, György Szarvas, and Noah A. Smith. The multilingual amazon reviews corpus. In EMNLP (1), pages 4563–4568. Association for Computational Linguistics, 2020

    2020

  41. [50]

    What makes good in-context examples for gpt-3? In DeeLIO@ACL, pages 100–114

    Jiachang Liu, Dinghan Shen, Yizhe Zhang, Bill Dolan, Lawrence Carin, and Weizhu Chen. What makes good in-context examples for gpt-3? In DeeLIO@ACL, pages 100–114. Association for Computational Linguistics, 2022

    2022

  42. [51]

    Compositional ex- emplars for in-context learning

    Jiacheng Ye, Zhiyong Wu, Jiangtao Feng, Tao Yu, and Lingpeng Kong. Compositional ex- emplars for in-context learning. In ICML, volume 202 of Proceedings of Machine Learning Research, pages 39818–39833. PMLR, 2023

    2023

  43. [52]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, Anirudh Goyal, Anthony Hartshorn, Aobo Yang, Archi Mitra, Archie Sravankumar, Artem Korenev, Arthur Hinsvark, Arun Rao, Aston Zha...

    2024 arXiv

  44. [53]

    Smith, and Tao Yu

    Hongjin Su, Jungo Kasai, Chen Henry Wu, Weijia Shi, Tianlu Wang, Jiayi Xin, Rui Zhang, Mari Ostendorf, Luke Zettlemoyer, Noah A. Smith, and Tao Yu. Selective annotation makes language models better few-shot learners. In ICLR. OpenReview.net, 2023

    2023

  45. [54]

    IDEAL: influence-driven selective annotations empower in-context learners in large language models

    Shaokun Zhang, Xiaobo Xia, Zhaoqing Wang, Ling-Hao Chen, Jiale Liu, Qingyun Wu, and Tongliang Liu. IDEAL: influence-driven selective annotations empower in-context learners in large language models. In ICLR. OpenReview.net, 2024

    2024

  46. [55]

    Let’s verify step by step

    Hunter Lightman, Vineet Kosaraju, Yuri Burda, Harrison Edwards, Bowen Baker, Teddy Lee, Jan Leike, John Schulman, Ilya Sutskever, and Karl Cobbe. Let’s verify step by step. In ICLR. OpenReview.net, 2024

    2024

  47. [56]

    David Rein, Betty Li Hou, Asa Cooper Stickland, Jackson Petty, Richard Yuanzhe Pang, Julien Dirani, Julian Michael, and Samuel R. Bowman. GPQA: A graduate-level google-proof q&a benchmark. In First Conference on Language Modeling, 2024

    2024

  48. [57]

    A.", "B

    Naman Jain, King Han, Alex Gu, Wen-Ding Li, Fanjia Yan, Tianjun Zhang, Sida Wang, Ar- mando Solar-Lezama, Koushik Sen, and Ion Stoica. Livecodebench: Holistic and contamination free evaluation of large language models for code. In ICLR. OpenReview.net, 2025. 13 Algorithm 1 Inf...

    2025

  49. [2017]

    A.", "B

    In January 2017, it was stated that the sixth season would end the main storyline, and for a seventh season, the series would be softly rebooted with a new storyline. Question: is there a season six of once upon a time Label:True ... Definition: Given a question from a scienti...

    2017

Pith tools

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