{"work":{"id":"2676656a-67bd-4ad5-bad6-cb6f5fcdbfbe","openalex_id":null,"doi":null,"arxiv_id":"2411.15594","raw_key":null,"title":"A Survey on LLM-as-a-Judge","authors":null,"authors_text":"Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, Saizhuo Wang, Kun Zhang, Yuanzhuo Wang, Wen Gao, Lionel Ni, and Jian Guo","year":2024,"venue":"cs.CL","abstract":"Accurate and consistent evaluation is crucial for decision-making across numerous fields, yet it remains a challenging task due to inherent subjectivity, variability, and scale. Large Language Models (LLMs) have achieved remarkable success across diverse domains, leading to the emergence of \"LLM-as-a-Judge,\" where LLMs are employed as evaluators for complex tasks. With their ability to process diverse data types and provide scalable, cost-effective, and consistent assessments, LLMs present a compelling alternative to traditional expert-driven evaluations. However, ensuring the reliability of LLM-as-a-Judge systems remains a significant challenge that requires careful design and standardization. This paper provides a comprehensive survey of LLM-as-a-Judge, addressing the core question: How can reliable LLM-as-a-Judge systems be built? We explore strategies to enhance reliability, including improving consistency, mitigating biases, and adapting to diverse assessment scenarios. Additionally, we propose methodologies for evaluating the reliability of LLM-as-a-Judge systems, supported by a novel benchmark designed for this purpose. To advance the development and real-world deployment of LLM-as-a-Judge systems, we also discussed practical applications, challenges, and future directions. This survey serves as a foundational reference for researchers and practitioners in this rapidly evolving field.","external_url":"https://arxiv.org/abs/2411.15594","cited_by_count":null,"metadata_source":"pith","metadata_fetched_at":"2026-05-25T07:40:28.724343+00:00","pith_arxiv_id":"2411.15594","created_at":"2026-05-08T21:39:24.641446+00:00","updated_at":"2026-05-25T07:40:28.724343+00:00","title_quality_ok":false,"display_title":"A Survey on LLM-as-a-Judge","render_title":"A Survey on LLM-as-a-Judge"},"hub":{"state":{"work_id":"2676656a-67bd-4ad5-bad6-cb6f5fcdbfbe","tier":"super_hub","tier_reason":"100+ Pith inbound or 10,000+ external citations","pith_inbound_count":114,"external_cited_by_count":null,"distinct_field_count":18,"first_pith_cited_at":"2023-12-06T19:07:38+00:00","last_pith_cited_at":"2026-05-21T17:57:38+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":"2026-06-05T15:59:09.508591+00:00","tier_text":"super_hub"},"tier":"super_hub","role_counts":[{"context_role":"background","n":18},{"context_role":"method","n":5}],"polarity_counts":[{"context_polarity":"background","n":16},{"context_polarity":"use_method","n":5},{"context_polarity":"unclear","n":2}],"runs":{"ask_index":{"job_type":"ask_index","status":"succeeded","result":{"title":"A Survey on LLM-as-a-Judge","claims":[{"claim_text":"Accurate and consistent evaluation is crucial for decision-making across numerous fields, yet it remains a challenging task due to inherent subjectivity, variability, and scale. Large Language Models (LLMs) have achieved remarkable success across diverse domains, leading to the emergence of \"LLM-as-a-Judge,\" where LLMs are employed as evaluators for complex tasks. With their ability to process diverse data types and provide scalable, cost-effective, and consistent assessments, LLMs present a compelling alternative to traditional expert-driven evaluations. However, ensuring the reliability of L","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"LLaMA [4] have transformed AI by enabling human-like text generation and advanced natural language processing, spurring innovation in conversational agents, automated content cre- ation, and real-time translation [5]. Recent enhancements have extended their utility to multimodal tasks, including text-to- image and text-to-video generation that broaden the scope of generative AI applications [6], [7]. However, their dependence on static pre-training data can lead to outdated outputs and hallucina","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Table Retrieval (TR) [3, 7, 33] is a fundamental