Pith. sign in

REVIEW 3 major objections 5 minor 56 references

Insights into a radiology-specialised multimodal large language model with sparse autoencoders

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper applies a nested sparse autoencoder to the radiology model MAIRA-2 and shows a small set of its internal features are interpretable as clinical concepts, while steering generations along those features mostly produces off-target…

desk verdict Useful first SAE study of a medical MLLM with honest negative results, but the headline interpretability numbers are confounded by text predictability and a disclosed SAE configuration bug. read the letter →

arxiv 2507.12950 v2 pith:Z7WKV7QP submitted 2025-07-17 cs.LG

Kenza Bouzid , Shruthi Bannur , Felix Meissen , Daniel Coelho de Castro , Anton Schwaighofer , Javier Alvarez-Valle , Stephanie L. Hyland This is my paper
classification cs.LG
keywords sparseautoencodersmechanisticinterpretabilitymultimodallargelanguagemodelsradiologyreportgenerationfeaturesteeringchestX-rayautomatedMAIRA-2
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper asks whether mechanistic interpretability tools developed for large language models can open up a radiology-specialised multimodal model that writes chest X-ray reports. The authors train a Matryoshka sparse autoencoder on the internal representations of MAIRA-2 and find that a small minority of its 16,384 learned features — 288, or 1.8% — are consistently interpretable, capturing concepts like pleural effusion, cardiomegaly, line and tube placements, and even stylistic markers such as the word 'however'. Nearly half of the features resist interpretation entirely. The paper goes on to test whether adding or subtracting these features steers the model's reports as intended, and finds only partial, feature-dependent control: off-target changes are more common than on-target ones. If the approach is sound, it offers a route toward transparency and targeted control for medical image-language models, while exposing how much of the SAE interpretability pipeline remains unreliable in specialised domains.

What carries the argument

The load-bearing object is the Matryoshka sparse autoencoder (an SAE variant that trains several nested dictionary sizes at once to reduce feature splitting), trained on 34.7 million token representations drawn from MAIRA-2's residual stream at layer 15 and filtered to remove boilerplate and keep only the final image token of each chest X-ray. The decoder column of each activated feature doubles as the steering vector, added to the hidden states at every token during decoding. Interpretability is measured by an automated pipeline in which a text-only LLM first writes a description of the activation pattern and a second-stage LLM then predicts, from samples of text with image tokens replaced by '<image>', whether the feature is active; the detection $F_1$ of that prediction is the interpretability score. Steering success is scored by a third LLM judge that separates on-target changes from off-target changes in the generated report.

What would settle it

Take the features with detection $F_1$ above 0.85 and re-run the interpretation and scoring pipeline on samples where only the prompt tokens are shown, with the findings text withheld and images still replaced by '<image>'. If the clinical features, such as the pleural effusion or chest tube features, keep high detection $F_1$, they track something in the prompt context or in genuine image-driven computation; if their $F_1$ collapses toward 0.5, they were encoding the ground-truth findings text itself, and the claim that clinically relevant visual concepts were discovered would be substantially weakened.

Watch

Extended reading notes

Core claim

The central claim is that MAIRA-2's internal representations at the middle residual layer contain a sparse set of features that are human-interpretable through an automated LLM-based pipeline, and that these features include genuinely clinical concepts — medical devices, pathologies such as pleural effusion and cardiomegaly, longitudinal changes, and textual conventions. The claim is deliberately quantitative: of 16,384 features, 288 (1.8%) score above 0.75 detection $F_1$, while 7,500 (46%) score at or below chance, showing that interpretable features exist but are rare. The steering experiments then establish a second claim: manipulating these features by adding the SAE decoder vector to the residual stream produces directional control with mixed success, and the modal outcome is an off-target change rather than a clean on-target one. The authors present this as an initial map of the concepts a radiology-adapted multimodal model actually uses, with both the successes and the failure modes made explicit.

Load-bearing premise

The whole interpretation rests on a text-only LLM being able to explain what a multimodal feature means from samples in which the images were replaced with the placeholder '<image>'; because the findings text that the model reads and writes is present in those samples, a feature that scores as 'clinically meaningful' may actually be tracking a wording pattern rather than a visual concept the model learned.

Editorial extensions

If this is right

  • Some MAIRA-2 features are more fine-grained than standard label sets like CheXpert classes, so SAEs could supply concept vocabularies for auditing what a radiology model has learned.
  • Steering along a feature that activates frequently in training data is more likely to produce the desired on-target effect, suggesting activation frequency is a usable predictor of steerability.
  • Features that steer well in the positive direction also steer well in the negative direction, indicating the learned feature directions remain approximately linear well below their usual activation threshold.
  • Because most steering interventions produce off-target changes — including confabulated or omitted findings — SAE-based control is not yet safe for clinical use without further safeguards.
  • The public release of the trained SAEs and interpretations lets other researchers audit the same features and test alternative steering or interpretation schemes.

