Pith. sign in

REVIEW 3 major objections 6 minor 41 references

From Profiles to Steering Vectors: Global Sparse Priors and Local Semantic Calibration for Personalized Text Generation

T0 review · 3 major / 6 minor · reviewed 2026-08-02 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a user's writing style can be separated from topic content in a sparse-autoencoder feature space, and that injecting one global style vector plus small per-scenario contrastive vectors at inference time personalizes t

desk verdict A credible, well-tested personalization method with a solid empirical win, but the SAE disentanglement claim is thinner than the headline suggests. read the letter →

arxiv 2607.21620 v1 pith:YGMM7CD4 submitted 2026-06-01 cs.AI

Liuji Chen , Zeyu Zhang , Xinyuan Zhang , Shuai Nie , Qiang Liu , Shu Wu , Liang Wang This is my paper
classification cs.AI
keywords personalizedtextgenerationactivationsteeringsparseautoencodersstyledisentanglementcontrastivetraining-freepersonalizationinference-timeinterventionvectors
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a user's writing style can be extracted from their past texts and injected into a large language model at inference time without retrieval, fine-tuning, or large per-user storage. The key move is to do this extraction in a sparse autoencoder latent space rather than in the model's dense activations, because sparse features separate style from topic more cleanly. The method builds a global style vector by averaging sparse activations over all of a user's response tokens, and local style vectors by contrasting activations under positively and negatively selected contexts within clustered interaction scenarios. At generation time it adds these vectors to different model layers. On standard personalization benchmarks, this outperforms retrieval-based, fine-tuning-based, and dense-steering baselines on most metrics.

What carries the argument

The central object is the sparse autoencoder used as a bidirectional style-content separator: an encoder maps each token's dense activation into a high-dimensional sparse feature vector, an average over all user tokens forms a global sparse profile, a decoder maps that profile back to a residual-stream steering vector, and a second contrastive pass over clustered scenarios produces local vectors. The work the machinery does is to filter out topic-specific semantic residue before vectors are injected, which is what distinguishes the method from simple dense activation averaging.

What would settle it

Take a set of users with identical topic distributions but contrasting styles and compare the global vectors produced by GLASS against dense activation averaging; if the SAE vector is not measurably closer to the user's true style direction (for example, it decodes to topic words, or the performance gap over dense steering vanishes on topic-balanced data), the central claim fails.

Watch

Extended reading notes

Core claim

On its own terms, the paper claims that the reason activation steering has underperformed for personalization is semantic residue: dense activations from user histories mix style with topic, so steering vectors imitate content rather than voice. GLASS replaces the dense space with the latent space of a sparse autoencoder. It averages sparse activations over every token of a user's historical responses and decodes the average into a global steering vector vG; it then clusters the user's prompts, and within each cluster contrasts answer activations under positively selected versus negatively sampled contexts to produce a local vector vC. At inference, vG is added at an upper layer and vC at an

Load-bearing premise

The method stands or falls on whether sparse-autoencoder features separate style from semantic content well enough that averaging them over a user's history yields a global vector with little topic contamination.

Editorial extensions

If this is right

  • Personalization becomes a per-user vector bank: one global vector plus a handful of cluster vectors, replaceable without retraining or retrieval at query time.
  • Sparse-space steering should degrade less than dense-space steering when a user's history shifts in topic or length, because the extracted vectors carry less semantic residue.
  • Because offline cost is a few forward passes and per-user storage is below ten kilobytes, the approach scales to user populations where per-user fine-tuning would be impractical.
  • The method transfers across short- and long-form generation and across different SAE implementations, so the result is not tied to one pretrained sparse-autoencoder checkpoint.

