yuyangbai/DKGI-Trajectories

DKGI Trajectories on STaRK

Agent trajectories, answers and scores for every test configuration of the paper Knowledge Graphs as Files: An Agent Harness Where Tool Creation and Tool Use Co-Evolve (Yuyang Bai, Yu Wang, Yu Zhang).

• Code: https://github.com/leopoldwhite/DKGI
• Project page: https://yuyangbai.com/dkgi/
• The graph files the agents worked on: https://huggingface.co/datasets/yuyangbai/DKGI-STaRK-Graphs

In DKGI (direct knowledge graph interaction), an LLM agent receives the complete node and edge files of a STaRK knowledge graph and a shell. It writes and revises its own Python programs over these files to follow relations, combine node sets and select textual evidence. The agent returns a ranked list of node IDs.

Configurations

Each configuration is a subset (config_name), with one JSONL file per STaRK dataset (amazon, mag, prime) and one line per question.

Subset	Configuration	Questions	Paper
full_dkgi_pi_luna	Full DKGI (Pi), GPT-6-Luna	7,108 (test)	Table 2; Table 3 (test subset); analysis
full_dkgi_codex_luna	Full DKGI (Codex), GPT-6-Luna	7,108 (test)	Table 2; Table 3 (test subset)
full_dkgi_pi_astra	Full DKGI (Pi), GPT-6-Astra	7,108 (test)	Table 2; analysis
full_dkgi_codex_astra	Full DKGI (Codex), GPT-6-Astra	710 (test subset)	harness comparison (appendix)
shell_only	Shell only	710 (test subset)	Table 3
graph_tools_code	Graph tools + code	710 (test subset)	Table 3
graph_tools	Graph tools	710 (test subset)	Table 3
one_program	One program	710 (test subset)	Table 3
retrieval_only	Retrieval only	710 (test subset)	Table 3
no_edges	No edges	710 (test subset)	Table 3
id_probe	Closed-book ID probe	7,108 (test)	appendix (probes)
name_probe	Closed-book name probe (MAG, Prime)	5,466 (test)	appendix (probes)

test is STaRK's synthetic test set (7,108 questions). The test subset is STaRK's official test-0.1 split (710 questions). Its questions carry in_test_subset: true in every configuration. The rows of the paper's Table 3 (ablation) use the test subset, so the Full DKGI rows there are the test-subset questions of full_dkgi_pi_luna and full_dkgi_codex_luna.

Averaging hit_at_1, recall_at_20 and mrr over the questions of a dataset (and multiplying by 100) gives the numbers in the paper's tables.

Fields

Field	Meaning
configuration, configuration_id	paper name and subset name of the configuration
harness, model	pi (Pi coding agent, via DCI-Agent-Lite) or codex (Codex CLI); gpt-6-luna or gpt-6-astra
dataset, split, query_id, in_test_subset	the STaRK question
query, answer_ids	question text and gold node IDs (from STaRK)
status, attempts	scored or failed; the number of attempts (an attempt that failed was retried)
final_answer	the agent's final message
node_ids	the ranked node IDs parsed from it (names for the name probe)
hit_at_1, recall_at_20, mrr	scores of the question as used in the paper; a failed question scores 0
input_tokens, output_tokens	recorded token usage of the scored attempt
trajectory	list of steps of the scored attempt (below)

Each step has type, text, name, call_id and is_error:

type	text
user	the prompt (with Codex, also Codex's environment context message)
reasoning_summary	the model's reasoning summary, where the API returned one
assistant	assistant text, including the final answer
tool_call	the call: JSON arguments for Pi tools (read, bash, and in the graph-tool variants search, neighbors, get_node, get_schema); the JavaScript program of Codex's exec tool
tool_result	the output returned to the model (name and call_id link it to its call)

How the trajectories were extracted

• Pi configurations: Pi's message list of the scored attempt. Tool outputs appear as they were returned to the model, after the harness's head-and-tail truncation and the truncation of large outputs.
• Codex configurations: the input items of all model requests of the scored attempt, in order of first appearance, and the final agent message. Codex's own developer messages and tool definitions are left out because they are the same for every question. They are part of Codex CLI 0.156.1.
• Encrypted reasoning is not included. The system prompt of Pi and the prompts of every configuration are given in the paper's appendix and in the prompts/ directory of the code release.
• Paths in the trajectories are the absolute paths of our machine (the graph files were under .../experiments/stark/corpus/<dataset>).

License

CC BY 4.0, following the license of the STaRK data. Please cite STaRK and our paper.

@misc{bai2026dkgi,
  title  = {Knowledge Graphs as Files: An Agent Harness Where Tool Creation and Tool Use Co-Evolve},
  author = {Bai, Yuyang and Wang, Yu and Zhang, Yu},
  year   = {2026}
}

@inproceedings{wu2024stark,
  title     = {{STaRK}: Benchmarking {LLM} Retrieval on Textual and Relational Knowledge Bases},
  author    = {Wu, Shirley and Zhao, Shiyu and Yasunaga, Michihiro and Huang, Kexin and Cao, Kaidi and
               Huang, Qian and Ioannidis, Vassilis N. and Subbian, Karthik and Zou, James and Leskovec, Jure},
  booktitle = {Advances in Neural Information Processing Systems},
  volume    = {37},
  pages     = {127129--127153},
  year      = {2024}
}