Pith. sign in

REVIEW 4 major objections 4 minor 38 references

SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features via Representation Verbalization

T0 review · 4 major / 4 minor · reviewed 2026-08-14 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning an LLM on feature–explanation pairs lets it explain sparse autoencoder features directly from decoder directions, generalizing to unseen features, transferring across SAE dictionaries, and extending to other LLMs through a…

desk verdict Within-dictionary verbalization of SAE decoder directions is real and well-engineered; the cross-dictionary transfer headline is likely confounded by Matryoshka nesting and needs an independent test before it can stand. read the letter →

arxiv 2608.13538 v1 pith:6ZKRGTT7 submitted 2026-08-13 cs.CL

Weihan Meng , Hongzhu Guo , Yi Jing , Dewen Liu , Zijun Yao , Xiaozhi Wang , Lei Hou , Juanzi Li This is my paper · ORCID
classification cs.CL
keywords sparseautoencodersfeatureinterpretationrepresentationverbalizationdecoderdirectionsLLMinterpretabilitycross-LLMtransferfine-tuninglanguagemodelsreferenceagreement
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

The paper claims that a sparse autoencoder feature can be explained in natural language by a fine-tuned LLM that never sees the feature's activating text. The method injects the feature's decoder direction into the LLM's hidden states during prompt prefilling and fine-tunes only the layers downstream of the injection to complete a fixed sentence naming the concept. If correct, feature explanation becomes a one-pass inference from the internal representation itself, removing the corpus-scale activation search that current bottom-up interpretability pipelines require. The authors demonstrate that the trained verbalizer generalizes to unseen features, transfers to a different SAE dictionary over the same representation space without retraining, and—through a lightweight affine adapter trained on unlabeled aligned representations—explains features from other LLMs. They also show that jointly injecting two directions yields combined meanings and that reversing a direction shifts the generated meaning, indicating sensitivity to compositional and signed structure in decoder directions.

What carries the argument

The load-bearing object is the decoder direction: the column of the SAE decoder that defines a feature as a direction in the LLM's representation space. The paper uses it as the only feature-specific signal. At inference the direction is normalized to unit norm, scaled by the mean representation norm of the injection span, and added to every token representation in that span at a chosen layer; the prompt is a fixed sentence ending in 'The target concept is:', and the downstream layers generate the explanation. Partial fine-tuning freezes layers at or before the injection layer, preserving the representation space in which decoder directions live, and trains only layers after it plus the final norm and language-model head. For cross-LLM transfer, a single affine layer maps source-layer representations into the injection-layer space; because injection is additive, the adapter's bias cancels and only its weight matrix $W$ maps the direction itself. This whole interface is what makes explanation a one-pass readout from internal geometry rather than a search over activation examples.

What would settle it

Construct a gold set of features with independently verified semantics by having several human annotators label features from both their top-activating examples and targeted behavioral probes, retaining only features where annotators converge. If the verbalizer's Reference Agreement with the filtered references stays high while its agreement with those gold labels is near chance, the central claim would be falsified; if the verbalizer tracks the gold labels even where they disagree with the filtered references, the claim would be strengthened. A complementary strong test is synthetic: add a known concept direction to representations at the injection layer and check whether the verbalizer names that concept at above-chance rates.

Watch

Extended reading notes

Core claim

The paper's central claim is that the semantic content of an SAE feature is accessible in its decoder direction alone, and that an LLM can be trained to read that direction aloud. SAEVerbalizer injects a normalized decoder direction into the token representations of a fixed prompt at a chosen layer, then fine-tunes only the layers after that layer to continue the prompt with an explanation. The trained verbalizer produces explanations for features it never saw during training, transfers directly to a different SAE dictionary over the same representation space, and, via an affine adapter trained on aligned hidden states from unlabeled text, explains features from a different LLM. In the main evaluation, the best configuration attains Reference Agreement—the fraction of generated explanations an LLM judge considers to agree with the reference explanation—of 52.3 percent on a global train-standard set, 80.5 percent on low-index gold features, and 56.1 percent on a global gold set. The authors interpret these results as evidence that representation verbalization is a trainable, partially reusable capability rather than a byproduct of studying activation examples.

Load-bearing premise

The load-bearing premise is that the filtered reference explanations used as training targets and evaluation references correctly capture what each SAE feature means; if those references are wrong or too broad, high Reference Agreement does not show the verbalizer produces correct explanations.