Reading between the lines

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

  • Beyond the paper: if sparse features really separate style from content, then the same machinery could compose or negate style vectors to adjust one dimension (e.g., formality) while holding topic fixed, without collecting new user data.
  • Beyond the paper: clustering on TF-IDF prompt features is only one way to define scenarios; clustering on stylistic or dialog-act features might produce sharper local vectors and is a direct extension testable on the same benchmarks.
  • Beyond the paper: the paper's disentanglement evidence is indirect—robustness to perturbation rather than direct feature inspection; a stronger test would ablate individual sparse features and observe targeted style changes in generated text.
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 / 6 minor

Summary. The paper proposes GLASS, a training-free personalized text generation framework that combines global and local activation steering in sparse autoencoder (SAE) latent space. The global vector is obtained by averaging SAE feature activations over all tokens of a user's history and decoding the average; local vectors are constructed by clustering user prompts, forming positive/negative contrastive examples, and subtracting mean-pooled activations. At inference, GLASS injects the global vector and the nearest-cluster local vector into different layers of LLaMA-3-8B. Experiments on LaMP-4, LaMP-5, LongLaMP-3, and LongLaMP-4 compare GLASS against retrieval-, PEFT-, and steering-based baselines using ROUGE-1/ROUGE-L and a GPT-5 judge score. The paper reports that GLASS ranks first on 11 of 12 main metrics, and it includes ablations, robustness analyses, sensitivity studies, efficiency measurements, and a case study. The central mechanistic claim is that SAE-based sparse features reduce semantic residue in style vectors, yielding better disentanglement of style from content.

Significance. If the claims are supported, GLASS would be a practically attractive method: it requires no user-specific training, stores only a small vector bank per user, and appears to outperform strong baselines on standard benchmarks. The paper provides a clear algorithmic description, pseudocode, ablation studies, and an analysis of layer/strength sensitivity, which are useful for reproducibility. However, the headline result rests on the assertion that SAE sparse features separate style from semantic content, and this mechanism is not directly verified. In addition, the main evaluation lacks error bars and significance tests, so the extent of the improvement over the strongest baselines is uncertain. The robustness analysis in Figure 1 is too underspecified to serve as evidence for the disentanglement claim. These issues make the current evidence suggestive rather than conclusive.

major comments (3)
  1. [§4.3, Table 1] The central empirical claim that GLASS 'consistently outperforms' baselines is not supported by any error bar, confidence interval, or significance test. Several margins are very small; for example, on LongLaMP-4 the R-1 difference vs. OPPU is 0.1284 vs. 0.1277 and the R-L difference is 0.0947 vs. 0.0944, and on LaMP-4 the judge gap vs. StyleVector is 2.09 vs. 1.95. In addition, Table 4 shows that all intervention layers and strengths (α, β, γ) are tuned per dataset on validation data. Without paired significance tests or multiple-seed variation, the '11 of 12 metrics first' claim is not demonstrably robust. Please report confidence intervals or paired bootstrap p-values across the 200 users, and ideally a small seed/layer perturbation analysis.
  2. [§3.1, Eq. (6)–(7); §4.5.1, Figure 1] The core mechanism—that SAE sparse features separate style from semantic content—is asserted rather than demonstrated. The global vector vG is the average of SAE feature activations over all tokens of all user responses (Eq. 6) decoded back into residual space (Eq. 7). Any topic feature that co-occurs frequently in the user's history will survive this averaging, so the purported freedom from semantic residue is not guaranteed. The only direct evidence offered is Figure 1, but the perturbation procedure is described in a single sentence ('perturb user histories along length and topic'), with no details on how perturbations were constructed, what exactly is measured, or whether the plotted values are aggregated with error bars. Moreover, downstream performance drop under perturbed histories is an indirect proxy; it does not measure the semantic content of vG itself. Table 2 shows that remo
  3. [§3.2.3, Eq. (9)] The construction of positive examples for local contrastive extraction uses ROUGE-L of a one-shot prompted generation against the user's reference as a utility score. This is an ad-hoc proxy: ROUGE-L is a lexical-overlap measure, and the paper does not provide evidence that this selection criterion is aligned with the stylistic signal the local vector is intended to capture. Since the local vector is a substantial component of the method (removing vC degrades performance, Table 2), the validity of this selection procedure is load-bearing. I recommend ablating alternative utility scores (e.g., embedding similarity or an LLM judge score) or at least reporting the sensitivity of the method to the choice of the top-n positive examples.
minor comments (6)
  1. [Table 1, Fints row] The Fints row contains a formatting error: '2.040.20340.1212' should be split into separate judge, R-1, and R-L columns. This makes the table difficult to read.
  2. [Figure 5] The efficiency figure is visually unclear. It uses bubble size to indicate latency but does not label the vertical axis or provide a legend for bubble-size units. Please add explicit axis labels and a scale.
  3. [§4.5.1, Figure 1] The robustness analysis would be far more informative with a precise description of the perturbation protocol: how much are lengths shifted, how are topics changed, how are positive/negative pairs created, and are results averaged over multiple perturbation seeds? The current one-sentence description prevents readers from reproducing or interpreting the comparison.
  4. [§4.1.1, Datasets] The exclusion of the email subset (for privacy) and LaMP-7 (as fully synthetic) is reasonable, but it should be stated whether this was decided before or after seeing the results, and whether the same 200-user sampling protocol was used for all baselines. A short note would remove ambiguity.
  5. [§3.2.1, Clustering] Adaptive K via silhouette score is compared against a fixed K, but the candidate range of K and the number of anchors per cluster are not reported. These details matter for reproducibility of the local vector bank.
  6. [Appendix D.1, Judge] The GPT-5 judge scores are reported as averages over a single deterministic run. Given the known variability of LLM judges, I recommend also reporting standard deviations or an agreement measure (e.g., Cohen's kappa with a second judge) even on a small sample.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: GLASS's steering vectors are empirical constructions validated on held-out benchmarks; the only author-overlap citation is a transparent, non-load-bearing baseline reference.