task in Information Retrieval (IR) that identifies relevant tables from large-scale reposi- tories to satisfy user needs. Traditionally, TR has been formulated as a conventional ad-hoc retrieval problem [8], primarily focused on semantic matching to support downstream applications such as Open Table Question Answering [14] or Data Integration [53, 56]. However, the proliferation of LLMs and intelligent agents [2, 39, 51] 1https://github.com/AriKin","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"prior cancer prediction studies [4, 5] and recommended in recent work on early detec- tion settings [49]. Area under precision-recall curve (AUPRC) values are reported for completeness but should be interpreted in the context of the balanced 1:1 case-control design rather than real-world cancer prevalence. For the LLM-as-a-judge evaluation, we adopted a pairwise comparison framework [50], using GPT-5 with high reasoning effort as the judge model. For each rubric, the judge was presented with two","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"To address this, we applied an additional annotation phase to ensure data quality. Specifically, we adopted a two-stage labeling process that combines automatic pre-annotation with large language models (LLMs) and manual verification. (1) LLM-based pre-annotation. In this stage, we drew inspiration from prior work on voting strategies [ 45] and LLM-as-a-judge methods [21]. We employed two advanced LLMs, DeepSeek [3] and Qwen [7], to independently provide preliminary labels along with justificati","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"To further enhance efficiency, both the annotation model and the pool of tasks are periodically refreshed, ensuring that data collection consistently targets the weaknesses of the most recent agent. 2.5 Multi-turn Reinforcement Learning To train agents capable of long-horizon reasoning and interactive decision-making, we adopt a multi-turn RL framework built on RLVR (Reinforcement Learning with Verifiable Rewards) [21]. We construct domain- specific pipelines that automatically synthesize large-","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"management, and memory interactions across long episodes) suffers in comparison to shorter-term tasks [242, 243]. Recent computer use testbeds such as OSWorld [229] and OSWorld-Human [230] expose substantial performance and reliability gaps across long workflows and heterogeneous apps, even for strong agents. Similarly, agents struggle to manage memory appropriately over long horizons, leading to erroneous or redundant context [244]. This deficiency is only compounded for security considerations","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks A Survey on LLM-as-a-Judge 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":5,"context_role":"method"}]},"error":null,"updated_at":"2026-05-21T05:42:38.324917+00:00"},"author_expand":{"job_type":"author_expand","status":"succeeded","result":{"authors_linked":[{"id":"5da846d7-2658-4e24-af08-1e5fa573d8be","orcid":null,"display_name":"Jiawei Gu"},{"id":"54bfda4e-8558-4c1c-9648-36b756197322","orcid":null,"display_name":"Xuhui Jiang"},{"id":"03cdede3-5595-4d0d-a39b-6f8b50e25bd3","orcid":null,"display_name":"Zhichao Shi"},{"id":"9f56b9b2-f3d0-43f2-85d0-b0ca1a71a65b","orcid":null,"display_name":"Hexiang Tan"},{"id":"d69de932-3b3b-4396-befc-9a89a7fccc51","orcid":null,"display_name":"Xuehao Zhai"},{"id":"593622de-185b-49f8-8eec-5e6d02c3fc35","orcid":null,"display_name":"Chengjin Xu"}]},"error":null,"updated_at":"2026-05-21T05:42:37.924463+00:00"},"context_extract":{"job_type":"context_extract","status":"succeeded","result":{"enqueued_papers":25},"error":null,"updated_at":"2026-05-14T10:49:17.018350+00:00"},"graph_features":{"job_type":"graph_features","status":"succeeded","result":{"co_cited":[{"title":"Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena","work_id":"d0c30cd7-81e1-4159-a87f-f6adca77ff08","shared_citers":9},{"title":"LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods","work_id":"377b190f-39ed-4db3-9747-433802e5303c","shared_citers":8},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":8},{"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","shared_citers":7},{"title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","work_id":"c5006563-f3ec-438a-9e35-b7b484f34828","shared_citers":6},{"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","shared_citers":6},{"title":"LLaMA: