{"work":{"id":"1aabd84d-3779-4ba9-ba2f-15ce264a9b1e","openalex_id":"https://openalex.org/W2546505767","doi":"10.1136/bcr-2013-201554","arxiv_id":"2209.07858","raw_key":null,"title":"Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned","authors":null,"authors_text":"Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath","year":2022,"venue":"cs.CL","abstract":"We describe our early efforts to red team language models in order to simultaneously discover, measure, and attempt to reduce their potentially harmful outputs. We make three main contributions. First, we investigate scaling behaviors for red teaming across 3 model sizes (2.7B, 13B, and 52B parameters) and 4 model types: a plain language model (LM); an LM prompted to be helpful, honest, and harmless; an LM with rejection sampling; and a model trained to be helpful and harmless using reinforcement learning from human feedback (RLHF). We find that the RLHF models are increasingly difficult to red team as they scale, and we find a flat trend with scale for the other model types. Second, we release our dataset of 38,961 red team attacks for others to analyze and learn from. We provide our own analysis of the data and find a variety of harmful outputs, which range from offensive language to more subtly harmful non-violent unethical outputs. Third, we exhaustively describe our instructions, processes, statistical methodologies, and uncertainty about red teaming. We hope that this transparency accelerates our ability to work together as a community in order to develop shared norms, practices, and technical standards for how to red team language models.","external_url":"https://arxiv.org/abs/2209.07858","cited_by_count":664,"metadata_source":"pith","metadata_fetched_at":"2026-08-05T02:28:24.338817+00:00","pith_arxiv_id":"2209.07858","created_at":"2026-05-08T21:39:24.720020+00:00","updated_at":"2026-08-05T02:28:24.338817+00:00","title_quality_ok":true,"display_title":"Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned","render_title":"Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned"},"hub":{"state":{"work_id":"1aabd84d-3779-4ba9-ba2f-15ce264a9b1e","tier":"super_hub","tier_reason":"100+ Pith inbound or 10,000+ external citations","pith_inbound_count":150,"external_cited_by_count":664,"distinct_field_count":14,"first_pith_cited_at":"2022-07-11T22:59:39+00:00","last_pith_cited_at":"2026-07-08T15:39:10+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":"2026-08-21T01:39:40.226255+00:00","tier_text":"super_hub"},"tier":"super_hub","role_counts":[{"context_role":"background","n":24},{"context_role":"method","n":3},{"context_role":"dataset","n":1}],"polarity_counts":[{"context_polarity":"background","n":24},{"context_polarity":"use_method","n":2},{"context_polarity":"support","n":1},{"context_polarity":"use_dataset","n":1}],"runs":{"ask_index":{"job_type":"ask_index","status":"succeeded","result":{"title":"Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned","claims":[{"claim_text":"We describe our early efforts to red team language models in order to simultaneously discover, measure, and attempt to reduce their potentially harmful outputs. We make three main contributions. First, we investigate scaling behaviors for red teaming across 3 model sizes (2.7B, 13B, and 52B parameters) and 4 model types: a plain language model (LM); an LM prompted to be helpful, honest, and harmless; an LM with rejection sampling; and a model trained to be helpful and harmless using reinforcement learning from human feedback (RLHF). We find that the RLHF models are increasingly difficult to re","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"Retrieved from: https://blog.google/products/bard/google-bard-expansion-teens/. [74] Ganguli, D., Lovitt, L., Kernion, J., Askell, A., Bai, Y., Kadavath, S., Mann, B., Perez, E., Schiefer, N., Ndousse, K. & Jones, A. Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned. arXiv preprint arXiv:2209.07858 (2022). [75] OpenAI. GPT-4 Technical Report. arXiv preprint arXiv:2303.08774v4 (2023). [76] Bommasani, R., Klyman, K., Longpre, S., Kapoor, S., Maslej, N., X","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"exhibit harmful behaviors, hallucinations, leaking personal in- formation, and other shortcomings through adversarial probing. The models are susceptible to generating harmful responses even though they are aligned for safety [177, 178]. Red- teaming is a common approach to address illicit outputs, where the LLMs are prompted to generate harmful outputs [178, 179]. The dataset collected through red-teaming is used to fine-tune models for safety. While red-teaming largely relies on human annotato","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Conformalized survival analysis with adaptive cut-offs.Biometrika, 111(2):459-477, 2024. [15] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv preprint arXiv:2204.05862, 2022. [16] Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer","claim_type":"dataset","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"training and inference cost, making MoE central to the next gener- ation of LLMs [2, 10, 19, 39, 48, 49, 52]. At the same time, serving these increasingly large models is becoming bottlenecked not by arithmetic throughput, but by the cost of moving billions of parameters through the memory hier- archy [12]. This growing imbalance, known as the AI memory wall [17], has motivated the adoption of Processing-in-Memory (PIM) technology. PIM architectures embed lightweight compute units near DRAM bank","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"(GopherCite) [41], and general dialogue (Sparrow) [42]. Aside from applications, a number of seminal papers defined key areas for the future of RLHF, including those on: 1. Reward model over-optimization [43]: The ability for RL optimizers to over-fit to models trained on preference data, 2. Language models as a general area of study for alignment [23], and 3. Red teaming [44] - the process of assessing the safety of a language model. Work continued on refining RLHF for application to chat model","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"},{"claim_text":"Poesia, Davide Ghilardi, Anna Goldie, Federico Bianchi, Dan Jurafsky, and Christopher D Manning. h4rm3l: A language for composable jailbreak attack synthesis.arXiv preprint arXiv:2408.04811, 2024. 2 [11] Igor Fedorov, Kate Plawiak, Lemeng Wu, Tarek Elgamal, Naveen Suda, et al. Llama guard 3-1b-int4: Compact and 9 efficient safeguard for human-ai conversations.arXiv preprint arXiv:2411.17713, 2024. 3 [12] Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (24 contexts).","role_counts":[{"n":24,"context_role":"background"},{"n":3,"context_role":"method"},{"n":1,"context_role":"dataset"}]},"error":null,"updated_at":"2026-06-27T04:25:44.407079+00:00"},"author_expand":{"job_type":"author_expand","status":"succeeded","result":{"authors_linked":[{"id":"20146a94-3ea4-4ed5-8389-126a1613d880","orcid":null,"display_name":"Deep Ganguli"},{"id":"97c891e8-0ac2-40a8-8fb9-c49436fa1c87","orcid":null,"display_name":"Liane Lovitt"},{"id":"6ff3739a-293b-4060-bd4d-4c63b8f88381","orcid":null,"display_name":"Jackson Kernion"},{"id":"fa166f46-a486-47b9-b4c7-b3c0d3548ac0","orcid":null,"display_name":"Amanda Askell"},{"id":"f15a23ad-8224-4c50-a63d-24a766b946a8","orcid":null,"display_name":"Yuntao Bai"},{"id":"155c969c-3372-4b52-b881-15f051250399","orcid":null,"display_name":"Saurav Kadavath"}]},"error":null,"updated_at":"2026-06-27T04:25:44.403440+00:00"},"context_extract":{"job_type":"context_extract","status":"succeeded","result":{"enqueued_papers":25},"error":null,"updated_at":"2026-05-14T17:50:00.478760+00:00"},"graph_features":{"job_type":"graph_features","status":"succeeded","result":{"co_cited":[{"title":"Universal and Transferable Adversarial Attacks on Aligned Language Models","work_id":"3322fa86-1768-4677-8425-dd326b45e078","shared_citers":17},{"title":"Constitutional AI: Harmlessness from AI Feedback","work_id":"faaaa4e0-2676-4fac-a0b4-99aef10d2095","shared_citers":16},{"title":"Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback","work_id":"a1f2574b-a899-4713-be60-c87ba332656c","shared_citers":14},{"title":"Red Teaming Language Models with Language Models","work_id":"d1274c54-508f-42f9-aeb3-91db13f3a622","shared_citers":12},{"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","shared_citers":10},{"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","shared_citers":8},{"title":"Ethical