{
  "schema_version": "1.0",
  "scope": "Source-linked facets for the eight credit-assignment method instantiations in the accompanying comparison. Paper versions are explicit; task domains and dependencies describe those versions. Selection uses documented properties and does not rank methods.",
  "facet_definitions": {
    "credit_granularity": {
      "description": "Credit scales and units explicitly described by the cited method. An action segment can share one advantage across many tokens.",
      "values": [
        "episode",
        "interaction_step",
        "interaction_turn",
        "assistant_action_segment"
      ]
    },
    "required_supervision": {
      "description": "Reward signals and auxiliary scoring or annotation inputs used by the cited credit-estimation instantiation. Positive entries establish use; omitted entries do not establish absence. This facet does not enumerate data-construction or task-verifier dependencies.",
      "values": [
        "terminal_outcome",
        "turn_verifier",
        "gold_answer",
        "frozen_reference_model",
        "dependency_annotator"
      ]
    },
    "learned_value_critic": {
      "description": "Whether the cited credit estimator requires a separately trained state/action value critic. A no value permits frozen reference models, dependency annotators and independently specified task judges. Unknown denotes insufficient primary evidence.",
      "values": [
        "yes",
        "no",
        "unknown"
      ]
    },
    "evaluation_setting": {
      "description": "Task domains evaluated in the cited paper version. These categories describe evaluation settings and do not establish performance equivalence across benchmarks.",
      "values": [
        "embodied_planning",
        "web_shopping",
        "search_question_answering",
        "software_issue_resolution",
        "cli_filesystem",
        "mathematical_reasoning",
        "code_generation",
        "application_tool_use",
        "user_service_dialogue",
        "function_calling",
        "interactive_games"
      ]
    }
  },
  "methods": [
    {
      "id": "gigpo",
      "name": "GiGPO",
      "paper": {
        "arxiv_id": "2505.10978",
        "version": "2505.10978v1",
        "title": "Group-in-Group Policy Optimization for LLM Agent Training",
        "authors": [
          "Feng, Lang",
          "Xue, Zhenghai",
          "Liu, Tingcong",
          "An, Bo"
        ],
        "publication_date": "2025-05-16",
        "url": "https://arxiv.org/abs/2505.10978v1",
        "bibtex": "@misc{arxiv250510978,\n  title = {Group-in-Group Policy Optimization for LLM Agent Training},\n  author = {Feng, Lang and Xue, Zhenghai and Liu, Tingcong and An, Bo},\n  year = {2025},\n  eprint = {2505.10978v1},\n  archivePrefix = {arXiv},\n  doi = {10.48550/arXiv.2505.10978},\n  url = {https://arxiv.org/abs/2505.10978v1}\n}"
      },
      "credit_granularity": [
        {
          "value": "episode",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.10978v1#S4.SS3.p1",
              "support": "The estimator combines episode-relative and step-relative advantages."
            }
          ]
        },
        {
          "value": "interaction_step",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.10978v1#S4.SS3.p1",
              "support": "Individual environment actions receive a step-relative advantage."
            }
          ]
        }
      ],
      "required_supervision": [
        {
          "value": "terminal_outcome",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.10978v1#S4.SS1.p1",
              "support": "Episode return supplies outcome feedback; binary terminal success is an explicitly supported reward setting."
            }
          ]
        }
      ],
      "learned_value_critic": {
        "value": "no",
        "evidence": [
          {
            "url": "https://arxiv.org/html/2505.10978v1#S1.p6",
            "support": "GiGPO explicitly operates without a critic or auxiliary value model."
          }
        ]
      },
      "evaluation_settings": [
        {
          "value": "embodied_planning",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.10978v1#S5.SS1.p1",
              "support": "ALFWorld evaluates sequential household tasks in a simulated environment."
            }
          ]
        },
        {
          "value": "web_shopping",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.10978v1#S5.SS1.p1",
              "support": "WebShop evaluates search, navigation and purchase decisions."
            }
          ]
        },
        {
          "value": "interactive_games",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.10978v1#A4.SS3.p1",
              "support": "Appendix experiments evaluate vision-language agents on Sokoban and Gym Cards EZPoints."