Open and Efficient Foundation Language Models","work_id":"c018fc23-6f3f-4035-9d02-28a2173b2b9d","shared_citers":6},{"title":"Evaluating Large Language Models Trained on Code","work_id":"042493e9-b26f-4b4e-bbde-382072ca9b08","shared_citers":5},{"title":"Self-Consistency Improves Chain of Thought Reasoning in Language Models","work_id":"8c6d5a6b-b5cc-4105-9c84-9c34bb9375bb","shared_citers":5},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":5},{"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","shared_citers":5},{"title":"Constitutional AI: Harmlessness from AI Feedback","work_id":"faaaa4e0-2676-4fac-a0b4-99aef10d2095","shared_citers":4},{"title":"H., Chen, S., Liu, Z., Jiang, F., and Wang, B","work_id":"f26248cc-fd46-4eb4-94cf-a455fa482305","shared_citers":4},{"title":"Holistic Evaluation of Language Models","work_id":"cc02a01e-7218-47dc-8e66-3333e7e4adec","shared_citers":4},{"title":"On the Opportunities and Risks of Foundation Models","work_id":"a18039e9-928d-47c9-a836-32656a71bf71","shared_citers":4},{"title":"OpenAI o1 System Card","work_id":"68d3c334-0fc9-49e3-b7b0-a69afae933e2","shared_citers":4},{"title":"Reflexion: Language Agents with Verbal Reinforcement Learning","work_id":"778f739e-5f55-4961-8a2a-e4736a2757f4","shared_citers":4},{"title":"s1: Simple test-time scaling","work_id":"806265b1-8f22-48dd-b8ad-a99823b18fa4","shared_citers":4},{"title":"Scaling Laws for Neural Language Models","work_id":"b7dd8749-9c45-4977-ab9b-64478dce1ae8","shared_citers":4},{"title":"arXiv preprint arXiv:2305.15324 , year=","work_id":"12dff4e7-a81d-4b95-aba7-885a457042c6","shared_citers":3},{"title":"arXiv preprint arXiv:2412.12509 , year =","work_id":"fbc020bc-197e-43ce-92fd-3dca29743d27","shared_citers":3},{"title":"DeepSeek-V3 Technical Report","work_id":"57d2791d-2219-4c31-a077-afc04b12a75c","shared_citers":3},{"title":"Gemini: A Family of Highly Capable Multimodal Models","work_id":"83f7c85b-3f11-450f-ac0c-64d9745220b2","shared_citers":3},{"title":"Justice or prejudice? quantifying biases in llm-as-a-judge","work_id":"226fa5a3-d2b0-46fc-b73d-cfb5ca7b6a4c","shared_citers":3}],"time_series":[{"n":3,"year":2025},{"n":57,"year":2026}],"dependency_candidates":[]},"error":null,"updated_at":"2026-05-14T10:49:23.470868+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":"2026-05-14T10:49:23.448824+00:00"},"role_polarity":{"job_type":"role_polarity","status":"succeeded","result":{"title":"A Survey on LLM-as-a-Judge","claims":[{"claim_text":"Accurate and consistent evaluation is crucial for decision-making across numerous fields, yet it remains a challenging task due to inherent subjectivity, variability, and scale. Large Language Models (LLMs) have achieved remarkable success across diverse domains, leading to the emergence of \"LLM-as-a-Judge,\" where LLMs are employed as evaluators for complex tasks. With their ability to process diverse data types and provide scalable, cost-effective, and consistent assessments, LLMs present a compelling alternative to traditional expert-driven evaluations. However, ensuring the reliability of L","claim_type":"abstract","evidence_strength":"source_metadata"},{"claim_text":"LLaMA [4] have transformed AI by enabling human-like text generation and advanced natural language processing, spurring innovation in conversational agents, automated content cre- ation, and real-time translation [5]. Recent enhancements have extended their utility to multimodal tasks, including text-to- image and text-to-video generation that broaden the scope of generative AI applications [6], [7]. However, their dependence on static pre-training data can lead to outdated outputs and hallucina","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"Table Retrieval (TR) [3, 7, 33] is a fundamental task in Information Retrieval (IR) that identifies relevant tables from large-scale reposi- tories to satisfy user needs. Traditionally, TR has been formulated as a conventional ad-hoc retrieval problem [8], primarily focused on semantic matching