{"work":{"id":"d525fe06-5560-4e97-86fc-7a0e551f5b17","openalex_id":null,"doi":null,"arxiv_id":"2308.10248","raw_key":null,"title":"Steering Language Models With Activation Engineering","authors":null,"authors_text":"Alexander Matt Turner, Lisa Thiergart, Gavin Leech, David Udell, Juan J. Vazquez, Ulisse Mini","year":2023,"venue":"cs.CL","abstract":"Prompt engineering and finetuning aim to maximize language model performance on a given metric (like toxicity reduction). However, these methods do not fully elicit a model's capabilities. To reduce this gap, we introduce activation engineering: the inference-time modification of activations in order to control (or steer) model outputs. Specifically, we introduce the Activation Addition (ActAdd) technique, which contrasts the intermediate activations on prompt pairs (such as \"Love\" versus \"Hate\") to compute a steering vector (Subramani et al. 2022). By tactically adding in e.g. the \"Love\" - \"Hate\" steering vector during the forward pass, we achieve SOTA on negative-to-positive sentiment shift and detoxification using models including LLaMA-3 and OPT. ActAdd yields inference-time control over high-level output properties (like topic and sentiment) while preserving performance on off-target tasks. ActAdd is lightweight: it does not require any machine optimization and works with a single pair of data points, which enables rapid iteration over steering. ActAdd demonstrates the power of activation engineering.","external_url":"https://arxiv.org/abs/2308.10248","cited_by_count":null,"metadata_source":"pith","metadata_fetched_at":"2026-05-25T07:05:26.771610+00:00","pith_arxiv_id":"2308.10248","created_at":"2026-05-08T21:29:13.729021+00:00","updated_at":"2026-06-05T21:23:00.469572+00:00","title_quality_ok":true,"display_title":"Steering Language Models With Activation Engineering","render_title":"Steering Language Models With Activation Engineering"},"hub":{"state":{"work_id":"d525fe06-5560-4e97-86fc-7a0e551f5b17","tier":"super_hub","tier_reason":"100+ Pith inbound or 10,000+ external citations","pith_inbound_count":115,"external_cited_by_count":null,"distinct_field_count":8,"first_pith_cited_at":"2023-11-07T01:59:11+00:00","last_pith_cited_at":"2026-05-22T01:52:04+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":"2026-06-12T03:59:19.121432+00:00","tier_text":"super_hub"},"tier":"super_hub","role_counts":[{"context_role":"background","n":16},{"context_role":"baseline","n":2},{"context_role":"method","n":2},{"context_role":"dataset","n":1}],"polarity_counts":[{"context_polarity":"background","n":13},{"context_polarity":"unclear","n":3},{"context_polarity":"baseline","n":2},{"context_polarity":"use_method","n":2},{"context_polarity":"use_dataset","n":1}],"runs":{"ask_index":{"job_type":"ask_index","status":"succeeded","result":{"title":"Steering Language Models With Activation Engineering","claims":[{"claim_text":"Prompt engineering and finetuning aim to maximize language model performance on a given metric (like toxicity reduction). However, these methods do not fully elicit a model's capabilities. To reduce this gap, we introduce activation engineering: the inference-time modification of activations in order to control (or steer) model outputs. Specifically, we introduce the Activation Addition (ActAdd) technique, which contrasts the intermediate activations on prompt pairs (such as \"Love\" versus \"Hate\") to compute a steering vector (Subramani et al. 2022). By tactically adding in e.g. the \"Love\" - \"H","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"[6] find that the default Assistant persona derives from an amalgamation of character archetypes learned during pretraining, suggesting that post-training steers models toward a particular region of a pre-existing persona space rather than constructing the Assistant from scratch. Theory of mind in LLMs.LLMs demonstrate substantial capacity to model mental states. Strachan et al. [34] found GPT-4 performs at or above human levels on some theory of mind measures, and Street et al. [35] showed GPT-","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"ference on Language Modeling, 2024. URL https: //openreview.net/forum?id=nGCMLATBit. [65] Ziqian Bi, Keyu Chen, Chiung-Yi Tseng, Danyang Zhang, Tianyang Wang, Hongying Luo, Lu Chen, Jun- ming Huang, Jibin Guan, Junfeng Hao, Xinyuan