and social risks of harm from Language Models","work_id":"b4ce1c45-ef69-445a-a872-dbb785b485e9","shared_citers":7},{"title":"HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal","work_id":"b0b0303f-2444-4789-a979-8153624312ff","shared_citers":7},{"title":"Llama 2: Open Foundation and Fine-Tuned Chat Models","work_id":"68a5177f-d644-44c1-bd4f-4e5278c22f5d","shared_citers":7},{"title":"Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations","work_id":"93844332-869b-448c-a1be-35466150b1b2","shared_citers":7},{"title":"Proximal Policy Optimization Algorithms","work_id":"240c67fe-d14d-4520-91c1-38a4e272ca19","shared_citers":7},{"title":"A General Language Assistant as a Laboratory for Alignment","work_id":"a43f9ea0-01be-47d5-b8ee-a1a9f73381c5","shared_citers":6},{"title":"Evaluating Large Language Models Trained on Code","work_id":"042493e9-b26f-4b4e-bbde-382072ca9b08","shared_citers":6},{"title":"Improving alignment of dialogue agents via targeted human judgements","work_id":"6ad5970e-7550-4ae8-a158-7084dec7e3cc","shared_citers":6},{"title":"LLaMA: Open and Efficient Foundation Language Models","work_id":"c018fc23-6f3f-4035-9d02-28a2173b2b9d","shared_citers":6},{"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","shared_citers":5},{"title":"Fine-Tuning Language Models from Human Preferences","work_id":"4f54aad1-f3b6-404f-b9c7-e21ba0a33b99","shared_citers":5},{"title":"Holistic Evaluation of Language Models","work_id":"cc02a01e-7218-47dc-8e66-3333e7e4adec","shared_citers":5},{"title":"Measuring Massive Multitask Language Understanding","work_id":"e87ec49a-544b-4ec8-8991-75298c64ff5e","shared_citers":5},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":5},{"title":"Scaling Laws for Neural Language Models","work_id":"b7dd8749-9c45-4977-ab9b-64478dce1ae8","shared_citers":5},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":5},{"title":"Training language models to follow instructions with human feedback","work_id":"52aff42f-4fa9-4fcf-bdb3-1459b9bebf65","shared_citers":5},{"title":"arXiv preprint arXiv:2305.15324 , year=","work_id":"12dff4e7-a81d-4b95-aba7-885a457042c6","shared_citers":4}],"time_series":[{"n":1,"year":2022},{"n":4,"year":2023},{"n":2,"year":2024},{"n":32,"year":2026}],"dependency_candidates":[]},"error":null,"updated_at":"2026-05-14T17:49:48.797617+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":"2026-05-14T17:49:23.201516+00:00"},"role_polarity":{"job_type":"role_polarity","status":"succeeded","result":{"title":"Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned","claims":[{"claim_text":"We describe our early efforts to red team language models in order to simultaneously discover, measure, and attempt to reduce their potentially harmful outputs. We make three main contributions. First, we investigate scaling behaviors for red teaming across 3 model sizes (2.7B, 13B, and 52B parameters) and 4 model types: a plain language model (LM); an LM prompted to be helpful, honest, and harmless; an LM with rejection sampling; and a model trained to be helpful and harmless using reinforcement learning from human feedback (RLHF). We find that the RLHF models are increasingly difficult to re","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"Retrieved from: https://blog.google/products/bard/google-bard-expansion-teens/. [74] Ganguli, D., Lovitt, L., Kernion, J., Askell, A., Bai, Y., Kadavath, S., Mann, B., Perez, E., Schiefer, N., Ndousse, K. & Jones, A. Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned. arXiv preprint arXiv:2209.07858 (2022). [75] OpenAI. GPT-4 Technical Report. arXiv preprint arXiv:2303.08774v4 (2023). [76] Bommasani, R., Klyman, K., Longpre, S., Kapoor, S., Maslej, N., X","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"exhibit harmful behaviors, hallucinations, leaking personal in- formation, and other shortcomings through adversarial probing. The models are susceptible to generating harmful responses even though they are aligned for safety [177, 178]. Red- teaming is a common approach to address illicit outputs, where the LLMs are prompted to generate harmful outputs [178, 179]. The dataset collected through