full rationale

The paper's derivation chain (Eqs. 5-7, 11-16) constructs the global vector vG by averaging SAE features over the user's own history and decoding them (Eqs. 6-7), and constructs local contrastive vectors from positive/negative activations (Eqs. 11-14). These are modeling constructions, not derivations of the target from itself: the claim that the vectors encode style is tested against held-out LaMP/LongLaMP references, and all baselines are evaluated against the same user-specific references. That is the personalization task itself, not a circular step. The use of ROUGE-L in positive-sample selection (Eq. 9) is a training-time selection objective on user history; it does not consume test references, so it is not a fitted-input-called-prediction. The robustness probe in Figure 1 is an independent perturbation experiment; it may be underspecified or indirect, but that is an empirical-validity concern, not circularity. The only author-overlap citation is StyleVector (Zhang et al., 2025a), which shares co-authors; it is used as a comparative baseline and motivation, and no load-bearing claim or uniqueness theorem is imported from it. The Limitations section concedes that the construction 'still relies on fixed SAE representations and activation-level aggregation,' which flags an unverified assumption about disentanglement, not a circular step. No equation reduces to its own input; the score of 2 reflects only the minor, transparent self-citation.

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

The ledger shows the method rests on four domain assumptions with weak direct evidence, plus per-dataset hyperparameters. The SAE disentanglement assumption is the most load-bearing and is only indirectly probed.

free parameters (3)
  • Steering strengths α (global), β (local), γ (priming) = LaMP-4: 0.8/0.2/0.4; LaMP-5: 0.4/0.6/1.0; LongLaMP-3: 0.4/0.8/0.1; LongLaMP-4: 0.4/0.7/0.1
    Selected on validation data per dataset; directly control the injected vector magnitudes and are central to the results.
  • Intervention layers LG (global), LC (local), LSP (priming) = LG=19, LC=23, LSP=23 for all datasets
    Chosen on validation; different layers have very different effects per Figure 6.
  • Number of clusters K per user = not fixed; selected by maximizing silhouette score
    Controls granularity of local style vectors; adaptive K claimed to outperform fixed K (Table 9).
assumptions (4)
  • domain assumption SAE sparse features are approximately disentangled, so token-level averaging preserves style and cancels semantic content.
    Sec 3.1, Eq. (6)-(7); core premise for vG. Justified only indirectly via Figure 1.
  • domain assumption TF-IDF plus K-means partitions each user's prompts into coherent interaction scenarios such that contrastive directions within a cluster isolate style.
    Sec 3.2.1; cluster quality is not directly evaluated.
  • domain assumption Adding a constant vector to hidden states at certain layers linearly steers output style without destroying task competence.
    Sec 3.3, Eq. (15)-(16); standard linear representation hypothesis from prior work, not re-derived here.
  • ad hoc to paper ROUGE-L of a one-shot prompted generation against the user's reference is a valid utility score for selecting positive demonstration examples.
    Sec 3.2.3, Eq. (9); heuristic choice with no justification in the paper.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Profiles to Steering Vectors: Global Sparse Priors and Local Semantic Calibration for Personalized Text Generation." pith.science (2026). https://pith.science/paper/YGMM7CD4