Editorial extensions

If this is right

  • A single trained verbalizer can explain any feature of any SAE defined on the same LLM layer, since decoder directions from different dictionaries share that representation space.
  • Explanation cost drops to one forward pass on a short fixed prompt per feature, eliminating corpus-scale inference and example retrieval.
  • Cross-LLM explanation can be obtained from as little as one affine layer trained on aligned unlabeled text, so the method can be extended to new LLMs without feature-level supervision.
  • Adding supervision beyond a small amount (1.5k pairs already produce large gains) is what creates the capability; raw backbones show almost no spontaneous explanation ability, with Reference Agreement around 1–2 percent.
  • The verbalizer's sensitivity to joint injection and sign reversal offers a direct way to probe compositional and signed structure in decoder directions.

Reading between the lines

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

  • A natural extension the paper does not test: features whose activations are rare or absent in text corpora might still be explained from their decoder directions, because the method never depends on collecting activation examples; this could be checked by comparing verbalizer outputs on low-frequency features against human labels.
  • The same trained interface could be pushed beyond naming a concept, for example to produce test sentences that activate a feature or to answer questions about its behavior; the paper's fixed 'name the concept' prompt is only one instantiation of representation verbalization.
  • If adapters transfer direction semantics across LLMs, then decoder-direction geometry itself may be largely shared across models at matched depths; a stronger test would be whether adapter-mapped directions compose consistently, for instance whether $\text{adapter}(A+B) \approx \text{adapter}(A)+\text{adapter}(B)$ across many features.
  • An important caveat that follows from the paper's own evaluation choices: Reference Agreement measures consistency with filtered references, not correctness, so a human-gold validation study would be needed before treating verbalizer outputs as verified descriptions of model behavior.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. The paper proposes SAEVerbalizer, a method that fine-tunes the downstream layers of an LLM so that, given only a sparse autoencoder decoder direction injected into a designated prompt span, the model generates a natural-language explanation of the feature. The authors train on feature–explanation pairs from filtered Neuronpedia references, evaluate with Reference Agreement (RA) against the same reference pool, and report generalization to held-out features across Gemma 3 1B/4B/27B backbones and multiple Gemma Scope 2 layers (Table 1), reuse on a width-65k SAE on the same LLM and layer (Table 2), adapter-based cross-LLM transfer (Table 3), and ablations over supervision size, prompt, injection span, and injection strength. They also report qualitative cases for joint injection and sign reversal. The central claimed capability is that a single fine-tuned LLM can explain unseen SAE features directly from decoder directions without per-feature activation analysis.

Significance. If the capability claims hold, the paper makes a useful practical contribution: it replaces corpus-scale activation mining with a single forward pass on a fixed prompt for each feature, and it demonstrates that explanation of unseen features can be learned from a few thousand pairs. Strengths include the disjoint test splits (GTS/LIG/GG), the clear zero-supervision backbone baseline in Table 4, the careful documentation of prompts and filtering, and the explicit limitations section. The main quantitative effects—for example, 1.6% backbone versus 52.3% RA for the default 27B-L16 configuration—are large and support the existence of a learned verbalization capability. However, the cross-dictionary transfer claim is weakened by the Matryoshka nesting of the two Gemma Scope 2 dictionaries, and the evaluation metric partly measures agreement with the same filtered Neuronpedia pool used for supervision, so the headline 'explains features' wording overstates what RA establishes.