to support downstream applications such as Open Table Question Answering [14] or Data Integration [53, 56]. However, the proliferation of LLMs and intelligent agents [2, 39, 51] 1https://github.com/AriKin","claim_type":"background","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"prior cancer prediction studies [4, 5] and recommended in recent work on early detec- tion settings [49]. Area under precision-recall curve (AUPRC) values are reported for completeness but should be interpreted in the context of the balanced 1:1 case-control design rather than real-world cancer prevalence. For the LLM-as-a-judge evaluation, we adopted a pairwise comparison framework [50], using GPT-5 with high reasoning effort as the judge model. For each rubric, the judge was presented with two","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"To address this, we applied an additional annotation phase to ensure data quality. Specifically, we adopted a two-stage labeling process that combines automatic pre-annotation with large language models (LLMs) and manual verification. (1) LLM-based pre-annotation. In this stage, we drew inspiration from prior work on voting strategies [ 45] and LLM-as-a-judge methods [21]. We employed two advanced LLMs, DeepSeek [3] and Qwen [7], to independently provide preliminary labels along with justificati","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"To further enhance efficiency, both the annotation model and the pool of tasks are periodically refreshed, ensuring that data collection consistently targets the weaknesses of the most recent agent. 2.5 Multi-turn Reinforcement Learning To train agents capable of long-horizon reasoning and interactive decision-making, we adopt a multi-turn RL framework built on RLVR (Reinforcement Learning with Verifiable Rewards) [21]. We construct domain- specific pipelines that automatically synthesize large-","claim_type":"method","confidence":0.9,"evidence_strength":"citation_context"},{"claim_text":"management, and memory interactions across long episodes) suffers in comparison to shorter-term tasks [242, 243]. Recent computer use testbeds such as OSWorld [229] and OSWorld-Human [230] expose substantial performance and reliability gaps across long workflows and heterogeneous apps, even for strong agents. Similarly, agents struggle to manage memory appropriately over long horizons, leading to erroneous or redundant context [244]. This deficiency is only compounded for security considerations","claim_type":"background","confidence":0.85,"evidence_strength":"citation_context"}],"why_cited":"Pith tracks A Survey on LLM-as-a-Judge 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":5,"context_role":"method"}]},"error":null,"updated_at":"2026-05-21T05:42:37.636480+00:00"},"summary_claims":{"job_type":"summary_claims","status":"succeeded","result":{"title":"A Survey on LLM-as-a-Judge","claims":[{"claim_text":"Accurate and consistent evaluation is crucial for decision-making across numerous fields, yet it remains a challenging task due to inherent subjectivity, variability, and scale. Large Language Models (LLMs) have achieved remarkable success across diverse domains, leading to the emergence of \"LLM-as-a-Judge,\" where LLMs are employed as evaluators for complex tasks. With their ability to process diverse data types and provide scalable, cost-effective, and consistent assessments, LLMs present a compelling alternative to traditional expert-driven evaluations. However, ensuring the reliability of L","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks A Survey on LLM-as-a-Judge because it crossed a citation-hub threshold.","role_counts":[]},"error":null,"updated_at":"2026-05-14T10:49:15.033479+00:00"}},"summary":{"title":"A Survey on LLM-as-a-Judge","claims":[{"claim_text":"Accurate and consistent evaluation is crucial for decision-making across numerous fields, yet it remains a challenging task due to inherent subjectivity, variability, and scale. Large Language Models (LLMs) have achieved remarkable success across diverse domains, leading to the emergence of \"LLM-as-a-Judge,\" where LLMs are employed as evaluators for complex tasks. With their ability to process diverse data types and provide