            }
          ]
        }
      ]
    },
    {
      "id": "turn-level-rl",
      "name": "Turn-level RL (MT-GRPO)",
      "paper": {
        "arxiv_id": "2505.11821",
        "version": "2505.11821v1",
        "title": "Reinforcing Multi-Turn Reasoning in LLM Agents via Turn-Level Credit Assignment",
        "authors": [
          "Zeng, Siliang",
          "Wei, Quan",
          "Brown, William",
          "Frunza, Oana",
          "Nevmyvaka, Yuriy",
          "Hong, Mingyi"
        ],
        "publication_date": "2025-05-17",
        "url": "https://arxiv.org/abs/2505.11821v1",
        "bibtex": "@misc{arxiv250511821,\n  title = {Reinforcing Multi-Turn Reasoning in LLM Agents via Turn-Level Credit Assignment},\n  author = {Zeng, Siliang and Wei, Quan and Brown, William and Frunza, Oana and Nevmyvaka, Yuriy and Hong, Mingyi},\n  year = {2025},\n  eprint = {2505.11821v1},\n  archivePrefix = {arXiv},\n  doi = {10.48550/arXiv.2505.11821},\n  url = {https://arxiv.org/abs/2505.11821v1}\n}"
      },
      "credit_granularity": [
        {
          "value": "interaction_turn",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.11821v1#S4.SS2.p1",
              "support": "Each agent-environment interaction receives a turn-level advantage."
            }
          ]
        }
      ],
      "required_supervision": [
        {
          "value": "terminal_outcome",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.11821v1#S3.SS2.p3",
              "support": "Outcome verifiers score final-answer correctness and response formatting."
            }
          ]
        },
        {
          "value": "turn_verifier",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.11821v1#S3.SS2.p2",
              "support": "The first-turn verifiers check tool execution and answer presence in retrieved results."
            }
          ]
        },
        {
          "value": "gold_answer",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.11821v1#S3.SS2.p2",
              "support": "Search-result supervision checks whether an accepted reference answer appears in retrieved text."
            }
          ]
        }
      ],
      "learned_value_critic": {
        "value": "no",
        "evidence": [
          {
            "url": "https://arxiv.org/html/2505.11821v1#S4.SS2.p2",
            "support": "The published MT-GRPO instance computes both advantages directly from normalized group rewards in Equations 9-10."
          },
          {
            "url": "https://arxiv.org/html/2505.11821v1#S2.SS2.p2",
            "support": "The paper explicitly describes GRPO as eliminating the separate value function."
          }
        ],
        "note": "This value applies to the published MT-GRPO instantiation. The general turn-level formulation also permits other RL algorithms."
      },
      "evaluation_settings": [
        {
          "value": "search_question_answering",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2505.11821v1#S6.p1",
              "support": "The reported experiments use a two-turn Wikipedia search agent answering TriviaQA questions."
            }
          ]
        }
      ]
    },
    {
      "id": "hgpo",
      "name": "HGPO",
      "paper": {
        "arxiv_id": "2602.22817",
        "version": "2602.22817v1",
        "title": "Hierarchy-of-Groups Policy Optimization for Long-Horizon Agentic Tasks",
        "authors": [
          "He, Shuo",
          "Feng, Lang",
          "Wei, Qi",
          "Cheng, Xin",
          "Feng, Lei",
          "An, Bo"
        ],
        "publication_date": "2026-02-26",
        "url": "https://arxiv.org/abs/2602.22817v1",
        "bibtex": "@misc{arxiv260222817,\n  title = {Hierarchy-of-Groups Policy Optimization for Long-Horizon Agentic Tasks},\n  author = {He, Shuo and Feng, Lang and Wei, Qi and Cheng, Xin and Feng, Lei and An, Bo},\n  year = {2026},\n  eprint = {2602.22817v1},\n  archivePrefix = {arXiv},\n  doi = {10.48550/arXiv.2602.22817},\n  url = {https://arxiv.org/abs/2602.22817v1}\n}"
      },
      "credit_granularity": [
        {
          "value": "interaction_step",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2602.22817v1#S4.SS2.p4",
              "support": "Hierarchical group-relative advantages are estimated for individual action steps."
            }
          ]
        }
      ],
      "required_supervision": [
        {
          "value": "terminal_outcome",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2602.22817v1#S3.p1",
              "support": "The problem setup provides a scalar reward at the trajectory's final step."