Reading between the lines

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

  • A direct test the paper leaves implicit: retrain or reinterpret the SAE on inputs stripped of the findings text (image tokens plus only the prompt), and check whether the 'clinical' features survive; if they do not, many high-$F_1$ features may be tracking textual style rather than visual concepts.
  • The correlation between activation frequency and steering success suggests that rarer features, including some of the most interpretable ones (a hemidiaphragm feature active on 0.03% of samples), may be poorly grounded in the model; targeting such features with counterfactual training data could test whether they can be made steerable.
  • The paper's negative result aligns with broader evidence that SAE steering underperforms prompting and fine-tuning on general text tasks; an extension would be to compare prompting-based control such as 'mention the pleural effusion' against feature steering on the same 67 features, which the paper does not do.
  • The strong correlation between positive and negative steering effects hints that the steering vector acts on a shared direction the model already uses in both directions; probing with intermediate coefficients rather than only ±10 could reveal whether the effect saturates or flips sign.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper trains a Matryoshka sparse autoencoder (SAE) on the layer-15 residual-stream activations of MAIRA-2, a radiology-specialised multimodal large language model, using token representations filtered from MIMIC-CXR. It then performs large-scale automated interpretability with GPT-4o, scoring each feature by detection F1 against held-out activations, and reports that 288 of 16,384 features (1.8%) are interpretable above 0.75, with descriptions covering medical devices, pathologies, temporal changes, and textual patterns. The paper further steers MAIRA-2 by adding SAE decoder vectors to all token positions and evaluates the effect with an LLM judge, finding that pure on-target changes are rare and that off-target changes dominate in most cases. The authors release the trained SAEs and automated interpretations.

Significance. If the interpretability results hold, this is one of the first large-scale mechanistic interpretability studies of a radiology-specific multimodal model, and the release of SAE checkpoints and feature interpretations is a valuable community resource. The paper is commendably transparent about negative results: the large proportion of uninterpretable features, the frequency of off-target steering effects, and the acknowledged methodological limitations. The systematic evaluation of steering side-effects, including clinically significant confabulations, is a useful contribution beyond the interpretability claims themselves. However, the significance depends on whether the high-F1 features genuinely reflect image-grounded concepts of MAIRA-2 rather than textual regularities of radiology reports, which is not established by the current pipeline.

major comments (3)
  1. [§3.6, Appendix C.1] The detection-F1 interpretability score is confounded by the next-100-characters context. Because the SAE is trained on token positions from both the prompt and the ground-truth findings text (§3.3), a feature that activates on a prompt token or on the retained final image token is almost always followed by the corresponding findings text. A text-only interpretation LLM can then label the feature with whatever pathology or device is mentioned in that future text and score high F1 even if the feature encodes nothing about the image. The authors explicitly note in Appendix C.1 that 'in many cases the input text to the model may already contain some part of the image description,' and Figure C.1 (adding a findings description does not improve interpretability) is exactly what this confound predicts. To support the claim in §4.1 that the 288 high-F1 features include clinically relevant visual concepts, the authors should provide a control that masks or removes the future text in the exemplars, or that evaluates interpretability on image-only token positions; otherwise the reported F1 distribution may reflect text prediction rather than multimodal representation content.
  2. [Appendix A.2] The acknowledged reversed Matryoshka group sizes mean that all downstream results in Section 4 were obtained with a configuration that is not the intended Matryoshka-SAE. The statement that this 'appears to have slightly reduced the overall interpretability' and the promise to 'quantify the impact on steering in a revision' leave the central quantitative claims unsupported for the correct configuration. The authors should either re-run the SAE training with the correct group ordering and repeat the interpretability and steering analyses, or provide the promised quantification and show that the proportions in Figures 2 and 3 are materially unchanged.
  3. [§3.5 and Table A.2] There is an internal inconsistency in the SAE size: Section 3.5 states that an expansion factor of ef=4 was selected, yielding 16,384 features, while Table A.2 lists 'Expansion factor 8'. For the 4096-dimensional activation space, ef=8 would give 32,768 features, which conflicts with the 16,384-feature dictionary used throughout the paper and in the release description. The authors should correct the table or the main text and ensure that the released checkpoint metadata matches the stated configuration.
minor comments (5)
  1. [§4.1] The phrase 'f730 and 16 others' is ambiguous: it is unclear whether these 17 features share identical low-quality descriptions or whether 16 other features have similar but not identical descriptions; please rephrase for clarity.
  2. [Figure 2] The x-axis label 'Interpretability (detection)' is vague; please specify that the metric is detection F1 as defined in §3.6.
  3. [Appendix C.1] The finding that truncating to 100 characters after the current token improves scores relative to 0, 10, or 250 characters is itself consistent with the future-text confound; the authors should discuss this dependence explicitly in the main text rather than only in the appendix.
  4. [Appendix E.1] The steering-feature selection deliberately restricts to F1 > 0.85 and adds highly activating features, so the percentages in Figure 3 are best-case estimates. The paper should state more prominently that these results are not representative of the full feature population, even though the selection is described in the appendix.
  5. [Table E.1] The feature description for f10643 ('Immediate notification of findings by telephone upon discovery') is a textual/operational pattern rather than an image-derived concept; given the paper's emphasis on clinical relevance, please clarify in the table or the main text which features are expected to be visual versus textual.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: detection F1 is scored against ground-truth SAE activations and steering against independent LLM judgment; the acknowledged text-context confound is a validity limitation, not a circular reduction.