Song, and Junhao Song. Is gpt-oss good? a comprehensive evaluation of openai's latest open source models, 2025. URLhttps://arxiv.org/abs/2508.12461. [66] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInternational Confer- ence on Learn","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"in Section 3.4 (Safety Prompt) of MM-SafetyBench [ 52], incorporates a safety prompt prepended to the model input to mitigate unsafe behaviors. ActAdd [50, 75, 101] directly ed- its the target activations by removing the undesirable concept vectors d∈ I − (notation in §3.1) and adding desirable con- cept vectors d∈ I + following the strength settings in [101] and [50]. MOP applies the oblique projection from con- cept engineering [56] to the decoder transformer in MLLMs. We further evaluate the ","claim_type":"baseline","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Steering language models with activation engineering.arXiv preprint arXiv:2308.10248, 2023. 13 [56] Muhammed Ustaomeroglu, Baris Askin, Gauri Joshi, Carlee Joe-Wong, and Guannan Qu. Internal planning in language models: Characterizing horizon and branch awareness. In The Fourteenth International Conference on Learning Representations, 2026. URL https: //openreview.net/forum?id=dqGWQdFdTC. [57] Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretabili","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Timescale Consciousness Mechanism Example References Short-term Explicit Prompt-Based Context GPT-2 [22], GPT-3 [ 23], Prefix-Tuning [ 24], Prompt-Tuning [25], P-Tuning [26, 27], InstructGPT [28] Implicit Key-Value Cache Mechanism vLLM [ 29], StreamingLLM[ 30], H2O[ 31], LESS [ 32], KVQuant [33], RetrievalAttention [34], Memory 3 [1] Hidden State Steering Steer [ 35], ICV [ 36], ActAdd [ 37], StyleVec [38], CAA [ 39], FreeCtrl [40], EasyEdit2 [41] Activation Circuit Modula- tion SAC [42], DESTEI","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"},{"claim_text":"No prior work connects Damasio to LLMs or uses the model's own internal representa- tions (SAE features) as somatic marker analogues. Mechanistic Interpretability.We build on Bricken et al. [ 6] for SAEs, Rajamanoharan et al. [ 16] for the JumpReLU architecture, Gemma Scope [ 12, 13] for pretrained features, and activa- tion steering methods including ActAdd [ 22], con- trastive activation addition [ 17], and SAE-targeted steering [7]. 3 Method Our approach has four stages. First, wediscover whi","claim_type":"method","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Steering Language Models With Activation Engineering because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (16 contexts).","role_counts":[{"n":16,"context_role":"background"},{"n":2,"context_role":"baseline"},{"n":2,"context_role":"method"},{"n":1,"context_role":"dataset"}]},"error":null,"updated_at":"2026-05-21T04:42:33.203480+00:00"},"author_expand":{"job_type":"author_expand","status":"succeeded","result":{"authors_linked":[{"id":"81159144-209c-459a-8e7e-e36492783c23","orcid":null,"display_name":"Alexander Matt Turner"},{"id":"66b221ce-7e8c-4796-baf3-66b4380566c9","orcid":null,"display_name":"Lisa Thiergart"},{"id":"d14f96dd-10cf-4130-a704-a70a60c44e55","orcid":null,"display_name":"Gavin Leech"},{"id":"e89fb1f6-aeb1-48a7-8e0d-eeecceafaab1","orcid":null,"display_name":"David Udell"},{"id":"781cec06-aca1-4809-8253-4dddb914c661","orcid":null,"display_name":"Juan J. Vazquez"},{"id":"f1b9e5dc-18b9-4625-9090-f498377ab490","orcid":null,"display_name":"Ulisse Mini"}]},"error":null,"updated_at":"2026-05-21T04:42:33.834653+00:00"},"context_extract":{"job_type":"context_extract","status":"succeeded","result":{"enqueued_papers":25},"error":null,"updated_at":"2026-05-14T07:57:53.482638+00:00"},"graph_features":{"job_type":"graph_features","status":"succeeded","result":{"co_cited":[{"title":"Representation Engineering: A Top-Down Approach to AI Transparency","work_id":"45b326e2-e962-41a5-a542-2559e103a19b","shared_citers":44},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":28},{"title":"Steering Llama 2 via Contrastive Activation Addition","work_id":"3317feaa-e788-45fc-95aa-4ea20028b55b","shared_citers":18},{"title":"Persona Vectors: Monitoring and Controlling Character Traits in Language Models","work_id":"cf32dbef-9132-4648-abcb-0ebf3ac3af80","shared_citers":16},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":16},{"title":"The