@misc{pith2026260721620,
  author       = {Pith},
  title        = {Pith review of: From Profiles to Steering Vectors: Global Sparse Priors and Local Semantic Calibration for Personalized Text Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/YGMM7CD4}},
  note         = {Machine review of arXiv:2607.21620}
}
read the original abstract

Personalized text generation requires models to capture user-specific writing styles from historical data. Existing approaches based on retrieval, parameter-efficient fine-tuning, or activation steering either introduce inference and storage overhead or struggle to separate stylistic signals from semantic content. We propose GLASS, a training-free framework for personalized generation via Global-Local Activation Steering with Sparse priors. GLASS uses sparse autoencoders to extract a global user-style prior from historical responses and constructs local contrastive style vectors over clustered interaction scenarios. During inference, it jointly injects global and local vectors into different model layers, enabling context-aware personalization without retrieval or parameter updates. Experiments on LaMP and LongLaMP show that GLASS outperforms retrieval-, fine-tuning-, and steering-based baselines across ROUGE metrics and LLM-as-judge evaluations. Further analyses show that SAE-based representations are more robust to topic and length shifts, suggesting better disentanglement of stylistic information from semantic residue.

Figures

Figures reproduced from arXiv: 2607.21620 by the authors.

Figure 1
Figure 1. Performance degradation on downstream tasks [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of GLASS. the user’s historical interaction data, where xi rep￾resents a prompt specifying a particular task and yi denotes the corresponding text written in the user’s authentic style. Given a new user request x, a model M is expected to produce a personalized output yˆ = M(x, Du) that satisfies the generation requirements of user u while preserving their stylis￾tic nuances. 2.2 Sparse Autoencoders Sparse … view at source ↗
Figure 3
Figure 3. The inference procedure of GLASS. For negative sample construction, we sample n histories from other users to form the negative n￾shot context S −(xa) = [(xj , yj )]n j=1, where each (xj , yj ) is drawn from S u′̸=u Du′ without any clus￾ter constraint. To filter noisy pairs, we define the personaliza￾tion score under positive and negative contexts as s + a = score(M˜ ([S +(xa); xa]), ya), s − a = score(M([S −(xa); x… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Impact of contrastive sample construction on [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Inference-time overhead of different methods. [PITH_FULL_IMAGE:figures/full_fig_p008_5.png]
Figure 6
Figure 6. Figure 6: Performance comparison under different steering layers and intervention strength settings. [PITH_FULL_IMAGE:figures/full_fig_p015_6.png]
Figure 7
Figure 7. Figure 7: Pairwise interaction between the global style vector [PITH_FULL_IMAGE:figures/full_fig_p017_7.png]
Figure 8
Figure 8. Figure 8: Pairwise interaction between the global style vector [PITH_FULL_IMAGE:figures/full_fig_p018_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 1 linked inside Pith

  1. [1]

    Aho and Jeffrey D

    Alfred V. Aho and Jeffrey D. Ullman , title =. 1972

    1972

  2. [2]

    Publications Manual , year = "1983", publisher =

    1983

  3. [3]

    Chandra and Dexter C

    Ashok K. Chandra and Dexter C. Kozen and Larry J. Stockmeyer , year = "1981", title =. doi:10.1145/322234.322243

    arXiv 1981

  4. [4]

    Scalable training of

    Andrew, Galen and Gao, Jianfeng , booktitle=. Scalable training of

  5. [5]

    Dan Gusfield , title =. 1997

    1997

  6. [6]

    Tetreault , title =

    Mohammad Sadegh Rasooli and Joel R. Tetreault , title =. Computing Research Repository , volume =. 2015 , url =

    2015

  7. [7]