full rationale

The claimed derivation chain is empirical and self-validating rather than circular. SAE features are trained on residual-stream activations of the released MAIRA-2 model (Sections 3.3-3.5). Interpretations are generated by GPT-4o from activation exemplars and then scored by detection F1 against held-out ground-truth feature activations (Section 3.6; Appendix C.3), so interpretability is not defined by the authors' own labels. Steering is evaluated by an LLM judge comparing original versus steered reports for on-target and off-target changes (Section 3.8), an independent behavioral test; the global coefficient α=10 and the iterative feature selection in Appendix E.1 are disclosed selection choices, not fitted parameters renamed as predictions, and the paper's main steering finding is the prevalence of off-target effects. The strongest candidate concern is Appendix C.1, where the interpretation LLM receives the next 100 characters and the paper notes the SAE is trained on prompt and target (Findings) tokens, so input text often already contains part of the image description; this could inflate F1 for features that track report-text regularities, and the paper's own textual-feature examples ('however', 'possible') show such features exist. However, this is an acknowledged validity limitation about visual grounding, not a circular step: the F1 label is the actual feature activation, and the same 100-character window is used for both interpretation and scoring in a held-out detection task, so the metric measures predictive power of the interpretation for activations. Self-citations (e.g., Bannur et al. for MAIRA-2) identify a publicly available model and are not load-bearing for the interpretability or steering results. No equation or definition reduces the paper's outputs to its inputs.

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

The central claims rest on standard SAE assumptions (features are locally linear, interpretable) and on the validity of LLM-based interpretation and evaluation. No mathematical axioms beyond standard ML background; no invented entities. The main free parameters are SAE hyperparameters and the steering coefficient, chosen by hand after experiments.

free parameters (5)
  • SAE expansion factor (ef) = 4
    Chosen over {2,4,8,16} as a trade-off between reconstruction quality, sparsity, and auto-interpretation compute cost; ef=16 gave best reconstruction. Section 3.5.
  • SAE active features per token (k) = 256
    Selected over {32,64,128,256} as producing the best reconstruction fidelity and fewest dead features. Section 3.5.
  • Steering coefficient alpha = 10 and -10
    Picked after testing {1,2.5,5,10,15,25,50,100} as balancing noticeable targeted changes with staying in-distribution. Section 3.7.
  • Representation normalization factor = 22.34
    Computed as the mean L2 norm of training representations, following Gao et al.; it is a data-derived scale that the SAE and downstream results depend on. Section 3.5.
  • Detection F1 threshold for steering feature selection = 0.85
    Features with F1 > 0.85 were chosen for steering, an arbitrary threshold that affects which features were tested. Appendix E.1.
assumptions (5)
  • domain assumption SAE features can disentangle monosemantic concepts in the residual stream of a multimodal LLM.
    The entire interpretability pipeline assumes sparse dictionary learning produces meaningful local features; the paper relies on prior SAE literature. Invoked in Section 3.4.
  • domain assumption Automated interpretability with GPT-4o, using text-only samples with image tokens replaced by '<image>', yields valid interpretations of multimodal features.
    The paper notes generalist GPT-4* CXR understanding is questionable but uses text-only context. Sections 3.6 and Appendix C.1.
  • domain assumption Detection F1 against LLM-predicted activations measures human interpretability.
    Follows Paulo et al. (2024) detection scoring; the paper treats F1 > 0.75 as 'consistently interpretable'. Section 3.6 and Figure 2.
  • domain assumption Layer 15 is the appropriate layer for semantically rich features.
    Motivated by prior works on middle layers; no empirical comparison on MAIRA-2 is provided. Section 3.3.
  • domain assumption Steering by adding a scaled SAE decoder vector to all residual stream token positions changes generation in the intended direction.
    Assumes the linear representation hypothesis; the paper itself discusses this may fail (Section 5). Section 3.7.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Insights into a radiology-specialised multimodal large language model with sparse autoencoders." pith.science (2026). https://pith.science/paper/Z7WKV7QP

