{"work":{"id":"51960d72-c69f-4db8-8efd-e90e8b4d9524","openalex_id":null,"doi":null,"arxiv_id":"2309.08600","raw_key":null,"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","authors":null,"authors_text":"Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey","year":2023,"venue":"cs.LG","abstract":"One of the roadblocks to a better understanding of neural networks' internals is \\textit{polysemanticity}, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is \\textit{superposition}, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identify those directions, using sparse autoencoders to reconstruct the internal activations of a language model. These autoencoders learn sets of sparsely activating features that are more interpretable and monosemantic than directions identified by alternative approaches, where interpretability is measured by automated methods. Moreover, we show that with our learned set of features, we can pinpoint the features that are causally responsible for counterfactual behaviour on the indirect object identification task \\citep{wang2022interpretability} to a finer degree than previous decompositions. This work indicates that it is possible to resolve superposition in language models using a scalable, unsupervised method. Our method may serve as a foundation for future mechanistic interpretability work, which we hope will enable greater model transparency and steerability.","external_url":"https://arxiv.org/abs/2309.08600","cited_by_count":null,"metadata_source":"pith","metadata_fetched_at":"2026-05-25T06:25:23.804135+00:00","pith_arxiv_id":"2309.08600","created_at":"2026-05-08T21:09:11.854994+00:00","updated_at":"2026-05-25T06:25:23.804135+00:00","title_quality_ok":true,"display_title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","render_title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models"},"hub":{"state":{"work_id":"51960d72-c69f-4db8-8efd-e90e8b4d9524","tier":"super_hub","tier_reason":"100+ Pith inbound or 10,000+ external citations","pith_inbound_count":106,"external_cited_by_count":null,"distinct_field_count":12,"first_pith_cited_at":"2023-09-27T21:53:56+00:00","last_pith_cited_at":"2026-05-21T21:13:14+00:00","author_build_status":"needed","summary_status":"needed","contexts_status":"needed","graph_status":"needed","ask_index_status":"needed","reader_status":"not_needed","recognition_status":"not_needed","updated_at":"2026-05-26T06:36:17.866057+00:00","tier_text":"super_hub"},"tier":"super_hub","role_counts":[{"context_role":"background","n":18},{"context_role":"method","n":6},{"context_role":"baseline","n":1},{"context_role":"dataset","n":1}],"polarity_counts":[{"context_polarity":"background","n":14},{"context_polarity":"use_method","n":6},{"context_polarity":"unclear","n":3},{"context_polarity":"baseline","n":1},{"context_polarity":"support","n":1},{"context_polarity":"use_dataset","n":1}],"runs":{"ask_index":{"job_type":"ask_index","status":"succeeded","result":{"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","claims":[{"claim_text":"One of the roadblocks to a better understanding of neural networks' internals is \\textit{polysemanticity}, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is \\textit{superposition}, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identif","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"overlap, effect-size imbalance, or shared feature subspaces, making it essential to characterize these dynamics. In this section, we train SAEs on four DLMs spanning different objectives, architectures, and scales, and study how features encoding sentiment, topic, and style evolve and interact. 