    A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =

    Ando, Rie Kubota and Zhang, Tong , Issn =. A Framework for Learning Predictive Structures from Multiple Tasks and Unlabeled Data , Volume =. Journal of Machine Learning Research , Month = dec, Numpages =

  8. [8]

    2025 , eprint=

    A Survey of Large Language Models , author=. 2025 , eprint=

    2025

Show all 41 references
  1. [9]

    arXiv preprint arXiv:2502.11528 , year=

    A Survey of Personalized Large Language Models: Progress and Future Directions , author=. arXiv preprint arXiv:2502.11528 , year=

  2. [10]

    2025 , eprint=

    Personalized Text Generation with Contrastive Activation Steering , author=. 2025 , eprint=

    2025

  3. [11]

    Nature Machine Intelligence , volume=

    The benefits, risks and bounds of personalizing the alignment of large language models to individuals , author=. Nature Machine Intelligence , volume=. 2024 , publisher=

    2024

  4. [12]

    2025 , eprint=

    Personalized Graph-Based Retrieval for Large Language Models , author=. 2025 , eprint=

    2025

  5. [13]

    Alireza Salemi and Sheshera Mysore and Michael Bendersky and Hamed Zamani , year=. La. 2304.11406 , archivePrefix=

    arXiv

  6. [14]

    2024 , eprint=

    LongLaMP: A Benchmark for Personalized Long-form Text Generation , author=. 2024 , eprint=

    2024

  7. [15]

    2024 , eprint=

    Retrieval-Augmented Generation for Large Language Models: A Survey , author=. 2024 , eprint=

    2024

  8. [16]

    2024 , eprint=

    A Survey on In-context Learning , author=. 2024 , eprint=

    2024

  9. [17]

    2024 , eprint=

    Persona-SQ: A Personalized Suggested Question Generation Framework For Real-world Documents , author=. 2024 , eprint=

    2024

  10. [18]

    Personalized Pieces: Efficient Personalized Large Language Models through Collaborative Efforts

    Tan, Zhaoxuan and Liu, Zheyuan and Jiang, Meng. Personalized Pieces: Efficient Personalized Large Language Models through Collaborative Efforts. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. doi:10.18653/v1/2024.emnlp-main.371

    2024 doi

  11. [19]

    2024 , eprint=

    HYDRA: Model Factorization Framework for Black-Box LLM Personalization , author=. 2024 , eprint=

    2024

  12. [20]

    Steering off Course: Reliability Challenges in Steering Language Models

    Da Silva, Patrick Queiroz and Sethuraman, Hari and Rajagopal, Dheeraj and Hajishirzi, Hannaneh and Kumar, Sachin. Steering off Course: Reliability Challenges in Steering Language Models. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (V...

    2025 doi

  13. [21]

    2025 , eprint=

    Fints: Efficient Inference-Time Personalization for LLMs with Fine-Grained Instance-Tailored Steering , author=. 2025 , eprint=

    2025

  14. [22]

    SAE s Are Good for Steering -- If You Select the Right Features

    Arad, Dana and Mueller, Aaron and Belinkov, Yonatan. SAE s Are Good for Steering -- If You Select the Right Features. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025. doi:10.18653/v1/2025.emnlp-main.519

    2025 doi

  15. [23]

    S tyle R emix: Interpretable Authorship Obfuscation via Distillation and Perturbation of Style Elements

    Fisher, Jillian and Hallinan, Skyler and Lu, Ximing and Gordon, Mitchell L and Harchaoui, Zaid and Choi, Yejin. S tyle R emix: Interpretable Authorship Obfuscation via Distillation and Perturbation of Style Elements. Proceedings of the 2024 Conference on Empirical Methods in N...

    2024 doi

  16. [24]

    A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models

    Shu, Dong and Wu, Xuansheng and Zhao, Haiyan and Rai, Daking and Yao, Ziyu and Liu, Ninghao and Du, Mengnan. A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Models. Findings of the Association for Computational Linguistics: EMNLP 2025. 2...