@misc{pith2026250712950,
  author       = {Pith},
  title        = {Pith review of: Insights into a radiology-specialised multimodal large language model with sparse autoencoders},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Z7WKV7QP}},
  note         = {Machine review of arXiv:2507.12950}
}
read the original abstract

Interpretability can improve the safety, transparency and trust of AI models, which is especially important in healthcare applications where decisions often carry significant consequences. Mechanistic interpretability, particularly through the use of sparse autoencoders (SAEs), offers a promising approach for uncovering human-interpretable features within large transformer-based models. In this study, we apply Matryoshka-SAE to the radiology-specialised multimodal large language model, MAIRA-2, to interpret its internal representations. Using large-scale automated interpretability of the SAE features, we identify a range of clinically relevant concepts - including medical devices (e.g., line and tube placements, pacemaker presence), pathologies such as pleural effusion and cardiomegaly, longitudinal changes and textual features. We further examine the influence of these features on model behaviour through steering, demonstrating directional control over generations with mixed success. Our results reveal practical and methodological challenges, yet they offer initial insights into the internal concepts learned by MAIRA-2 - marking a step toward deeper mechanistic understanding and interpretability of a radiology-adapted multimodal large language model, and paving the way for improved model transparency. We release the trained SAEs and interpretations: https://huggingface.co/microsoft/maira-2-sae.

Figures

Figures reproduced from arXiv: 2507.12950 by the authors.

Figure 1
Figure 1. Illustration of the study. A: We train a Matryoshka-SAE using the output of the residual stream of MAIRA-2 at layer 15. Multimodal tokenization includes obtaining image tokens from an image encoder. We filter out (grey) token indices corresponding to intermediate image tokens and boilerplate parts of the prompt. B: We conduct automated interpretability with LLMs for both deriving interpretations and scoring interpre… view at source ↗
Figure 2
Figure 2. Interpretable features exist, but are rare. Using de￾tection F1 on the evaluation set as a measure of interpretability, we show that from 16,384 features, 288 (1.8%) score above 0.75, whereas 7,500 (46%) score below 0.5 (random performance for a balanced evaluation set). good balance between the appearance of noticeable targeted changes while remaining in-distribution. 3.8. Evaluating steering success Successful ste… view at source ↗
Figure 3
Figure 3. On-target and off-target effects of steering features fi with α = 10. The scores of the LLM judge are binarized at 0.1 and the bars show the proportion of samples where steering led only to on-target changes, only to off-target changes, both, or none at all. Results are shown for the full validation set including 7,906 studies. concept categories such as the CheXpert classes (Irvin et al., 2019). Among these highly … view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

56 extracted references · 22 canonical work pages

  1. [1]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...

  2. [2]

    C., and Castro, D

    Abdulaal, A., Fry, H., Monta \ n a-Brown, N., Ijishakin, A., Gao, J., Hyland, S., Alexander, D. C., and Castro, D. C. An X -ray is worth 15 features: Sparse autoencoders for interpretable radiology report generation. arXiv preprint arXiv:2410.03334, 2024

    arXiv

  3. [3]

    From Mechanistic Interpretability to Mechanistic Biology : Training , Evaluating , and Interpreting Sparse Autoencoders on Protein Language Models , February 2025

    Adams, E., Bai, L., Lee, M., Yu, Y., and AlQuraishi, M. From Mechanistic Interpretability to Mechanistic Biology : Training , Evaluating , and Interpreting Sparse Autoencoders on Protein Language Models , February 2025. URL https://www.biorxiv.org/content/10.1101/2025.02.06.636901v1. Pages: 2025.02.06.636901 Section: New Results

    work page doi:10.1101/2025.02.06.636901v1 2025

  4. [4]

    Anders, E., Neo, C., Hoelscher-Obermaier, J., and Howard, J. N. Sparse autoencoders find composed features in small toy models, 2024. URL https://www.lesswrong.com/posts/a5wwqza2cY3W7L9cj

    work page 2024

  5. [5]

    Bannur, S., Bouzid, K., Castro, D. C., Schwaighofer, A., Bond-Taylor, S., Ilse, M., Pérez-García, F., Salvatelli, V., Sharma, H., Meissen, F., Ranjit, M., Srivastav, S., Gong, J., Falck, F., Oktay, O., Thieme, A., Lungren, M. P., Wetscherek, M. T., Alvarez-Valle, J., and Hyland, S. L. MAIRA -2: Grounded Radiology Report Generation , June 2024. URL http://...