3.1 Training SAEs on Discrete Diffusion Models We train TopK sparse autoencoders [6, 26] on residual-stream activations from MDLM [2], SEDD [3], LLaDA [5], and DREAM [4]. The training layers are selected ","claim_type":"method","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 2126-2136. Association for Computational Linguistics, 2018. doi: 10.18653/V1/P18-1198. URLhttps://aclanthology.org/P18-1198/. [52] Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models.arXiv preprint arXiv:2309.08600, 2023. [53] Bartosz Cywiński, Bart Bussmann, Arthur Conmy, Josh Engels, Neel Nanda","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"InEuropean Conference on Computer Vision, pages 170-187. Springer, 2022. [9] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vidit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong Wang. Videoinr: Learning video implicit neural representation for continuous space-time super-resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2047-2057, 2022. [10] Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoen- coders f","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"unlearning [4], which aims to remove the influence of specific data or concepts from trained models. T2I diffusion unlearning methods can be broadly categorized into two approaches: modifying model parameters [11, 13, 27, 36] or suppressing the generation of target concepts without directly editing the model [9, 25]. Among them, SAeUron [9] employs sparse autoencoders (SAEs) [7] as auxiliary modules to extract interpretable latent features [8] and suppresses target concepts by ablating selected ","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"nonlinear functions, due to their expressive power guaranteed by the universal approximation theo- rem [3]. However, are MLPs the best nonlinear regressors we can build? Despite the prevalent use of MLPs, they have significant drawbacks. In transformers [4] for example, MLPs consume almost all non-embedding parameters and are typically less interpretable (relative to attention layers) without post-analysis tools [5]. We propose a promising alternative to MLPs, called Kolmogorov-Arnold Networks (","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Daniel Ziegler, Emmanuel Ameisen, Joshua Batson, Tim Belonax, et al. Auditing language models for hidden objectives.arXiv preprint arXiv:2503.10965, 2025. [78] Abhay Sheshadri, Aidan Ewart, Kai Fronsdal, Isha Gupta, Samuel R. Bowman, Sara Price, Samuel Marks, and Rowan Wang. Auditbench: Evaluating alignment auditing techniques on models with hidden behaviors, 2026. URLhttps://arxiv.org/abs/2602.22755. [79] Jan Leike, Miljan Martic, Victoria Krakovna, Pedro A Ortega, Tom Everitt, Andrew Lefrancq,","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Sparse Autoencoders Find Highly Interpretable Features in Language Models because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (18 contexts).","role_counts":[{"n":18,"context_role":"background"},{"n":6,"context_role":"method"},{"n":1,"context_role":"baseline"},{"n":1,"context_role":"dataset"}]},"error":null,"updated_at":"2026-05-21T18:43:02.632173+00:00"},"author_expand":{"job_type":"author_expand","status":"succeeded","result":{"authors_linked":[{"id":"17f5da87-6304-425b-88cd-ea9449038dd2","orcid":null,"display_name":"Hoagy Cunningham"},{"id":"c28c14b7-8e8d-457f-8a59-009fa944e58b","orcid":null,"display_name":"Aidan Ewart"},{"id":"8e2b5cbb-d191-4414-8a06-f54d16a3429a","orcid":null,"display_name":"Logan Riggs"},{"id":"6a9f0508-3474-410e-9f99-072f85d848b4","orcid":null,"display_name":"Robert Huben"},{"id":"c16df207-c24b-4227-917d-63af14a3f9ca","orcid":null,"display_name":"and Lee Sharkey"}]},"error":null,"updated_at":"2026-05-21T18:43:02.628201+00:00"},"context_extract":{"job_type":"context_extract","status":"succeeded","result":{"enqueued_papers":25},"error":null,"updated_at":"2026-05-14T09:58:45.756814+00:00"},"graph_features":{"job_type":"graph_features","status":"succeeded","result":{"co_cited":[{"title":"Scaling and evaluating sparse autoencoders","work_id":"f3faddeb-36ed-42bc-a7c9-9e764dc9b368","shared_citers":16},{"title":"Toy Models of Superposition","work_id":"43875dbe-bc2d-4ab5-af63-744411533ff7","shared_citers":16},{"title":"Steering Language Models With Activation Engineering","work_id":"d525fe06-5560-4e97-86fc-7a0e551f5b17","shared_citers":11},{"title":"Representation Engineering: A Top-Down Approach to AI Transparency","work_id":"45b326e2-e962-41a5-a542-2559e103a19b","shared_citers":10},{"title":"arXiv preprint arXiv:2408.05147 , year=","work_id":"90aa1ebe-ab2b-4462-adfd-047832493655","shared_citers":9},{"title":"Gemma 