README.md

metadata

language:
  - en
license: mit
pretty_name: Emergent Collusion in Long-Horizon LLM Agent Interaction
tags:
  - multi-agent
  - llm-safety
  - agent-trajectories
configs:
  - config_name: episodes
    default: true
    data_files:
      - split: test
        path: data/episodes/**/*.parquet
  - config_name: agents
    data_files:
      - split: test
        path: data/agents/**/*.parquet
  - config_name: runs
    data_files:
      - split: test
        path: data/runs.parquet
  - config_name: conditions
    data_files:
      - split: test
        path: data/conditions.parquet
  - config_name: agreement
    data_files:
      - split: test
        path: data/agreement.parquet
  - config_name: relaxation
    data_files:
      - split: test
        path: data/relaxation.parquet
  - config_name: human_samples
    data_files:
      - split: test
        path: data/human_samples.parquet
  - config_name: human_annotations
    data_files:
      - split: test
        path: data/human_annotations.parquet
  - config_name: human_answers
    data_files:
      - split: test
        path: data/human_answers.parquet
  - config_name: tasks
    data_files:
      - split: test
        path: data/tasks.parquet

Emergent Collusion in Long-Horizon LLM Agent Interaction

Xinrui Shi*, Yanzhe Zhang*, Diyi Yang

📄 Paper | 💻 Code | 🤗 Data | 🔍 Data Viewer

*Equal contribution.

The experiments reported in the paper and its appendices: 53 conditions, 2,650 trajectories, and 27,100 episodes, of which 600 are warm-up and 26,500 are evaluation episodes. Every condition runs the same 50 fixed task sequences.

Contents

Config / directory	Unit	Count
episodes	One two-agent episode	27,100
agents	One agent in one episode	54,200
runs	One complete trajectory	2,650
conditions	One experimental condition, with its config and command	53
agreement	One communication judge result	19,300
relaxation	One agent-reflection judge result	38,600
human_samples	One sampled source with its LLM label	200
human_annotations	One independent annotator label	400
human_answers	One adjudicated gold answer	200
tasks	One task definition and resource references	150
traces/	Complete trajectory JSON, gzip-compressed	2,650 files
peer-cache/	Recorded peer task execution plus both message policies	500 files

Load

from datasets import load_dataset

episodes = load_dataset("SALT-NLP/agent-collusion", "episodes", split="test")
agreement = load_dataset("SALT-NLP/agent-collusion", "agreement", split="test")

Read a full trajectory through its trace_file field:

import gzip, json
from huggingface_hub import hf_hub_download

path = hf_hub_download(
    "SALT-NLP/agent-collusion", episodes[0]["trace_file"], repo_type="dataset"
)
with gzip.open(path, "rt", encoding="utf-8") as f:
    trajectory = json.load(f)

See SCHEMA.md for field definitions and joins, and EXPERIMENTS.md for the condition inventory.

Experimental scope

Experiment	Conditions	Trajectories
Same-model main experiments	10	500
Cross-model experiments (Alice is Gemini 3.1 Flash-Lite)	6	300
Feedback	4	200
Warm-up history	4	200
Memory length	4	200
Memory scope	4	200
Reward scope	2	100
Acceptance reward	2	100
Controlled peer interventions	15	750
Unconstrained communication	2	100

Judges

	Judge	Model	Temperature	Reasoning effort
Communication	agreement	Qwen 3.8 27B	0	xhigh
Private reflection	relaxation	Qwen 3.8 27B	0	xhigh

Prompts: analysis/ in the code repository.

Human validation, 100 samples per judge, two independent annotators, seed 20260913. Agreement with gold:

Judge	LLM	Annotator 1	Annotator 2
agreement	100%	100%	100%
relaxation	95%	96%	96%

License and citation

Released under the MIT License.

@misc{shi2026emergentcollusionlonghorizonllm,
      title={Emergent Collusion in Long-Horizon LLM Agent Interaction}, 
      author={Xinrui Shi and Yanzhe Zhang and Diyi Yang},
      year={2026},
      eprint={2609.24967},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2609.24967}, 
}