    arXiv 2024

  6. [6]

    Language models can explain neurons in language models

    Bills, S., Cammarata, N., Mossing, D., Tillman, H., Gao, L., Goh, G., Sutskever, I., Leike, J., Wu, J., and Saunders, W. Language models can explain neurons in language models. https://openaipublic.blob.core.windows.net/neuron-explainer/paper/index.html, 2023

    work page 2023

  7. [7]

    E., Hume, T., Carter, S., Henighan, T., and Olah, C

    Bricken, T., Templeton, A., Batson, J., Chen, B., Jermyn, A., Conerly, T., Turner, N., Anil, C., Denison, C., Askell, A., Lasenby, R., Wu, Y., Kravec, S., Schiefer, N., Maxwell, T., Joseph, N., Hatfield-Dodds, Z., Tamkin, A., Nguyen, K., McLean, B., Burke, J. E., Hume, T., Carter, S., Henighan, T., and Olah, C. Towards monosemanticity: Decomposing languag...

    2023

  8. [8]

    BatchTopK sparse autoencoders

    Bussmann, B., Leask, P., and Nanda, N. BatchTopK sparse autoencoders. arXiv preprint arXiv:2412.06410, 2024

    arXiv 2024

Show all 56 references
  1. [9]

    Learning Multi - Level Features with Matryoshka Sparse Autoencoders , March 2025

    Bussmann, B., Nabeshima, N., Karvonen, A., and Nanda, N. Learning Multi - Level Features with Matryoshka Sparse Autoencoders , March 2025. URL http://arxiv.org/abs/2503.17547. arXiv:2503.17547 [cs]

    2025 arXiv

  2. [10]

    Chanin, D., Wilken-Smith , J., Dulka, T., Bhatnagar, H., and Bloom, J. I. A is for absorption: Studying feature splitting and absorption in sparse autoencoders. In NeurIPS 2024 Workshop on Interpretable AI, December 2024. URL https://openreview.net/forum?id=Wzav8fesTL

    2024

  3. [11]

    Chen, Z., Varma, M., Delbrouck, J.-B., Paschali, M., Blankemeier, L., Van Veen, D., Valanarasu, J. M. J., Youssef, A., Cohen, J. P., Reis, E. P., et al. CheXagent : Towards a foundation model for chest x-ray interpretation. arXiv preprint arXiv:2401.12208, 2024

    2024 arXiv

  4. [12]

    E., Stoica, I., and Xing, E

    Chiang, W.-L., Li, Z., Lin, Z., Sheng, Y., Wu, Z., Zhang, H., Zheng, L., Zhuang, S., Zhuang, Y., Gonzalez, J. E., Stoica, I., and Xing, E. P. Vicuna: An open-source chatbot impressing GPT-4 with 90\ URL https://lmsys.org/blog/2023-03-30-vicuna/

    2023

  5. [13]

    Sparse autoencoders find highly interpretable features in language models

    Cunningham, H., Ewart, A., Riggs, L., Huben, R., and Sharkey, L. Sparse autoencoders find highly interpretable features in language models. In The Twelfth International Conference on Learning Representations, 2024. URL https://openreview.net/forum?id=F76bwRSLeK

    2024

  6. [14]

    Evaluating feature steering: A case study in mitigating social biases, 2024

    Durmus, E., Tamkin, A., Clark, J., Wei, J., Marcus, J., Batson, J., Handa, K., Lovitt, L., Tong, M., McCain, M., Rausch, O., Huang, S., Bowman, S., Ritchie, S., Henighan, T., and Ganguli, D. Evaluating feature steering: A case study in mitigating social biases, 2024. URL https...

    2024

  7. [15]

    A mathematical framework for transformer circuits

    Elhage, N., Nanda, N., Olsson, C., Henighan, T., Joseph, N., Mann, B., Askell, A., Bai, Y., Chen, A., Conerly, T., et al. A mathematical framework for transformer circuits. Transformer Circuits Thread, 2021

    2021

  8. [16]

    R., Todd, E., Brinkmann, J., Juang, C., Pal, K., Rager, C., Mueller, A., Marks, S., Sharma, A

    Fiotto-Kaufman, J., Loftus, A. R., Todd, E., Brinkmann, J., Juang, C., Pal, K., Rager, C., Mueller, A., Marks, S., Sharma, A. S., et al. NNsight and NDIF : Democratizing access to foundation model internals. arXiv preprint arXiv:2407.14561, 2024

    2024 arXiv

  9. [17]

    D., Tillman, H., Goh, G., Troll, R., Radford, A., Sutskever, I., Leike, J., and Wu, J