            }
          ]
        }
      ],
      "learned_value_critic": {
        "value": "no",
        "evidence": [
          {
            "url": "https://arxiv.org/html/2602.22817v1#S5.SS3.p3",
            "support": "The computational analysis explicitly identifies HGPO as critic-free with a single actor LLM."
          }
        ]
      },
      "evaluation_settings": [
        {
          "value": "embodied_planning",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2602.22817v1#S5.SS1.p1",
              "support": "The reported agent benchmarks include ALFWorld."
            }
          ]
        },
        {
          "value": "web_shopping",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2602.22817v1#S5.SS1.p1",
              "support": "The reported agent benchmarks include WebShop."
            }
          ]
        }
      ]
    },
    {
      "id": "rtmc",
      "name": "RTMC",
      "paper": {
        "arxiv_id": "2604.11037",
        "version": "2604.11037v1",
        "title": "RTMC: Step-Level Credit Assignment via Rollout Trees",
        "authors": [
          "Wang, Tao",
          "Zheng, Suhang",
          "Xu, Xiaoxiao"
        ],
        "publication_date": "2026-04-13",
        "url": "https://arxiv.org/abs/2604.11037v1",
        "bibtex": "@misc{arxiv260411037,\n  title = {RTMC: Step-Level Credit Assignment via Rollout Trees},\n  author = {Wang, Tao and Zheng, Suhang and Xu, Xiaoxiao},\n  year = {2026},\n  eprint = {2604.11037v1},\n  archivePrefix = {arXiv},\n  doi = {10.48550/arXiv.2604.11037},\n  url = {https://arxiv.org/abs/2604.11037v1}\n}"
      },
      "credit_granularity": [
        {
          "value": "interaction_step",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2604.11037v1#S3.SS1.p3",
              "support": "The rollout tree aggregates return statistics for state-action steps."
            }
          ]
        },
        {
          "value": "assistant_action_segment",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2604.11037v1#S3.SS1.p1",
              "support": "One advantage is shared across the reasoning, tool-call and argument tokens of an action segment."
            }
          ]
        }
      ],
      "required_supervision": [
        {
          "value": "terminal_outcome",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2604.11037v1#S5.SS1.p2",
              "support": "The reported training configurations use task terminal reward."
            }
          ]
        },
        {
          "value": "turn_verifier",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2604.11037v1#S5.SS1.p2",
              "support": "The reported RTMC configuration also uses valid-action, invalid-action and validation feedback."
            }
          ]
        }
      ],
      "learned_value_critic": {
        "value": "no",
        "evidence": [
          {
            "url": "https://arxiv.org/html/2604.11037v1#S1.SS0.SSS0.Px3.p1",
            "support": "The method explicitly estimates step values and advantages without a critic network."
          }
        ]
      },
      "evaluation_settings": [
        {
          "value": "software_issue_resolution",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2604.11037v1#S5.SS1.p1",
              "support": "Evaluation uses SWE-bench Verified with a SWE-agent-style tool interface."
            }
          ]
        }
      ]
    },
    {
      "id": "a3",
      "name": "A3",
      "paper": {
        "arxiv_id": "2605.08013",
        "version": "2605.08013v1",
        "title": "Learning CLI Agents with Structured Action Credit under Selective Observation",
        "authors": [
          "Su, Haoyang",
          "Wen, Ying"
        ],
        "publication_date": "2026-05-08",
        "url": "https://arxiv.org/abs/2605.08013v1",
        "bibtex": "@misc{arxiv260508013,\n  title = {Learning CLI Agents with Structured Action Credit under Selective Observation},\n  author = {Su, Haoyang and Wen, Ying},\n  year = {2026},\n  eprint = {2605.08013v1},\n  archivePrefix = {arXiv},\n  doi = {10.48550/arXiv.2605.08013},\n  url = {https://arxiv.org/abs/2605.08013v1}\n}"
      },
      "credit_granularity": [
        {
          "value": "episode",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2605.08013v1#S3.SS3.p1",
              "support": "The estimator includes an episode-scope outcome component."
            }
          ]
        },
        {
          "value": "interaction_turn",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2605.08013v1#S3.SS3.p1",
              "support": "Each shell-action turn receives an advantage combining episode, turn and action-history tree scopes."
            }
          ]
        }
      ],
      "required_supervision": [
        {
          "value": "terminal_outcome",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2605.08013v1#S3.SS3.p1",
              "support": "The credit estimator takes one episode return for each completed rollout."