2: Improving Open Language Models at a Practical Size","work_id":"4dd94e2f-2b27-4cbf-88a0-4910f0772a57","shared_citers":9},{"title":"Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models","work_id":"fb24e7e7-f336-4706-bc2d-62d656b28d74","shared_citers":9},{"title":"Steering Llama 2 via Contrastive Activation Addition","work_id":"3317feaa-e788-45fc-95aa-4ea20028b55b","shared_citers":8},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":8},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":7},{"title":"The Linear Representation Hypothesis and the Geometry of Large Language Models","work_id":"a7b44adc-f2c2-4420-a27d-8ade97dd3b75","shared_citers":7},{"title":"In-context Learning and Induction Heads","work_id":"db2b0911-2758-4a2a-99dc-15b14b91bd5e","shared_citers":6},{"title":"K-sparse autoencoders","work_id":"b6541a25-f562-4327-b8d1-5aedaee60f5e","shared_citers":6},{"title":"Adam: A Method for Stochastic Optimization","work_id":"1910796d-9b52-4683-bf5c-de9632c1028b","shared_citers":5},{"title":"arXiv preprint arXiv:2501.16496 , year=","work_id":"f55f2189-55b1-4a1c-acfb-a5fa7bfa9e86","shared_citers":5},{"title":"Eliciting Latent Predictions from Transformers with the Tuned Lens","work_id":"a127314f-7424-488f-b6d7-8214650c420f","shared_citers":5},{"title":"Qwen2.5 Technical Report","work_id":"d8432992-4980-4a81-85c7-9fa2c2b87f85","shared_citers":5},{"title":"Stefan Heimersheim and Neel Nanda","work_id":"7aaa8958-6923-46a6-99cd-533ced88a8b2","shared_citers":5},{"title":"Tang, T., Luo, W., Huang, H., Zhang, D., Wang, X., Zhao, W","work_id":"4fd59f7a-f0c8-4d33-a3c8-cb7ca660cb17","shared_citers":5},{"title":"The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets","work_id":"400e017f-8643-4166-b6da-a75d4446da80","shared_citers":5},{"title":"Understanding intermediate layers using linear classifier probes","work_id":"bdc944db-4be2-44f7-950b-eaef12fab00e","shared_citers":5},{"title":"arXiv preprint arXiv:2308.09124 , year=","work_id":"25f5f724-b6d7-427f-a2f3-e8b72fd3b5e2","shared_citers":4},{"title":"Batchtopk sparse autoencoders","work_id":"812bea0e-6edc-46bc-8213-fbf0ebf4b662","shared_citers":4},{"title":"Discovering latent knowledge in language models without supervision","work_id":"a12b68bd-76a4-4837-ac7c-3ed5a60010d3","shared_citers":4}],"time_series":[{"n":3,"year":2024},{"n":1,"year":2025},{"n":58,"year":2026}],"dependency_candidates":[]},"error":null,"updated_at":"2026-05-14T10:08:51.418452+00:00"},"identity_refresh":{"job_type":"identity_refresh","status":"succeeded","result":{"items":[{"title":"Qwen3 Technical Report","outcome":"unchanged","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","resolver":"local_arxiv","confidence":0.98,"old_work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e"}],"counts":{"fixed":0,"merged":0,"unchanged":1,"quarantined":0,"needs_external_resolution":0},"errors":[],"attempted":1},"error":null,"updated_at":"2026-05-14T09:58:37.279299+00:00"},"role_polarity":{"job_type":"role_polarity","status":"succeeded","result":{"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","claims":[{"claim_text":"One of the roadblocks to a better understanding of neural networks' internals is \\textit{polysemanticity}, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is \\textit{superposition}, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identif","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"overlap, effect-size imbalance, or shared feature subspaces, making it essential to characterize these dynamics. In this section, we train SAEs on four DLMs spanning different objectives, architectures, and scales, and study how features encoding sentiment, topic, and style evolve and interact. 