    Gao, L., la Tour, T. D., Tillman, H., Goh, G., Troll, R., Radford, A., Sutskever, I., Leike, J., and Wu, J. Scaling and evaluating sparse autoencoders, June 2024. URL http://arxiv.org/abs/2406.04093. arXiv:2406.04093 [cs]

    2024 arXiv

  10. [18]

    Enhancing automated interpretability with output-centric feature descriptions

    Gur-Arieh, Y., Mayan, R., Agassy, C., Geiger, A., and Geva, M. Enhancing automated interpretability with output-centric feature descriptions. arXiv preprint arXiv:2501.08319, 2025

    2025 arXiv

  11. [19]

    L lama S cope: Extracting millions of features from Llama-3.1-8B with sparse autoencoders, 2024

    He, Z., Shu, W., Ge, X., Chen, L., Wang, J., Zhou, Y., Liu, F., Guo, Q., Huang, X., Wu, Z., Jiang, Y.-G., and Qiu, X. L lama S cope: Extracting millions of features from Llama-3.1-8B with sparse autoencoders, 2024. URL https://arxiv.org/abs/2410.20526

    2024 arXiv

  12. [20]

    L., Bannur, S., Bouzid, K., Castro, D

    Hyland, S. L., Bannur, S., Bouzid, K., Castro, D. C., Ranjit, M., Schwaighofer, A., P \'e rez-Garc \' a, F., Salvatelli, V., Srivastav, S., Thieme, A., et al. MAIRA -1: A specialised large multimodal model for radiology report generation. arXiv:2311.13668, 2023. URL https://ar...

    2023 arXiv

  13. [21]

    CheXpert : A large chest radiograph dataset with uncertainty labels and expert comparison

    Irvin, J., Rajpurkar, P., Ko, M., Yu, Y., Ciurea-Ilcus, S., Chute, C., Marklund, H., Haghgoo, B., Ball, R., Shpanskaya, K., et al. CheXpert : A large chest radiograph dataset with uncertainty labels and expert comparison. In Proceedings of the AAAI conference on artificial int...

    2019

  14. [22]

    M., and Tan, C

    Jiang, Y., Chen, C., Nguyen, D., Mervak, B. M., and Tan, C. GPT - 4V cannot generate radiology reports yet. ArXiv, abs/2407.12176, 2024. URL https://api.semanticscholar.org/CorpusID:271244474

    2024 arXiv

  15. [23]

    Johnson, A. E. W., Pollard, T. J., Berkowitz, S. J., Mark, R. G., and Horng, S. MIMIC-CXR database (version 2.0.0). PhysioNet, 2019

    2019

  16. [24]

    The remarkable robustness of LLMs : Stages of inference? arXiv preprint arXiv:2406.19384, 2024

    Lad, V., Gurnee, W., and Tegmark, M. The remarkable robustness of LLMs : Stages of inference? arXiv preprint arXiv:2406.19384, 2024

    2024 arXiv

  17. [25]

    M., Patel, N., Shen, C., Martin, B., Eng, A., Shah, C., Grullon, S., and Juyal, D

    Le, N. M., Patel, N., Shen, C., Martin, B., Eng, A., Shah, C., Grullon, S., and Juyal, D. Learning biologically relevant features in a pathology foundation model using sparse autoencoders. In NeurIPS 2024 Workshop on Advancements In Medical Foundation Models, December 2024. UR...

    2024

  18. [26]

    Efficient sparse coding algorithms

    Lee, H., Battle, A., Raina, R., and Ng, A. Efficient sparse coding algorithms. Advances in neural information processing systems, 19, 2006

    2006

  19. [27]

    Inference-time intervention: Eliciting truthful answers from a language model

    Li, K., Patel, O., Vi \'e gas, F., Pfister, H., and Wattenberg, M. Inference-time intervention: Eliciting truthful answers from a language model. Advances in Neural Information Processing Systems, 36: 0 41451--41530, 2023 a

    2023

  20. [28]

    Dynamic graph enhanced contrastive learning for chest x-ray report generation

    Li, M., Lin, B., Chen, Z., Lin, H., Liang, X., and Chang, X. Dynamic graph enhanced contrastive learning for chest x-ray report generation. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 3334--3343, 2023 b

    2023

  21. [29]

    G emma S cope: Open sparse autoencoders everywhere all at once on G emma 2

    Lieberum, T., Rajamanoharan, S., Conmy, A., Smith, L., Sonnerat, N., Varma, V., Kram \'a r, J., Dragan, A., Shah, R., and Nanda, N. G emma S cope: Open sparse autoencoders everywhere all at once on G emma 2. arXiv preprint arXiv:2408.05147, 2024