red-teaming is used to fine-tune models for safety. While red-teaming largely relies on human annotato","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Conformalized survival analysis with adaptive cut-offs.Biometrika, 111(2):459-477, 2024. [15] Yuntao Bai, Andy Jones, Kamal Ndousse, Amanda Askell, Anna Chen, Nova DasSarma, Dawn Drain, Stanislav Fort, Deep Ganguli, Tom Henighan, et al. Training a helpful and harmless assistant with reinforcement learning from human feedback.arXiv preprint arXiv:2204.05862, 2022. [16] Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben Mann, Ethan Perez, Nicholas Schiefer","claim_type":"dataset","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"training and inference cost, making MoE central to the next gener- ation of LLMs [2, 10, 19, 39, 48, 49, 52]. At the same time, serving these increasingly large models is becoming bottlenecked not by arithmetic throughput, but by the cost of moving billions of parameters through the memory hier- archy [12]. This growing imbalance, known as the AI memory wall [17], has motivated the adoption of Processing-in-Memory (PIM) technology. PIM architectures embed lightweight compute units near DRAM bank","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"(GopherCite) [41], and general dialogue (Sparrow) [42]. Aside from applications, a number of seminal papers defined key areas for the future of RLHF, including those on: 1. Reward model over-optimization [43]: The ability for RL optimizers to over-fit to models trained on preference data, 2. Language models as a general area of study for alignment [23], and 3. Red teaming [44] - the process of assessing the safety of a language model. Work continued on refining RLHF for application to chat model","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"},{"claim_text":"Poesia, Davide Ghilardi, Anna Goldie, Federico Bianchi, Dan Jurafsky, and Christopher D Manning. h4rm3l: A language for composable jailbreak attack synthesis.arXiv preprint arXiv:2408.04811, 2024. 2 [11] Igor Fedorov, Kate Plawiak, Lemeng Wu, Tarek Elgamal, Naveen Suda, et al. Llama guard 3-1b-int4: Compact and 9 efficient safeguard for human-ai conversations.arXiv preprint arXiv:2411.17713, 2024. 3 [12] Deep Ganguli, Liane Lovitt, Jackson Kernion, Amanda Askell, Yuntao Bai, Saurav Kadavath, Ben","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned because it crossed a citation-hub threshold. Current citing contexts most often use it as background evidence (24 contexts).","role_counts":[{"n":24,"context_role":"background"},{"n":3,"context_role":"method"},{"n":1,"context_role":"dataset"}]},"error":null,"updated_at":"2026-06-27T04:25:44.409580+00:00"},"summary_claims":{"job_type":"summary_claims","status":"succeeded","result":{"title":"Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned","claims":[{"claim_text":"We describe our early efforts to red team language models in order to simultaneously discover, measure, and attempt to reduce their potentially harmful outputs. We make three main contributions. First, we investigate scaling behaviors for red teaming across 3 model sizes (2.7B, 13B, and 52B parameters) and 4 model types: a plain language model (LM); an LM prompted to be helpful, honest, and harmless; an LM with rejection sampling; and a model trained to be helpful and harmless using reinforcement learning from human feedback (RLHF). We find that the RLHF models are increasingly difficult to re","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned because it crossed a citation-hub threshold.","role_counts":[]},"error":null,"updated_at":"2026-05-14T17:49:29.241745+00:00"}},"summary":{"title":"Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned","claims":[{"claim_text":"We describe our early efforts to red team language models in order to simultaneously discover, measure, and attempt to reduce their potentially harmful outputs. We make three main contributions. First, we investigate scaling behaviors for red teaming across 3 model sizes (2.7B, 13B, and 52B parameters) and 4 model types: a plain language model (LM); an LM prompted to be helpful, honest, and harmless; an LM with rejection sampling; and a model trained to be