    2025 doi

  17. [25]

    2024 , eprint=

    The Llama 3 Herd of Models , author=. 2024 , eprint=

    2024

  18. [26]

    2021 , url =

    Unsupervised Dense Information Retrieval with Contrastive Learning , author=. 2021 , url =

    2021

  19. [27]

    2025 , eprint=

    PRIME: Large Language Model Personalization with Cognitive Dual-Memory and Personalized Thought Process , author=. 2025 , eprint=

    2025

  20. [28]

    LLM -based Medical Assistant Personalization with Short- and Long-Term Memory Coordination

    Zhang, Kai and Kang, Yangyang and Zhao, Fubang and Liu, Xiaozhong. LLM -based Medical Assistant Personalization with Short- and Long-Term Memory Coordination. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Hum...

    2024 doi

  21. [30]

    Proceedings of the 2025 International ACM SIGIR Conference on Innovative Concepts and Theories in Information Retrieval (ICTIR) , pages =

    Salemi, Alireza and Zamani, Hamed , title =. Proceedings of the 2025 International ACM SIGIR Conference on Innovative Concepts and Theories in Information Retrieval (ICTIR) , pages =. 2025 , isbn =. doi:10.1145/3731120.3744584 , abstract =

    2025

  22. [31]

    2023 , eprint=

    Integrating Summarization and Retrieval for Enhanced Personalization via Large Language Models , author=. 2023 , eprint=

    2023

  23. [32]

    Guided Profile Generation Improves Personalization with Large Language Models

    Zhang, Jiarui. Guided Profile Generation Improves Personalization with Large Language Models. Findings of the Association for Computational Linguistics: EMNLP 2024. 2024. doi:10.18653/v1/2024.findings-emnlp.231

    2024 doi

  24. [33]

    2025 , eprint=

    Can Large Language Models Understand Preferences in Personalized Recommendation? , author=. 2025 , eprint=

    2025

  25. [34]

    LLM s + Persona-Plug = Personalized LLM s

    Liu, Jiongnan and Zhu, Yutao and Wang, Shuting and Wei, Xiaochi and Min, Erxue and Lu, Yu and Wang, Shuaiqiang and Yin, Dawei and Dou, Zhicheng. LLM s + Persona-Plug = Personalized LLM s. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (...

    2025 doi

  26. [35]

    Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1 , pages =

    Ding, Yucheng and Tan, Yangwenjian and Liu, Xiangyu and Niu, Chaoyue and Meng, Fandong and Zhou, Jie and Liu, Ning and Wu, Fan and Chen, Guihai , title =. Proceedings of the 31st ACM SIGKDD Conference on Knowledge Discovery and Data Mining V.1 , pages =. 2025 , isbn =. doi:10....

    2025

  27. [36]

    2023 , eprint=

    Representation Engineering: A Top-Down Approach to AI Transparency , author=. 2023 , eprint=

    2023

  28. [37]

    International Conference on Machine Learning , year=

    In-context vectors: Making in context learning more effective and controllable through latent space steering , author=. International Conference on Machine Learning , year=

  29. [38]

    Steering Llama 2 via Contrastive Activation Addition

    Rimsky, Nina and Gabrieli, Nick and Schulz, Julian and Tong, Meg and Hubinger, Evan and Turner, Alexander. Steering Llama 2 via Contrastive Activation Addition. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 202...

    2024 doi

  30. [39]

    2024 , eprint=

    Steering Language Models With Activation Engineering , author=. 2024 , eprint=

    2024

  31. [40]

    T ruth X : Alleviating Hallucinations by Editing Large Language Models in Truthful Space

    Zhang, Shaolei and Yu, Tian and Feng, Yang. T ruth X : Alleviating Hallucinations by Editing Large Language Models in Truthful Space. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2024. doi:10.18653/v1/2024.acl...

    2024 doi

  32. [41]

    2024 , eprint=

    Truth Forest: Toward Multi-Scale Truthfulness in Large Language Models through Intervention without Tuning , author=. 2024 , eprint=

    2024

  33. [42]

    Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning

    Tan, Zhaoxuan and Zeng, Qingkai and Tian, Yijun and Liu, Zheyuan and Yin, Bing and Jiang, Meng. Democratizing Large Language Models via Personalized Parameter-Efficient Fine-tuning. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing. 2024. d...

    2024 doi

Pith tools

Reviewed August 2, 2026 · model on record in the stance chip above.