3.1 Training SAEs on Discrete Diffusion Models We train TopK sparse autoencoders [6, 26] on residual-stream activations from MDLM [2], SEDD [3], LLaDA [5], and DREAM [4]. The training layers are selected ","claim_type":"method","confidence":0.95,"evidence_strength":"citation_context"},{"claim_text":"Computational Linguistics, ACL 2018, Melbourne, Australia, July 15-20, 2018, Volume 1: Long Papers, pages 2126-2136. Association for Computational Linguistics, 2018. doi: 10.18653/V1/P18-1198. URLhttps://aclanthology.org/P18-1198/. [52] Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoencoders find highly interpretable features in language models.arXiv preprint arXiv:2309.08600, 2023. [53] Bartosz Cywiński, Bart Bussmann, Arthur Conmy, Josh Engels, Neel Nanda","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"InEuropean Conference on Computer Vision, pages 170-187. Springer, 2022. [9] Zeyuan Chen, Yinbo Chen, Jingwen Liu, Xingqian Xu, Vidit Goel, Zhangyang Wang, Humphrey Shi, and Xiaolong Wang. Videoinr: Learning video implicit neural representation for continuous space-time super-resolution. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2047-2057, 2022. [10] Hoagy Cunningham, Aidan Ewart, Logan Riggs, Robert Huben, and Lee Sharkey. Sparse autoen- coders f","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"unlearning [4], which aims to remove the influence of specific data or concepts from trained models. T2I diffusion unlearning methods can be broadly categorized into two approaches: modifying model parameters [11, 13, 27, 36] or suppressing the generation of target concepts without directly editing the model [9, 25]. Among them, SAeUron [9] employs sparse autoencoders (SAEs) [7] as auxiliary modules to extract interpretable latent features [8] and suppresses target concepts by ablating selected ","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"nonlinear functions, due to their expressive power guaranteed by the universal approximation theo- rem [3]. However, are MLPs the best nonlinear regressors we can build? Despite the prevalent use of MLPs, they have significant drawbacks. In transformers [4] for example, MLPs consume almost all non-embedding parameters and are typically less interpretable (relative to attention layers) without post-analysis tools [5]. We propose a promising alternative to MLPs, called Kolmogorov-Arnold Networks (","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Daniel Ziegler, Emmanuel Ameisen, Joshua Batson, Tim Belonax, et al. Auditing language models for hidden objectives.arXiv preprint arXiv:2503.10965, 2025. [78] Abhay Sheshadri, Aidan Ewart, Kai Fronsdal, Isha Gupta, Samuel R. Bowman, Sara Price, Samuel Marks, and Rowan Wang. Auditbench: Evaluating alignment auditing techniques on models with hidden behaviors, 2026. URLhttps://arxiv.org/abs/2602.22755. [79] Jan Leike, Miljan Martic, Victoria Krakovna, Pedro A Ortega, Tom Everitt, Andrew Lefrancq,","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Sparse Autoencoders Find Highly Interpretable Features in Language Models because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (18 contexts).","role_counts":[{"n":18,"context_role":"background"},{"n":6,"context_role":"method"},{"n":1,"context_role":"baseline"},{"n":1,"context_role":"dataset"}]},"error":null,"updated_at":"2026-05-21T18:43:02.634971+00:00"},"summary_claims":{"job_type":"summary_claims","status":"succeeded","result":{"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","claims":[{"claim_text":"One of the roadblocks to a better understanding of neural networks' internals is \\textit{polysemanticity}, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is \\textit{superposition}, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identif","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks Sparse Autoencoders Find Highly Interpretable Features in Language Models because it crossed a citation-hub threshold.","role_counts":[]},"error":null,"updated_at":"2026-05-14T10:08:40.527817+00:00"}},"summary":{"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","claims":[{"claim_text":"One of the roadblocks to a better understanding of neural networks' internals is \\textit{polysemanticity}, where neurons appear to activate in multiple, semantically distinct contexts. Polysemanticity prevents us from identifying concise, human-understandable explanations for what neural networks are doing internally. One hypothesised cause of polysemanticity is \\textit{superposition}, where neural networks represent more features than