helpful and harmless using reinforcement learning from human feedback (RLHF). We find that the RLHF models are increasingly difficult to re","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks Red Teaming Language Models to Reduce Harms: Methods, Scaling Behaviors, and Lessons Learned because it crossed a citation-hub threshold.","role_counts":[]},"graph":{"co_cited":[{"title":"Universal and Transferable Adversarial Attacks on Aligned Language Models","work_id":"3322fa86-1768-4677-8425-dd326b45e078","shared_citers":17},{"title":"Constitutional AI: Harmlessness from AI Feedback","work_id":"faaaa4e0-2676-4fac-a0b4-99aef10d2095","shared_citers":16},{"title":"Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback","work_id":"a1f2574b-a899-4713-be60-c87ba332656c","shared_citers":14},{"title":"Red Teaming Language Models with Language Models","work_id":"d1274c54-508f-42f9-aeb3-91db13f3a622","shared_citers":12},{"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","shared_citers":10},{"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","shared_citers":8},{"title":"Ethical and social risks of harm from Language Models","work_id":"b4ce1c45-ef69-445a-a872-dbb785b485e9","shared_citers":7},{"title":"HarmBench: A Standardized Evaluation Framework for Automated Red Teaming and Robust Refusal","work_id":"b0b0303f-2444-4789-a979-8153624312ff","shared_citers":7},{"title":"Llama 2: Open Foundation and Fine-Tuned Chat Models","work_id":"68a5177f-d644-44c1-bd4f-4e5278c22f5d","shared_citers":7},{"title":"Llama Guard: LLM-based Input-Output Safeguard for Human-AI Conversations","work_id":"93844332-869b-448c-a1be-35466150b1b2","shared_citers":7},{"title":"Proximal Policy Optimization Algorithms","work_id":"240c67fe-d14d-4520-91c1-38a4e272ca19","shared_citers":7},{"title":"A General Language Assistant as a Laboratory for Alignment","work_id":"a43f9ea0-01be-47d5-b8ee-a1a9f73381c5","shared_citers":6},{"title":"Evaluating Large Language Models Trained on Code","work_id":"042493e9-b26f-4b4e-bbde-382072ca9b08","shared_citers":6},{"title":"Improving alignment of dialogue agents via targeted human judgements","work_id":"6ad5970e-7550-4ae8-a158-7084dec7e3cc","shared_citers":6},{"title":"LLaMA: Open and Efficient Foundation Language Models","work_id":"c018fc23-6f3f-4035-9d02-28a2173b2b9d","shared_citers":6},{"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","shared_citers":5},{"title":"Fine-Tuning Language Models from Human Preferences","work_id":"4f54aad1-f3b6-404f-b9c7-e21ba0a33b99","shared_citers":5},{"title":"Holistic Evaluation of Language Models","work_id":"cc02a01e-7218-47dc-8e66-3333e7e4adec","shared_citers":5},{"title":"Measuring Massive Multitask Language Understanding","work_id":"e87ec49a-544b-4ec8-8991-75298c64ff5e","shared_citers":5},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":5},{"title":"Scaling Laws for Neural Language Models","work_id":"b7dd8749-9c45-4977-ab9b-64478dce1ae8","shared_citers":5},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":5},{"title":"Training language models to follow instructions with human feedback","work_id":"52aff42f-4fa9-4fcf-bdb3-1459b9bebf65","shared_citers":5},{"title":"arXiv preprint arXiv:2305.15324 , year=","work_id":"12dff4e7-a81d-4b95-aba7-885a457042c6","shared_citers":4}],"time_series":[{"n":1,"year":2022},{"n":4,"year":2023},{"n":2,"year":2024},{"n":32,"year":2026}],"dependency_candidates":[]},"authors":[{"id":"fa166f46-a486-47b9-b4c7-b3c0d3548ac0","orcid":null,"display_name":"Amanda Askell","source":"manual","import_confidence":0.72},{"id":"20146a94-3ea4-4ed5-8389-126a1613d880","orcid":null,"display_name":"Deep Ganguli","source":"manual","import_confidence":0.72},{"id":"6ff3739a-293b-4060-bd4d-4c63b8f88381","orcid":null,"display_name":"Jackson Kernion","source":"manual","import_confidence":0.72},{"id":"97c891e8-0ac2-40a8-8fb9-c49436fa1c87","orcid":null,"display_name":"Liane Lovitt","source":"manual","import_confidence":0.72},{"id":"155c969c-3372-4b52-b881-15f051250399","orcid":null,"display_name":"Saurav Kadavath","source":"manual","import_confidence":0.72},{"id":"f15a23ad-8224-4c50-a63d-24a766b946a8","orcid":null,"display_name":"Yuntao Bai","source":"manual","import_confidence":0.72}]}}