Linear Representation Hypothesis and the Geometry of Large Language Models","work_id":"a7b44adc-f2c2-4420-a27d-8ade97dd3b75","shared_citers":16},{"title":"Qwen2.5 Technical Report","work_id":"d8432992-4980-4a81-85c7-9fa2c2b87f85","shared_citers":13},{"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":13},{"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","work_id":"51960d72-c69f-4db8-8efd-e90e8b4d9524","shared_citers":11},{"title":"Mistral 7B","work_id":"eb5e1305-ad11-4875-ad8d-ad8b8f697599","shared_citers":10},{"title":"Refusal in Language Models Is Mediated by a Single Direction","work_id":"fbb9538d-8e58-4902-9fbd-b11f044bc2d5","shared_citers":10},{"title":"Steering llama 2 via contrastive activation addition","work_id":"1c681d32-40e4-4e00-b5a1-ddec8430f6ca","shared_citers":10},{"title":"Eliciting Latent Predictions from Transformers with the Tuned Lens","work_id":"a127314f-7424-488f-b6d7-8214650c420f","shared_citers":9},{"title":"Llama 2: Open Foundation and Fine-Tuned Chat Models","work_id":"68a5177f-d644-44c1-bd4f-4e5278c22f5d","shared_citers":9},{"title":"2024 , month = feb, number =","work_id":"dbc5264a-aabf-4346-8cbe-123e943b13b8","shared_citers":8},{"title":"Gemma 2: Improving Open Language Models at a Practical Size","work_id":"4dd94e2f-2b27-4cbf-88a0-4910f0772a57","shared_citers":8},{"title":"Inference-time intervention: Eliciting truthful answers from a language model","work_id":"cea287bd-7a03-4e95-8b27-f9db5f278697","shared_citers":8},{"title":"Linear representations of sentiment in large language models","work_id":"6cb3c7a7-3301-449f-97b9-7e047edafdf9","shared_citers":8},{"title":"Understanding intermediate layers using linear classifier probes","work_id":"bdc944db-4be2-44f7-950b-eaef12fab00e","shared_citers":8},{"title":"Toy Models of Superposition","work_id":"43875dbe-bc2d-4ab5-af63-744411533ff7","shared_citers":7},{"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","shared_citers":7},{"title":"AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models","work_id":"3b676de6-edef-4976-a8b5-082d4ff50867","shared_citers":6},{"title":"Discovering latent knowledge in language models without supervision","work_id":"a12b68bd-76a4-4837-ac7c-3ed5a60010d3","shared_citers":6},{"title":"Gemma: Open Models Based on Gemini Research and Technology","work_id":"a9ea2870-df28-40b8-a9e0-a7e9a116f793","shared_citers":6}],"time_series":[{"n":2,"year":2023},{"n":1,"year":2024},{"n":1,"year":2025},{"n":69,"year":2026}],"dependency_candidates":[]},"error":null,"updated_at":"2026-05-14T07:57:53.540589+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":"2026-05-14T07:57:45.088062+00:00"},"role_polarity":{"job_type":"role_polarity","status":"succeeded","result":{"title":"Steering Language Models With Activation Engineering","claims":[{"claim_text":"Prompt engineering and finetuning aim to maximize language model performance on a given metric (like toxicity reduction). However, these methods do not fully elicit a model's capabilities. To reduce this gap, we introduce activation engineering: the inference-time modification of activations in order to control (or steer) model outputs. Specifically, we introduce the Activation Addition (ActAdd) technique, which contrasts the intermediate activations on prompt pairs (such as \"Love\" versus \"Hate\") to compute a steering vector (Subramani et al. 2022). By tactically adding in e.g. the \"Love\" - \"H","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"[6] find that the default Assistant persona derives from an amalgamation of character archetypes learned during pretraining, suggesting that post-training steers models toward a particular region of a pre-existing persona space rather than constructing the Assistant from scratch. Theory of mind in LLMs.LLMs demonstrate substantial capacity to model mental states. Strachan et al. [34] found GPT-4 performs at or above human levels on some theory of mind measures, and Street et al. [35] showed GPT-","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"ference on Language Modeling, 2024. URL https: //openreview.net/forum?id=nGCMLATBit. [65] Ziqian Bi, Keyu Chen, Chiung-Yi Tseng, Danyang Zhang, Tianyang Wang, Hongying Luo, Lu Chen, Jun- ming Huang, Jibin Guan, Junfeng Hao, Xinyuan Song, and Junhao Song. Is gpt-oss good? a comprehensive evaluation