they have neurons by assigning features to an overcomplete set of directions in activation space, rather than to individual neurons. Here, we attempt to identif","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks Sparse Autoencoders Find Highly Interpretable Features in Language Models because it crossed a citation-hub threshold.","role_counts":[]},"graph":{"co_cited":[{"title":"Scaling and evaluating sparse autoencoders","work_id":"f3faddeb-36ed-42bc-a7c9-9e764dc9b368","shared_citers":16},{"title":"Toy Models of Superposition","work_id":"43875dbe-bc2d-4ab5-af63-744411533ff7","shared_citers":16},{"title":"Steering Language Models With Activation Engineering","work_id":"d525fe06-5560-4e97-86fc-7a0e551f5b17","shared_citers":11},{"title":"Representation Engineering: A Top-Down Approach to AI Transparency","work_id":"45b326e2-e962-41a5-a542-2559e103a19b","shared_citers":10},{"title":"arXiv preprint arXiv:2408.05147 , year=","work_id":"90aa1ebe-ab2b-4462-adfd-047832493655","shared_citers":9},{"title":"Gemma 2: Improving Open Language Models at a Practical Size","work_id":"4dd94e2f-2b27-4cbf-88a0-4910f0772a57","shared_citers":9},{"title":"Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models","work_id":"fb24e7e7-f336-4706-bc2d-62d656b28d74","shared_citers":9},{"title":"Steering Llama 2 via Contrastive Activation Addition","work_id":"3317feaa-e788-45fc-95aa-4ea20028b55b","shared_citers":8},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":8},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":7},{"title":"The Linear Representation Hypothesis and the Geometry of Large Language Models","work_id":"a7b44adc-f2c2-4420-a27d-8ade97dd3b75","shared_citers":7},{"title":"In-context Learning and Induction Heads","work_id":"db2b0911-2758-4a2a-99dc-15b14b91bd5e","shared_citers":6},{"title":"K-sparse autoencoders","work_id":"b6541a25-f562-4327-b8d1-5aedaee60f5e","shared_citers":6},{"title":"Adam: A Method for Stochastic Optimization","work_id":"1910796d-9b52-4683-bf5c-de9632c1028b","shared_citers":5},{"title":"arXiv preprint arXiv:2501.16496 , year=","work_id":"f55f2189-55b1-4a1c-acfb-a5fa7bfa9e86","shared_citers":5},{"title":"Eliciting Latent Predictions from Transformers with the Tuned Lens","work_id":"a127314f-7424-488f-b6d7-8214650c420f","shared_citers":5},{"title":"Qwen2.5 Technical Report","work_id":"d8432992-4980-4a81-85c7-9fa2c2b87f85","shared_citers":5},{"title":"Stefan Heimersheim and Neel Nanda","work_id":"7aaa8958-6923-46a6-99cd-533ced88a8b2","shared_citers":5},{"title":"Tang, T., Luo, W., Huang, H., Zhang, D., Wang, X., Zhao, W","work_id":"4fd59f7a-f0c8-4d33-a3c8-cb7ca660cb17","shared_citers":5},{"title":"The Geometry of Truth: Emergent Linear Structure in Large Language Model Representations of True/False Datasets","work_id":"400e017f-8643-4166-b6da-a75d4446da80","shared_citers":5},{"title":"Understanding intermediate layers using linear classifier probes","work_id":"bdc944db-4be2-44f7-950b-eaef12fab00e","shared_citers":5},{"title":"arXiv preprint arXiv:2308.09124 , year=","work_id":"25f5f724-b6d7-427f-a2f3-e8b72fd3b5e2","shared_citers":4},{"title":"Batchtopk sparse autoencoders","work_id":"812bea0e-6edc-46bc-8213-fbf0ebf4b662","shared_citers":4},{"title":"Discovering latent knowledge in language models without supervision","work_id":"a12b68bd-76a4-4837-ac7c-3ed5a60010d3","shared_citers":4}],"time_series":[{"n":3,"year":2024},{"n":1,"year":2025},{"n":58,"year":2026}],"dependency_candidates":[]},"authors":[{"id":"c28c14b7-8e8d-457f-8a59-009fa944e58b","orcid":null,"display_name":"Aidan Ewart","source":"manual","import_confidence":0.72},{"id":"c16df207-c24b-4227-917d-63af14a3f9ca","orcid":null,"display_name":"and Lee Sharkey","source":"manual","import_confidence":0.72},{"id":"17f5da87-6304-425b-88cd-ea9449038dd2","orcid":null,"display_name":"Hoagy Cunningham","source":"manual","import_confidence":0.72},{"id":"8e2b5cbb-d191-4414-8a06-f54d16a3429a","orcid":null,"display_name":"Logan Riggs","source":"manual","import_confidence":0.72},{"id":"6a9f0508-3474-410e-9f99-072f85d848b4","orcid":null,"display_name":"Robert Huben","source":"manual","import_confidence":0.72}]}}