major comments (4)
  1. [§4.3.1 (Table 2)] The cross-dictionary transfer experiment is confounded by Matryoshka nesting. The width-65k and width-262k Gemma Scope 2 SAEs are from the same Matryoshka family, and the paper itself invokes the Matryoshka reconstruction objective in §4.1. In such dictionaries, narrower dictionaries are nested subsets of wider ones, so the width-65k test directions may coincide with low-index width-262k directions already present in the 48k training pairs. The paper never measures decoder-direction overlap between the two dictionaries. As a result, the high RA in Table 2 could reflect memorized directions rather than a learned capability that transfers to a genuinely separate dictionary. This is load-bearing because the abstract and §4.3.1 claim transfer 'across separately trained SAE dictionaries.' Please report the distribution of cosine similarities between the width-65k directions and the width-262k training directions, and repeat the transfer test either with an independently trained SAE or with the width-65k directions restricted to those with no near-neighbor in the training dictionary. The within-dictionary unseen-feature results and the adapter results are not affected by this particular confound.
  2. [§4.1 (Evaluation Protocol) and Limitations] RA is a permissive judge-based agreement metric against references drawn from the same filtered Neuronpedia pool used to construct the training supervision. The judge counts synonyms, subset/superset, and 'strong direct logical/contextual association' as agreement, and the model is explicitly fine-tuned to reproduce filtered Neuronpedia explanations. Consequently, the reported RA values partly measure reproduction of the reference distribution rather than explanation correctness. The Limitations section acknowledges that RA does not establish absolute correctness, but the concern is stronger: for the central claim that the verbalizer 'explains SAE features directly from decoder directions,' the evaluation should include at least a human-annotated subset, a hard-negative or distractor-reference condition, or an inter-judge agreement measure to show that RA is not merely style matching. This point affects the interpretation of every quantitative table.
  3. [Appendix F and Tables 1, 4, 5] All configurations are single runs, so the quantitative claims lack variability estimates. Differences such as 48.1% versus 52.3% on GTS across supervision sizes (Table 4) and 50.6% versus 52.3% across injection strengths (Table 5) may be within noise. The Limitations section notes this, but since the paper's central claim rests on these RA differences, please provide multiple seeds for at least the default configuration and the main comparison points, or bootstrap confidence intervals over test features. Without this, the claims about scaling with supervision and robustness to prompt and injection variation are not statistically supported.
  4. [§3.2 (Eq. 2)] The adapter transfer assumes that a linear map trained to reconstruct token representations can be applied to decoder directions via A(h+v)-A(h)=Wv. Decoder directions are not validated to behave as difference vectors in the source representation space, and the MSE objective only aligns pointwise representations. The empirical RA in Table 3 is encouraging, but the paper should provide a direct check that mapped directions land near target-space features (for example, nearest-neighbor directions in a target SAE) or that Wv produces the intended steering effect. Otherwise the cross-LLM transfer claim rests on an unverified linearity assumption.
minor comments (4)
  1. [Table 8] Please clarify that the cosine similarity column reports the cosine between the two decoder directions in each row (for example, the directions of features #40105 and #111800), and state this explicitly in the table caption.
  2. [§4.3.1] The width-65k SAE is described as 'unseen'; 'not used during training' would be more precise, since the dictionary belongs to the same Gemma Scope 2 Matryoshka family as the training dictionary.
  3. [Table 1] Configurations with very low RA, such as 1B-L22 with 3.9% on GTS at 12k pairs, are still summarized as 'generalize to unseen features across all configurations'; please report the corresponding backbone baseline for each configuration or add a statistical threshold to support this blanket statement.
  4. [Appendix D] The evaluation judge prompt permits YES for 'strong direct logical/contextual association,' which may inflate agreement; please report a stricter variant that requires synonymy or near-synonymy, or at least quantify how often the permissive criterion is the deciding factor.

Circularity Check

2 steps flagged · score 6.0 of 10

Cross-dictionary transfer is confounded by Matryoshka nesting, and the adapter LIG column is a checkpoint-selection value; the within-dictionary unseen-feature results remain independent.

  1. fitted input called prediction [Section 4.3.1 (Table 2); Section 4.1; Appendix A.1]
    "We directly apply the default verbalizer to decoder directions from an unseen width-65k Gemma Scope 2 SAE defined on the same LLM and layer. ... Because the two SAEs share the same representation space, transfer requires neither an adapter nor additional feature–explanation supervision. ... LIG and GG differ in index distribution; the latter comparison probes features prioritized by Gemma Scope 2’s Matryoshka reconstruction objective (McDougall et al., 2025; Bussmann et al., 2025)."

    Gemma Scope 2 SAEs are Matryoshka SAEs, so the width-65k dictionary is a nested subset of the width-262k dictionary and its decoder directions coincide with low-index width-262k directions. The paper itself reports that qualified feature–explanation pairs are concentrated among lower-index features and attributes this to the Matryoshka organization. It never checks whether the width-65k test features were among the 48k width-262k training pairs. Under the Matryoshka construction the paper cites, the alleged 'unseen' width-65k test features can be training features from the width-262k dictionary, so the Table 2 RA may measure memorized reproduction of supervision, not transfer to a genuinely separate dictionary.