of openai's latest open source models, 2025. URLhttps://arxiv.org/abs/2508.12461. [66] Diederik P. Kingma and Jimmy Ba. Adam: A method for stochastic optimization. InInternational Confer- ence on Learn","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"in Section 3.4 (Safety Prompt) of MM-SafetyBench [ 52], incorporates a safety prompt prepended to the model input to mitigate unsafe behaviors. ActAdd [50, 75, 101] directly ed- its the target activations by removing the undesirable concept vectors d∈ I − (notation in §3.1) and adding desirable con- cept vectors d∈ I + following the strength settings in [101] and [50]. MOP applies the oblique projection from con- cept engineering [56] to the decoder transformer in MLLMs. We further evaluate the ","claim_type":"baseline","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Steering language models with activation engineering.arXiv preprint arXiv:2308.10248, 2023. 13 [56] Muhammed Ustaomeroglu, Baris Askin, Gauri Joshi, Carlee Joe-Wong, and Guannan Qu. Internal planning in language models: Characterizing horizon and branch awareness. In The Fourteenth International Conference on Learning Representations, 2026. URL https: //openreview.net/forum?id=dqGWQdFdTC. [57] Kevin Ro Wang, Alexandre Variengien, Arthur Conmy, Buck Shlegeris, and Jacob Steinhardt. Interpretabili","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Timescale Consciousness Mechanism Example References Short-term Explicit Prompt-Based Context GPT-2 [22], GPT-3 [ 23], Prefix-Tuning [ 24], Prompt-Tuning [25], P-Tuning [26, 27], InstructGPT [28] Implicit Key-Value Cache Mechanism vLLM [ 29], StreamingLLM[ 30], H2O[ 31], LESS [ 32], KVQuant [33], RetrievalAttention [34], Memory 3 [1] Hidden State Steering Steer [ 35], ICV [ 36], ActAdd [ 37], StyleVec [38], CAA [ 39], FreeCtrl [40], EasyEdit2 [41] Activation Circuit Modula- tion SAC [42], DESTEI","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"},{"claim_text":"No prior work connects Damasio to LLMs or uses the model's own internal representa- tions (SAE features) as somatic marker analogues. Mechanistic Interpretability.We build on Bricken et al. [ 6] for SAEs, Rajamanoharan et al. [ 16] for the JumpReLU architecture, Gemma Scope [ 12, 13] for pretrained features, and activa- tion steering methods including ActAdd [ 22], con- trastive activation addition [ 17], and SAE-targeted steering [7]. 3 Method Our approach has four stages. First, wediscover whi","claim_type":"method","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Steering Language Models With Activation Engineering because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (16 contexts).","role_counts":[{"n":16,"context_role":"background"},{"n":2,"context_role":"baseline"},{"n":2,"context_role":"method"},{"n":1,"context_role":"dataset"}]},"error":null,"updated_at":"2026-05-21T04:42:33.837624+00:00"},"summary_claims":{"job_type":"summary_claims","status":"succeeded","result":{"title":"Steering Language Models With Activation Engineering","claims":[{"claim_text":"Prompt engineering and finetuning aim to maximize language model performance on a given metric (like toxicity reduction). However, these methods do not fully elicit a model's capabilities. To reduce this gap, we introduce activation engineering: the inference-time modification of activations in order to control (or steer) model outputs. Specifically, we introduce the Activation Addition (ActAdd) technique, which contrasts the intermediate activations on prompt pairs (such as \"Love\" versus \"Hate\") to compute a steering vector (Subramani et al. 2022). By tactically adding in e.g. the \"Love\" - \"H","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks Steering Language Models With Activation Engineering because it crossed a citation-hub threshold.","role_counts":[]},"error":null,"updated_at":"2026-05-14T07:57:40.311225+00:00"}},"summary":{"title":"Steering Language Models With Activation Engineering","claims":[{"claim_text":"Prompt engineering and finetuning aim to maximize language model performance on a given metric (like toxicity reduction). However, these methods do not fully elicit a model's capabilities. To reduce this gap, we introduce activation engineering: the inference-time modification of activations in order to control (or steer) model outputs. Specifically, we introduce the