scalable, cost-effective, and consistent assessments, LLMs present a compelling alternative to traditional expert-driven evaluations. However, ensuring the reliability of L","claim_type":"abstract","evidence_strength":"source_metadata"}],"why_cited":"Pith tracks A Survey on LLM-as-a-Judge because it crossed a citation-hub threshold.","role_counts":[]},"graph":{"co_cited":[{"title":"Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena","work_id":"d0c30cd7-81e1-4159-a87f-f6adca77ff08","shared_citers":9},{"title":"LLMs-as-Judges: A Comprehensive Survey on LLM-based Evaluation Methods","work_id":"377b190f-39ed-4db3-9747-433802e5303c","shared_citers":8},{"title":"Qwen3 Technical Report","work_id":"25a4e30c-1232-48e7-9925-02fa12ba7c9e","shared_citers":8},{"title":"GPT-4 Technical Report","work_id":"b928e041-6991-4c08-8c81-0359e4097c7b","shared_citers":7},{"title":"DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models","work_id":"c5006563-f3ec-438a-9e35-b7b484f34828","shared_citers":6},{"title":"DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning","work_id":"e6b75ad5-2877-4168-97c8-710407094d20","shared_citers":6},{"title":"LLaMA: Open and Efficient Foundation Language Models","work_id":"c018fc23-6f3f-4035-9d02-28a2173b2b9d","shared_citers":6},{"title":"Evaluating Large Language Models Trained on Code","work_id":"042493e9-b26f-4b4e-bbde-382072ca9b08","shared_citers":5},{"title":"Self-Consistency Improves Chain of Thought Reasoning in Language Models","work_id":"8c6d5a6b-b5cc-4105-9c84-9c34bb9375bb","shared_citers":5},{"title":"The Llama 3 Herd of Models","work_id":"1549a635-88af-4ac1-acfe-51ae7bb53345","shared_citers":5},{"title":"Training Verifiers to Solve Math Word Problems","work_id":"acab1aa8-b4d6-40e0-a3ee-25341701dca2","shared_citers":5},{"title":"Constitutional AI: Harmlessness from AI Feedback","work_id":"faaaa4e0-2676-4fac-a0b4-99aef10d2095","shared_citers":4},{"title":"H., Chen, S., Liu, Z., Jiang, F., and Wang, B","work_id":"f26248cc-fd46-4eb4-94cf-a455fa482305","shared_citers":4},{"title":"Holistic Evaluation of Language Models","work_id":"cc02a01e-7218-47dc-8e66-3333e7e4adec","shared_citers":4},{"title":"On the Opportunities and Risks of Foundation Models","work_id":"a18039e9-928d-47c9-a836-32656a71bf71","shared_citers":4},{"title":"OpenAI o1 System Card","work_id":"68d3c334-0fc9-49e3-b7b0-a69afae933e2","shared_citers":4},{"title":"Reflexion: Language Agents with Verbal Reinforcement Learning","work_id":"778f739e-5f55-4961-8a2a-e4736a2757f4","shared_citers":4},{"title":"s1: Simple test-time scaling","work_id":"806265b1-8f22-48dd-b8ad-a99823b18fa4","shared_citers":4},{"title":"Scaling Laws for Neural Language Models","work_id":"b7dd8749-9c45-4977-ab9b-64478dce1ae8","shared_citers":4},{"title":"arXiv preprint arXiv:2305.15324 , year=","work_id":"12dff4e7-a81d-4b95-aba7-885a457042c6","shared_citers":3},{"title":"arXiv preprint arXiv:2412.12509 , year =","work_id":"fbc020bc-197e-43ce-92fd-3dca29743d27","shared_citers":3},{"title":"DeepSeek-V3 Technical Report","work_id":"57d2791d-2219-4c31-a077-afc04b12a75c","shared_citers":3},{"title":"Gemini: A Family of Highly Capable Multimodal Models","work_id":"83f7c85b-3f11-450f-ac0c-64d9745220b2","shared_citers":3},{"title":"Justice or prejudice? quantifying biases in llm-as-a-judge","work_id":"226fa5a3-d2b0-46fc-b73d-cfb5ca7b6a4c","shared_citers":3}],"time_series":[{"n":3,"year":2025},{"n":57,"year":2026}],"dependency_candidates":[]},"authors":[{"id":"593622de-185b-49f8-8eec-5e6d02c3fc35","orcid":null,"display_name":"Chengjin Xu","source":"manual","import_confidence":0.72},{"id":"9f56b9b2-f3d0-43f2-85d0-b0ca1a71a65b","orcid":null,"display_name":"Hexiang Tan","source":"manual","import_confidence":0.72},{"id":"5da846d7-2658-4e24-af08-1e5fa573d8be","orcid":null,"display_name":"Jiawei Gu","source":"manual","import_confidence":0.72},{"id":"d69de932-3b3b-4396-befc-9a89a7fccc51","orcid":null,"display_name":"Xuehao Zhai","source":"manual","import_confidence":0.72},{"id":"54bfda4e-8558-4c1c-9648-36b756197322","orcid":null,"display_name":"Xuhui Jiang","source":"manual","import_confidence":0.72},{"id":"03cdede3-5595-4d0d-a39b-6f8b50e25bd3","orcid":null,"display_name":"Zhichao Shi","source":"manual","import_confidence":0.72}]}}