  2. fitted input called prediction [Appendix C (Validation and Checkpoint Selection); Table 3]
    "Every 1,000 steps, we evaluate a candidate checkpoint by mapping the source-SAE decoder directions into the default27B-L16 verbalizer and computing RA on the corresponding source SAE’s 200-feature LIG set. This selects step 3,000 for the 1B-to-27B adapter and step 4,000 for the 4B-to-27B adapter; GTS and GG are not used for checkpoint selection."

    The adapter checkpoint is selected by optimizing RA on the LIG set, and the same LIG numbers are then reported in Table 3 as part of the transfer result. The adapter-row LIG column is therefore a selection criterion rather than an independent held-out evaluation; its values are partly forced by the selection procedure. Only GTS and GG are free of this selection effect, so the cross-LLM transfer claim should rest on those columns alone, not on the self-referential LIG column.

full rationale

The paper is not globally circular. The core within-dictionary generalization result uses disjoint training and test splits for the same width-262k SAE, and the scaling experiment (Table 4) shows a clear dependence on the amount of supervision, while the cross-LLM adapter transfer is still supported by GTS and GG numbers that were not used for checkpoint selection. However, two evaluation choices introduce partial circularity. First, the 'unseen width-65k SAE' used to claim transfer across SAE dictionaries is, for Gemma Scope 2 Matryoshka SAEs, a nested subset of the width-262k dictionary on which the verbalizer was trained; the paper never verifies decoder-direction overlap, and its own observation that qualified pairs concentrate at low indices makes overlap with the 48k training pairs likely, so Table 2 may be a memorization probe rather than a cross-dictionary prediction. Second, adapter checkpoints are selected on the LIG set, and the same LIG numbers are reported as test results, making that column self-referential. The acknowledged RA-is-not-correctness limitation is not itself circularity, but the shared Neuronpedia reference pool compounds the concern. Overall, one advertised transfer claim is plausibly reduced to training-data reproduction, while the central unseen-feature generalization and the GTS/GG adapter results remain independent evidence for the method.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central empirical claims rest on three unproven premises: decoder directions are semantically decodable via additive injection, the filtered Neuronpedia labels are adequate supervision and evaluation references, and a linear representation-alignment adapter transfers to feature directions. The injection strength alpha and the data-quality thresholds are hand-chosen numbers that influence the reported RA. No new entities are postulated.

free parameters (2)
  • Injection strength alpha = 0.2 (default; evaluated from 0.01 to 1.00)
    Hand-chosen strength for adding the decoder direction to representations; at alpha=0.01 GTS RA drops to 18.9% from 52.3%, so reported results depend on this choice, though performance is stable from 0.1 to 1.0.
  • Filtering thresholds for supervision and test sets = Stage 1 scores >=3, coverage >=0.75; training standard scores >=4, coverage >=0.875; gold standard scores >=5…
    Hand-set quality criteria determine which Neuronpedia explanations enter training and test sets and therefore shape the RA numbers.
assumptions (4)
  • domain assumption Injected decoder directions retain feature-specific semantic information in the residual stream at the injection layer.
    The method assumes norm-matched additive injection of a decoder direction into L_inj representations is sufficient for downstream layers to identify the feature; no independent evidence is provided beyond the RA results.
  • domain assumption Filtered Neuronpedia explanations are valid feature labels, and RA against them measures explanation quality.
    Training and evaluation both rely on LLM-filtered observational explanations; the paper states RA does not establish absolute correctness (Section 4.1, Limitations).
  • ad hoc to paper A linear adapter trained on token-aligned representations transfers to SAE decoder directions via Wv.
    The equivalence A(h+v)-A(h)=Wv assumes features are linear displacements and that representation-space alignment transfers to decoder directions (Section 3.2). No feature-level supervision is used for the adapter.
  • domain assumption Cross-LLM representations at similar relative depths are approximately aligned.
    Adapter layer selection relies on this, citing Wolfram and Schein 2025; it is a background assumption from prior work.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features via Representation Verbalization." pith.science (2026). https://pith.science/paper/6ZKRGTT7

@misc{pith2026260813538,
  author       = {Pith},
  title        = {Pith review of: SAEVerbalizer: Generating Explanations for Sparse Autoencoder Features via Representation Verbalization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6ZKRGTT7}},
  note         = {Machine review of arXiv:2608.13538}
}
read the original abstract