    2024 arXiv

  22. [30]

    Liu, H., Li, C., Wu, Q., and Lee, Y. J. Visual instruction tuning. In Advances in Neural Information Processing Systems, volume 36, pp.\ 34892--34916, 2023

    2023

  23. [31]

    SAE - V : Interpreting Multimodal Models for Enhanced Alignment , February 2025

    Lou, H., Li, C., Ji, J., and Yang, Y. SAE - V : Interpreting Multimodal Models for Enhanced Alignment , February 2025. URL http://arxiv.org/abs/2502.17514. arXiv:2502.17514 [cs]

    2025 arXiv

  24. [32]

    dictionary\_learning

    Marks, S., Karvonen, A., and Mueller, A. dictionary\_learning. https://github.com/saprmarks/dictionary_learning, 2024

    2024

  25. [33]

    Robustly identifying concepts introduced during chat fine-tuning using crosscoders

    Minder, J., Dumas, C., Juang, C., Chugtai, B., and Nanda, N. Robustly identifying concepts introduced during chat fine-tuning using crosscoders. arXiv preprint arXiv:2504.02922, 2025

    2025

  26. [34]

    J., Tegmark, M., and de Witt, C

    Mudide, A., Engels, J., Michaud, E. J., Tegmark, M., and de Witt, C. S. Efficient dictionary learning with switch sparse autoencoders. arXiv preprint arXiv:2410.08201, 2024

    2024 arXiv

  27. [35]

    Steering language model refusal with sparse autoencoders

    O'Brien, K., Majercak, D., Fernandes, X., Edgar, R., Chen, J., Nori, H., Carignan, D., Horvitz, E., and Poursabzi-Sangde , F. Steering language model refusal with sparse autoencoders. arXiv preprint arXiv:2411.11296, 2024

    2024 arXiv

  28. [36]

    Sparse Autoencoders Learn Monosemantic Features in Vision - Language Models , April 2025

    Pach, M., Karthik, S., Bouniot, Q., Belongie, S., and Akata, Z. Sparse Autoencoders Learn Monosemantic Features in Vision - Language Models , April 2025. URL http://arxiv.org/abs/2504.02821. arXiv:2504.02821 [cs]

    2025

  29. [37]

    A concept-based explainability framework for large multimodal models

    Parekh, J., Khayatan, P., Shukor, M., Newson, A., and Cord, M. A concept-based explainability framework for large multimodal models. Advances in Neural Information Processing Systems, 37: 0 135783--135818, 2024

    2024

  30. [38]

    J., and Veitch, V

    Park, K., Choe, Y. J., and Veitch, V. The linear representation hypothesis and the geometry of large language models. In Proceedings of the 41st International Conference on Machine Learning, pp.\ 39643--39666. PMLR, July 2024. URL https://proceedings.mlr.press/v235/park24c.html

    2024

  31. [39]

    Automatically Interpreting Millions of Features in Large Language Models , December 2024

    Paulo, G., Mallen, A., Juang, C., and Belrose, N. Automatically Interpreting Millions of Features in Large Language Models , December 2024. URL http://arxiv.org/abs/2410.13928. arXiv:2410.13928 [cs]

    2024 arXiv

  32. [40]

    C., Schwaighofer, A., Lungren, M

    P \'e rez-Garc \' a, F., Sharma, H., Bond-Taylor, S., Bouzid, K., Salvatelli, V., Ilse, M., Bannur, S., Castro, D. C., Schwaighofer, A., Lungren, M. P., et al. Exploring scalable medical image encoders beyond text supervision. Nature Machine Intelligence, 7: 0 119--130, 2025. ...

    2025 doi

  33. [41]

    P., Senadeera, M., Jacobs, S., Coghlan, S., and Le, V

    Quinn, T. P., Senadeera, M., Jacobs, S., Coghlan, S., and Le, V. Trust and medical ai: the challenges we face and the expertise needed to overcome them. Journal of the American Medical Informatics Association, 28 0 (4): 0 890--894, 2021

    2021

  34. [42]

    R., Schwettmann, S., Wang, F., Rajaram, A., Hernandez, E., Andreas, J., and Torralba, A

    Shaham, T. R., Schwettmann, S., Wang, F., Rajaram, A., Hernandez, E., Andreas, J., and Torralba, A. A multimodal automated interpretability agent. In Forty-first International Conference on Machine Learning, 2024

    2024

  35. [43]

    and Zou, J

    Simon, E. and Zou, J. InterPLM : Discovering Interpretable Features in Protein Language Models via Sparse Autoencoders , November 2024. URL https://www.biorxiv.org/content/10.1101/2024.11.14.623630v1. Pages: 2024.11.14.623630 Section: New Results