            }
          ]
        }
      ],
      "learned_value_critic": {
        "value": "no",
        "evidence": [
          {
            "url": "https://arxiv.org/html/2605.08013v1#S2.SS2.p3",
            "support": "The paper states that A3 constructs its syntax-based advantage without auxiliary models."
          },
          {
            "url": "https://arxiv.org/html/2605.08013v1#S3.SS3.SSS3.p1",
            "support": "Tree state and branch values are averages of sampled episode returns."
          }
        ],
        "note": "This field describes the advantage estimator. Task reward computation and dataset construction have their own dependencies."
      },
      "evaluation_settings": [
        {
          "value": "cli_filesystem",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2605.08013v1#S4.SS1.p1",
              "support": "The evaluations use a shared shell interface over filesystem workspaces for operating-system, database and data-reasoning tasks."
            }
          ]
        }
      ]
    },
    {
      "id": "trace",
      "name": "TRACE",
      "paper": {
        "arxiv_id": "2607.13988",
        "version": "2607.13988v1",
        "title": "TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents",
        "authors": [
          "Tao, Leitian",
          "Peng, Baolin",
          "Yao, Wenlin",
          "Ge, Tao",
          "Cheng, Hao",
          "Wang, Mike Hang",
          "Gao, Jianfeng",
          "Li, Sharon"
        ],
        "publication_date": "2026-07-15",
        "url": "https://arxiv.org/abs/2607.13988v1",
        "bibtex": "@misc{arxiv260713988,\n  title = {TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents},\n  author = {Tao, Leitian and Peng, Baolin and Yao, Wenlin and Ge, Tao and Cheng, Hao and Wang, Mike Hang and Gao, Jianfeng and Li, Sharon},\n  year = {2026},\n  eprint = {2607.13988v1},\n  archivePrefix = {arXiv},\n  doi = {10.48550/arXiv.2607.13988},\n  url = {https://arxiv.org/abs/2607.13988v1}\n}"
      },
      "credit_granularity": [
        {
          "value": "episode",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2607.13988v1#S3.SS3.SSS0.Px3.p1",
              "support": "The objective retains a group-relative trajectory outcome advantage."
            }
          ]
        },
        {
          "value": "interaction_turn",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2607.13988v1#S3.SS1.p1",
              "support": "Credit targets correspond to tool-action and observation transitions at tool-call boundaries."
            }
          ]
        }
      ],
      "required_supervision": [
        {
          "value": "terminal_outcome",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2607.13988v1#S3.SS1.p1",
              "support": "A final-answer verifier supplies the terminal task outcome."
            }
          ]
        },
        {
          "value": "gold_answer",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2607.13988v1#S3.SS2.SSS0.Px1.p1",
              "support": "Prefix scoring measures the training-time gold answer's conditional log-probability."
            }
          ]
        },
        {
          "value": "frozen_reference_model",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2607.13988v1#S3.SS2.SSS0.Px1.p1",
              "support": "A frozen copy of the initial policy supplies prefix answer-likelihood scores."
            }
          ]
        }
      ],
      "learned_value_critic": {
        "value": "no",
        "evidence": [
          {
            "url": "https://arxiv.org/html/2607.13988v1#S3.SS2.SSS0.Px1.p1",
            "support": "TRACE explicitly replaces critic learning with a frozen reference-model probe."
          }
        ]
      },
      "evaluation_settings": [
        {
          "value": "search_question_answering",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2607.13988v1#S4.SS1.SSS0.Px6.p1",
              "support": "Evaluation covers BrowseComp-Plus, BrowseComp, GAIA and xbench-DeepSearch with closed-corpus and open-web retrieval."
            }
          ]
        }
      ]
    },
    {
      "id": "tcpo",
      "name": "TCPO",
      "paper": {
        "arxiv_id": "2608.01667",
        "version": "2608.01667v1",
        "title": "TCPO: Turn-Level Credit Policy Optimization",
        "authors": [
          "Liao, Sicong",
          "Chen, Zhi",
          "Tang, Yaohua"
        ],
        "publication_date": "2026-08-03",
        "url": "https://arxiv.org/abs/2608.01667v1",
        "bibtex": "@misc{arxiv260801667,\n  title = {TCPO: Turn-Level Credit Policy Optimization},\n  author = {Liao, Sicong and Chen, Zhi and Tang, Yaohua},\n  year = {2026},\n  eprint = {2608.01667v1},\n  archivePrefix = {arXiv},\n  doi = {10.48550/arXiv.2608.01667},\n  url = {https://arxiv.org/abs/2608.01667v1}\n}"
      },
      "credit_granularity": [
        {
          "value": "interaction_turn",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2608.01667v1#S3.SS0.SSS1.Px3.p3",
              "support": "Turn-level advantages are broadcast to tokens in the corresponding response."