Sparse autoencoders (SAEs) are proposed to extract numerous features from large language model (LLM) representations, yet explaining these features still relies primarily on external observation. This reliance leads to superficial explanations inferred from observed model behavior and computational inefficiency from collecting such behavioral evidence at scale. We introduce SAEVerbalizer, a framework that injects SAE decoder directions into an LLM's representations and fine-tunes the LLM's downstream layers to generate natural-language explanations of the injected features. Once trained, the resulting verbalizer explains SAE features directly from decoder directions, addressing both limitations. Our experiments show that the learned verbalization capability generalizes to unseen features, transfers across separately trained SAE dictionaries, and, with a lightweight adapter, extends to SAE features from different LLMs. Intervention experiments show that injecting multiple directions yields an explanation combining their meanings, while reversing individual directions produces corresponding meaning shifts.

Figures

Figures reproduced from arXiv: 2608.13538 by the authors.

Figure 1
Figure 1. Overview of SAEVERBALIZER. Given an SAE decoder direction and a fixed verbalization prompt, the verbalizer generates an explanation of the feature. corpus, identifies each feature’s top-activating ex￾amples, and prompts an LLM to summarize their shared pattern (Bricken et al., 2023; Paulo et al., 2025). However, these methods face two limi￾tations. Superficial Explanations. Without di￾rectly examining how a feature … view at source ↗
Figure 2
Figure 2. Overall design of SAEVERBALIZER. (a) The Verbalizer. During prompt prefilling, an SAE decoder direction is injected into the token representations of a designated injection span at Linj, conditioning explanation generation. (b) The Adapter. An adapter maps SAE decoder directions in a source LLM’s representation space into the verbalizer’s injection-layer representation space. where ds and dt are the representation d… view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

38 extracted references · 23 canonical work pages

  1. [1]

    Transformer Circuits Thread , note=

    Elhage, Nelson and Hume, Tristan and Olsson, Catherine and Schiefer, Nicholas and Henighan, Tom and Kravec, Shauna and Hatfield-Dodds, Zac and Lasenby, Robert and Drain, Dawn and Chen, Carol and Grosse, Roger and McCandlish, Sam and Kaplan, Jared and Amodei, Dario and Wattenberg, Martin and Olah, Christopher , year=. Transformer Circuits Thread , note=

    work page

  2. [2]

    Bricken, Trenton and Templeton, Adly and Batson, Joshua and Chen, Brian and Jermyn, Adam and Conerly, Tom and Turner, Nick and Anil, Cem and Denison, Carson and Askell, Amanda and Lasenby, Robert and Wu, Yifan and Kravec, Shauna and Schiefer, Nicholas and Maxwell, Tim and Joseph, Nicholas and Hatfield-Dodds, Zac and Tamkin, Alex and Nguyen, Karina and McL...

    work page

  3. [3]

    Huben, Robert and Cunningham, Hoagy and Smith, Logan and Ewart, Aidan and Sharkey, Lee , booktitle =

    work page

  4. [4]

    Gao, Leo and Dupre la Tour, Tom and Tillman, Henk and Goh, Gabriel and Troll, Rajan and Radford, Alec and Sutskever, Ilya and Leike, Jan and Wu, Jeffrey , booktitle =

    work page

  5. [5]

    Daniel and Sumers, Theodore R

    Templeton, Adly and Conerly, Tom and Marcus, Jonathan and Lindsey, Jack and Bricken, Trenton and Chen, Brian and Pearce, Adam and Citro, Craig and Ameisen, Emmanuel and Jones, Andy and Cunningham, Hoagy and Turner, Nicholas L and McDougall, Callum and MacDiarmid, Monte and Freeman, C. Daniel and Sumers, Theodore R. and Rees, Edward and Batson, Joshua and ...

    work page

  6. [6]

    2025 , type =

    McDougall, Callum and Conmy, Arthur and Kram. 2025 , type =

    work page 2025

  7. [7]

    2605.11887 , archivePrefix=

    Boyi Deng and Xu Wang and Yaoning Wang and Yu Wan and Yubo Ma and Baosong Yang and Haoran Wei and Jialong Tang and Huan Lin and Ruize Gao and Tianhao Li and Qian Cao and Xuancheng Ren and Xiaodong Deng and An Yang and Fei Huang and Dayiheng Liu and Jingren Zhou , year=. 2605.11887 , archivePrefix=

    arXiv

  8. [8]