    2024 doi

  36. [44]

    Sparse autoencoders for scientifically rigorous interpretation of vision models

    Stevens, S., Chao, W.-L., Berger-Wolf, T., and Su, Y. Sparse autoencoders for scientifically rigorous interpretation of vision models. arXiv preprint arXiv:2502.06755, 2025

    2025

  37. [45]

    L., McDougall, C., MacDiarmid, M., Freeman, C

    Templeton, A., Conerly, T., Marcus, J., Lindsey, J., Bricken, T., Chen, B., Pearce, A., Citro, C., Ameisen, E., Jones, A., Cunningham, H., Turner, N. L., McDougall, C., MacDiarmid, M., Freeman, C. D., Sumers, T. R., Rees, E., Batson, J., Jermyn, A., Carter, S., Olah, C., and H...

    2024

  38. [46]

    Towards generalist biomedical AI

    Tu, T., Azizi, S., Driess, D., Schaekermann, M., Amin, M., Chang, P.-C., Carroll, A., Lau, C., Tanno, R., Ktena, I., Palepu, A., Mustafa, B., Chowdhery, A., Liu, Y., Kornblith, S., Fleet, D., Mansfield, P., Prakash, S., Wong, R., Virmani, S., et al. Towards generalist biomedic...

    2024 doi

  39. [47]

    M., Thiergart, L., Leech, G., Udell, D., Vazquez, J

    Turner, A. M., Thiergart, L., Leech, G., Udell, D., Vazquez, J. J., Mini, U., and MacDiarmid, M. Steering language models with activation engineering. arXiv preprint arXiv:2308.10248, 2023

    2023 arXiv

  40. [48]

    Interpretability in the wild: a circuit for indirect object identification in GPT-2 small

    Wang, K., Variengien, A., Conmy, A., Shlegeris, B., and Steinhardt, J. Interpretability in the wild: a circuit for indirect object identification in GPT-2 small. arXiv preprint arXiv:2211.00593, 2022

    2022 arXiv

  41. [49]

    Metransformer: Radiology report generation by transformer with multiple learnable expert tokens

    Wang, Z., Liu, L., Wang, L., and Zhou, L. Metransformer: Radiology report generation by transformer with multiple learnable expert tokens. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp.\ 11558--11567, 2023

    2023

  42. [50]

    and Viégas, F

    Wattenberg, M. and Viégas, F. B. Relational composition in neural networks: A survey and call to action, 2024. URL https://arxiv.org/abs/2407.14662

    2024 arXiv

  43. [51]

    D., and Potts, C

    Wu, Z., Arora, A., Geiger, A., Wang, Z., Huang, J., Jurafsky, D., Manning, C. D., and Potts, C. AxBench : Steering LLMs ? Even Simple Baselines Outperform Sparse Autoencoders , March 2025. URL http://arxiv.org/abs/2501.17148. arXiv:2501.17148 [cs]

    2025 arXiv

  44. [52]

    Yan, Q., He, X., Yue, X., and Wang, X. E. Worse than random? an embarrassingly simple probing evaluation of large multimodal models in medical VQA . ArXiv, abs/2405.20421, 2024. URL https://api.semanticscholar.org/CorpusID:270199350

    2024 arXiv

  45. [53]

    Advancing multimodal medical capabilities of G emini

    Yang, L., Xu, S., Sellergren, A., Kohlberger, T., Zhou, Y., Ktena, I., Kiraly, A., Ahmed, F., Hormozdiari, F., Jaroensri, T., et al. Advancing multimodal medical capabilities of G emini. arXiv preprint arXiv:2405.03162, 2024

    2024 arXiv

  46. [54]

    T., Bajwa, J., Jacob, J., Pinnock, M

    Yildirim, N., Richardson, H., Wetscherek, M. T., Bajwa, J., Jacob, J., Pinnock, M. A., Harris, S., Coelho De Castro, D., Bannur, S., Hyland, S., et al. Multimodal healthcare AI : Identifying and designing clinically relevant vision-language applications for radiology. In Proce...

    2024

  47. [55]

    Large Multi -modal Models Can Interpret Features in Large Multi -modal Models , November 2024

    Zhang, K., Shen, Y., Li, B., and Liu, Z. Large Multi -modal Models Can Interpret Features in Large Multi -modal Models , November 2024. URL http://arxiv.org/abs/2411.14982. arXiv:2411.14982 [cs]

    2024

  48. [56]

    N., Topol, E

    Zhou, H.-Y., Adithan, S., Acosta, J. N., Topol, E. J., and Rajpurkar, P. A generalist learner for multifaceted medical image interpretation. arXiv preprint arXiv:2405.07988, 2024

    2024 arXiv

Pith tools

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