Activation Addition (ActAdd) technique, which contrasts the intermediate activations on prompt pairs (such as \"Love\" versus \"Hate\") to compute a steering vector (Subramani et al. 2022). By tactically adding in e.g. the \"Love\" - \"H","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks Steering Language Models With Activation Engineering because it crossed a citation-hub threshold.","role_counts":[]},"graph":{"co_cited":[{"title":"Representation Engineering: A Top-Down Approach to AI Transparency","work_id":"45b326e2-e962-41a5-a542-2559e103a19b","shared_citers":44},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":28},{"title":"Steering Llama 2 via Contrastive Activation Addition","work_id":"3317feaa-e788-45fc-95aa-4ea20028b55b","shared_citers":18},{"title":"Persona Vectors: Monitoring and Controlling Character Traits in Language Models","work_id":"cf32dbef-9132-4648-abcb-0ebf3ac3af80","shared_citers":16},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":16},{"title":"The Linear Representation Hypothesis and the Geometry of Large Language Models","work_id":"a7b44adc-f2c2-4420-a27d-8ade97dd3b75","shared_citers":16},{"title":"Qwen2.5 Technical Report","work_id":"d8432992-4980-4a81-85c7-9fa2c2b87f85","shared_citers":13},{"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":13},{"title":"Sparse Autoencoders Find Highly Interpretable Features in Language Models","work_id":"51960d72-c69f-4db8-8efd-e90e8b4d9524","shared_citers":11},{"title":"Mistral 7B","work_id":"eb5e1305-ad11-4875-ad8d-ad8b8f697599","shared_citers":10},{"title":"Refusal in Language Models Is Mediated by a Single Direction","work_id":"fbb9538d-8e58-4902-9fbd-b11f044bc2d5","shared_citers":10},{"title":"Steering llama 2 via contrastive activation addition","work_id":"1c681d32-40e4-4e00-b5a1-ddec8430f6ca","shared_citers":10},{"title":"Eliciting Latent Predictions from Transformers with the Tuned Lens","work_id":"a127314f-7424-488f-b6d7-8214650c420f","shared_citers":9},{"title":"Llama 2: Open Foundation and Fine-Tuned Chat Models","work_id":"68a5177f-d644-44c1-bd4f-4e5278c22f5d","shared_citers":9},{"title":"2024 , month = feb, number =","work_id":"dbc5264a-aabf-4346-8cbe-123e943b13b8","shared_citers":8},{"title":"Gemma 2: Improving Open Language Models at a Practical Size","work_id":"4dd94e2f-2b27-4cbf-88a0-4910f0772a57","shared_citers":8},{"title":"Inference-time intervention: Eliciting truthful answers from a language model","work_id":"cea287bd-7a03-4e95-8b27-f9db5f278697","shared_citers":8},{"title":"Linear representations of sentiment in large language models","work_id":"6cb3c7a7-3301-449f-97b9-7e047edafdf9","shared_citers":8},{"title":"Understanding intermediate layers using linear classifier probes","work_id":"bdc944db-4be2-44f7-950b-eaef12fab00e","shared_citers":8},{"title":"Toy Models of Superposition","work_id":"43875dbe-bc2d-4ab5-af63-744411533ff7","shared_citers":7},{"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","shared_citers":7},{"title":"AutoDAN: Generating Stealthy Jailbreak Prompts on Aligned Large Language Models","work_id":"3b676de6-edef-4976-a8b5-082d4ff50867","shared_citers":6},{"title":"Discovering latent knowledge in language models without supervision","work_id":"a12b68bd-76a4-4837-ac7c-3ed5a60010d3","shared_citers":6},{"title":"Gemma: Open Models Based on Gemini Research and Technology","work_id":"a9ea2870-df28-40b8-a9e0-a7e9a116f793","shared_citers":6}],"time_series":[{"n":2,"year":2023},{"n":1,"year":2024},{"n":1,"year":2025},{"n":69,"year":2026}],"dependency_candidates":[]},"authors":[{"id":"81159144-209c-459a-8e7e-e36492783c23","orcid":null,"display_name":"Alexander Matt Turner","source":"manual","import_confidence":0.72},{"id":"e89fb1f6-aeb1-48a7-8e0d-eeecceafaab1","orcid":null,"display_name":"David Udell","source":"manual","import_confidence":0.72},{"id":"d14f96dd-10cf-4130-a704-a70a60c44e55","orcid":null,"display_name":"Gavin Leech","source":"manual","import_confidence":0.72},{"id":"781cec06-aca1-4809-8253-4dddb914c661","orcid":null,"display_name":"Juan J. Vazquez","source":"manual","import_confidence":0.72},{"id":"66b221ce-7e8c-4796-baf3-66b4380566c9","orcid":null,"display_name":"Lisa Thiergart","source":"manual","import_confidence":0.72},{"id":"f1b9e5dc-18b9-4625-9090-f498377ab490","orcid":null,"display_name":"Ulisse Mini","source":"manual","import_confidence":0.72}]}}