    2023 , howpublished =

    Language models can explain neurons in language models , author=. 2023 , howpublished =

    2023

Show all 38 references
  1. [9]

    2023 , url =

    Lin, Johnny , title =. 2023 , url =

    2023

  2. [10]

    Proceedings of the 42nd International Conference on Machine Learning , pages =

    Paulo, Gon. Proceedings of the 42nd International Conference on Machine Learning , pages =. 2025 , editor =

    2025

  3. [11]

    SAGE : An Agentic Explainer Framework for Interpreting SAE Features in Language Models

    Han, Jiaojiao and Xu, Wujiang and Jin, Mingyu and Du, Mengnan. SAGE : An Agentic Explainer Framework for Interpreting SAE Features in Language Models. Proceedings of the 19th Conference of the E uropean Chapter of the A ssociation for C omputational L inguistics (Volume 5: Ind...

    2026 doi

  4. [12]

    Enhancing Automated Interpretability with Output-Centric Feature Descriptions

    Gur-Arieh, Yoav and Mayan, Roy and Agassy, Chen and Geiger, Atticus and Geva, Mor. Enhancing Automated Interpretability with Output-Centric Feature Descriptions. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 20...

    2025 doi

  5. [13]

    Ma, George and Pfrommer, Samuel and Sojoudi, Somayeh , booktitle =

  6. [14]

    FADE : Why Bad Descriptions Happen to Good Features

    Puri, Bruno and Jain, Aakriti and Golimblevskaia, Elena and Kahardipraja, Patrick and Wiegand, Thomas and Samek, Wojciech and Lapuschkin, Sebastian. FADE : Why Bad Descriptions Happen to Good Features. Findings of the Association for Computational Linguistics: ACL 2025. 2025. ...

    2025 doi

  7. [15]

    McCann , year=

    Jordan F. McCann , year=. 2605.12874 , archivePrefix=

    arXiv

  8. [16]

    2018 , editor =

    Kim, Been and Wattenberg, Martin and Gilmer, Justin and Cai, Carrie and Wexler, James and Viegas, Fernanda and Sayres, Rory , booktitle =. 2018 , editor =

    2018

  9. [17]

    Unveiling L anguage- S pecific F eatures in L arge L anguage M odels via S parse A utoencoders

    Deng, Boyi and Wan, Yu and Yang, Baosong and Zhang, Yidan and Feng, Fuli. Unveiling L anguage- S pecific F eatures in L arge L anguage M odels via S parse A utoencoders. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pap...

    2025 doi

  10. [18]

    L ingua L ens: Towards Interpreting Linguistic Mechanisms of Large Language Models via Sparse Auto-Encoder

    Jing, Yi and Yao, Zijun and Guo, Hongzhu and Ran, Lingxu and Wang, Xiaozhi and Hou, Lei and Li, Juanzi. L ingua L ens: Towards Interpreting Linguistic Mechanisms of Large Language Models via Sparse Auto-Encoder. Proceedings of the 2025 Conference on Empirical Methods in Natura...

    2025 doi

  11. [19]

    Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steering

    Zhao, Haiyan and Wu, Xuansheng and Yang, Fan and Shen, Bo and Liu, Ninghao and Du, Mengnan. Denoising Concept Vectors with Sparse Autoencoders for Improved Language Model Steering. Findings of the A ssociation for C omputational L inguistics: EACL 2026. 2026. doi:10.18653/v1/2...

    2026 doi

  12. [20]

    Self-explaining

    Kharlapenko, Dmitrii and neverix and Nanda, Neel and Conmy, Arthur , year =. Self-explaining

  13. [21]

    2606.08496 , archivePrefix=

    Jingyi He and Haiyan Zhao and Ruxue Shi and Yanguang Liu and Xin Wang and Fei Sun and Mengnan Du , year=. 2606.08496 , archivePrefix=

    arXiv

  14. [22]

    2024 , editor =

    Ghandeharioun, Asma and Caciularu, Avi and Pearce, Adam and Dixon, Lucas and Geva, Mor , booktitle =. 2024 , editor =

    2024

  15. [23]