            }
          ]
        }
      ],
      "required_supervision": [
        {
          "value": "turn_verifier",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2608.01667v1#S3.SS0.SSS0.Px1.p1",
              "support": "A verifier scores each turn's candidate output; the rollout retains this score history."
            }
          ]
        }
      ],
      "learned_value_critic": {
        "value": "no",
        "evidence": [
          {
            "url": "https://arxiv.org/html/2608.01667v1#S1.p4",
            "support": "The method explicitly requires no learned critic."
          }
        ]
      },
      "evaluation_settings": [
        {
          "value": "mathematical_reasoning",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2608.01667v1#S4.SS1.SSS0.Px1.p1",
              "support": "Math evaluation uses MATH-500 and AIME with answer verification."
            }
          ]
        },
        {
          "value": "code_generation",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2608.01667v1#S4.SS1.SSS0.Px1.p1",
              "support": "Code evaluation uses LiveCodeBench and HumanEval with execution tests."
            }
          ]
        },
        {
          "value": "application_tool_use",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2608.01667v1#S4.SS1.SSS0.Px1.p1",
              "support": "AppWorld supplies stateful tool interactions and native task-completion verification."
            }
          ]
        }
      ]
    },
    {
      "id": "iapo",
      "name": "IAPO",
      "paper": {
        "arxiv_id": "2608.24588",
        "version": "2608.24588v2",
        "title": "IAPO: Influence-Aware Policy Optimization for Credit Assignment in Multi-Turn Service Agents",
        "authors": [
          "Ren, Bo",
          "Mao, Yirong",
          "Yang, Yi",
          "Que, Wenhui"
        ],
        "publication_date": "2026-08-25",
        "url": "https://arxiv.org/abs/2608.24588v2",
        "bibtex": "@misc{arxiv260824588,\n  title = {IAPO: Influence-Aware Policy Optimization for Credit Assignment in Multi-Turn Service Agents},\n  author = {Ren, Bo and Mao, Yirong and Yang, Yi and Que, Wenhui},\n  year = {2026},\n  eprint = {2608.24588v2},\n  archivePrefix = {arXiv},\n  doi = {10.48550/arXiv.2608.24588},\n  url = {https://arxiv.org/abs/2608.24588v2}\n}"
      },
      "credit_granularity": [
        {
          "value": "assistant_action_segment",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2608.24588v2#Sx4.p1",
              "support": "The trajectory advantage is multiplied by an action-specific weight shared across that action's tokens."
            }
          ]
        }
      ],
      "required_supervision": [
        {
          "value": "terminal_outcome",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2608.24588v2#Sx4.p1",
              "support": "IAPO routes the original GRPO trajectory advantage across trainable actions."
            }
          ]
        },
        {
          "value": "dependency_annotator",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2608.24588v2#Sx4.SSx3.SSS0.Px3.p1",
              "support": "A frozen annotator extracts the completed rollout's dependency graph for each policy update."
            }
          ]
        }
      ],
      "learned_value_critic": {
        "value": "no",
        "evidence": [
          {
            "url": "https://arxiv.org/html/2608.24588v2#Sx1.p5",
            "support": "The paper explicitly preserves the critic-free group-based update."
          }
        ],
        "note": "The dependency annotator is an auxiliary model with a different role from a learned value critic."
      },
      "evaluation_settings": [
        {
          "value": "user_service_dialogue",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2608.24588v2#Sx5.SSx1.p1",
              "support": "Evaluation includes tau-squared-Bench, UserBench and AgentChangeBench user-interaction tasks."
            }
          ]
        },
        {
          "value": "function_calling",
          "evidence": [
            {
              "url": "https://arxiv.org/html/2608.24588v2#Sx5.SSx1.p1",
              "support": "BFCL-v4 Multi-Turn evaluates retained function-calling capability."
            }
          ]
        }
      ]
    }
  ]
}