    2024 , editor =

    Chen, Haozhe and Vondrick, Carl and Mao, Chengzhi , booktitle =. 2024 , editor =

    2024

  16. [24]

    Transformer Circuits Thread , year=

    Lindsey, Jack , title=. Transformer Circuits Thread , year=

  17. [25]

    Pan, Alexander and Chen, Lijie and Steinhardt, Jacob , booktitle =

  18. [26]

    CoRR , volume =

    Adam Karvonen and James Chua and Cl. CoRR , volume =. 2025 , url =. doi:10.48550/ARXIV.2512.15674 , eprinttype =. 2512.15674 , timestamp =

    2025 doi

  19. [27]

    Li and Zifan Carl Guo and Vincent Huang and Jacob Steinhardt and Jacob Andreas , year=

    Belinda Z. Li and Zifan Carl Guo and Vincent Huang and Jacob Steinhardt and Jacob Andreas , year=. 2511.08579 , archivePrefix=

  20. [28]

    and Ameisen, Emmanuel and Chen, James and Kishylau, Dzmitry and Pearce, Adam and Tarng, Julius and Wu, Alex and Wu, Jeff and Zhang, Yang and Ziegler, Daniel M

    Fraser-Taliente, Kit and Kantamneni, Subhash and Ong, Euan and Mossing, Dan and Lu, Christina and Bogdan, Paul C. and Ameisen, Emmanuel and Chen, James and Kishylau, Dzmitry and Pearce, Adam and Tarng, Julius and Wu, Alex and Wu, Jeff and Zhang, Yang and Ziegler, Daniel M. and...

  21. [29]

    2025 , url=

    Christopher Wolfram and Aaron Schein , booktitle=. 2025 , url=

    2025

  22. [30]

    2410.06981 , archivePrefix=

    Michael Lan and Philip Torr and Austin Meek and Ashkan Khakzar and David Krueger and Fazl Barez , year=. 2410.06981 , archivePrefix=

    arXiv

  23. [31]

    Transferring

    Chen, Alan and Merullo, Jack and Stolfo, Alessandro and Pavlick, Ellie , booktitle =. Transferring

  24. [32]

    Word Embeddings Are Steers for Language Models

    Han, Chi and Xu, Jialiang and Li, Manling and Fung, Yi and Sun, Chenkai and Jiang, Nan and Abdelzaher, Tarek and Ji, Heng. Word Embeddings Are Steers for Language Models. Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Pa...

    2024 doi

  25. [33]

    Multi-property Steering of Large Language Models with Dynamic Activation Composition

    Scalena, Daniel and Sarti, Gabriele and Nissim, Malvina. Multi-property Steering of Large Language Models with Dynamic Activation Composition. Proceedings of the 7th BlackboxNLP Workshop: Analyzing and Interpreting Neural Networks for NLP. 2024. doi:10.18653/v1/2024.blackboxnlp-1.34

    2024 doi

  26. [34]

    Zhu, Xudong and Khalili, Mohammad Mahdi and Zhu, Zhihui , booktitle =

  27. [35]

    Proceedings of the 29th Symposium on Operating Systems Principles , pages =

    Kwon, Woosuk and Li, Zhuohan and Zhuang, Siyuan and Sheng, Ying and Zheng, Lianmin and Yu, Cody Hao and Gonzalez, Joseph and Zhang, Hao and Stoica, Ion , title =. Proceedings of the 29th Symposium on Operating Systems Principles , pages =. 2023 , isbn =. doi:10.1145/3600006.36...

    2023

  28. [36]

    Advances in Neural Information Processing Systems , doi =

    Penedo, Guilherme and Kydl\'. Advances in Neural Information Processing Systems , doi =

  29. [37]

    Transformers: State-of-the-Art Natural Language Processing

    Wolf, Thomas and Debut, Lysandre and Sanh, Victor and Chaumond, Julien and Delangue, Clement and Moi, Anthony and Cistac, Pierric and Rault, Tim and Louf, Remi and Funtowicz, Morgan and Davison, Joe and Shleifer, Sam and von Platen, Patrick and Ma, Clara and Jernite, Yacine an...

    2020

  30. [38]

    Learning

    Bussmann, Bart and Nabeshima, Noa and Karvonen, Adam and Nanda, Neel , booktitle =. Learning. 2025 , editor =

    2025

Pith tools

Reviewed August 14, 2026 · model on record in the stance chip above.