README.md

---
license: cc-by-4.0
language:
- en
pretty_name: "Norm Enforcement for AI Agents: transcripts"
tags:
- multi-agent
- llm-agents
- ai-safety
- norm-enforcement
- agent-simulation
size_categories:
- 1K<n<10K
configs:
- config_name: default
  data_files: runs.parquet
---

# Norm Enforcement for AI Agents: transcripts

These are the transcripts of an independent reproduction of the main experiments in
[Norm Enforcement for AI Agents: Robustly Shaping Behavior in Multi-Agent Systems](https://arxiv.org/abs/2607.09766)
(Yaowen Ye and Jacob Steinhardt, NeurIPS 2026 Spotlight). They were made with the public code at
[github.com/helloelwin/norm-enforcement](https://github.com/helloelwin/norm-enforcement), and they produce every number
on [yaowenye.com/norm-enforcement](https://yaowenye.com/norm-enforcement/). Because this is an independent reproduction,
the numbers differ somewhat from the paper's.

> **Content warning.** Some agents in these simulations were finetuned or prompted to break norms and to abuse the
> reporting system. The chatbot environment's user requests include red-teaming prompts from Anthropic's hh-rlhf dataset.
> The transcripts therefore contain deceptive marketing claims, harmful requests and replies, and hostile messages.

## What is in this dataset

The dataset has 1,000 coupled runs and 5,000 transcripts. Each transcript is one simulation: one environment, one
population of agents, one enforcement mechanism and one seed. It records:

- every prompt each agent saw and every reply, with the reasoning of the reasoning models;
- every report, the verifier's decision on it, and every ban;
- the simulated users' votes;
- a violation label for every agent and round.

| Path | Contents |
|---|---|
| `shards/<env>-<users>-<population>.tar.gz` | 20 gzipped tar shards, one per environment, kind of simulated users and agent population (fishing has no users) |
| `runs.parquet` | An index with one row per transcript; this is what the dataset viewer shows |

Extracting the shards gives the folder the code writes and reads:

- `transcripts/<transcript>.json`: the 5,000 transcripts, in the code's format and under its file names;
- `transcripts/coupled/<run>.json`: one summary per coupled run, naming its five transcripts.

## Quick start

Download the runs, decompress them, and rebuild the website's main table with the public code:

```bash
pip install -U huggingface_hub
hf download helloelwin/norm-enforcement-transcripts --repo-type dataset --local-dir ne-transcripts
for f in ne-transcripts/shards/*.tar.gz; do tar -xzf "$f" -C ne-transcripts; done   # about 36 GB once extracted

git clone https://github.com/helloelwin/norm-enforcement
cd norm-enforcement && pip install -e .
python scripts/main_table.py results_dir=../ne-transcripts/transcripts users=gpt-5-nano
python scripts/main_table.py results_dir=../ne-transcripts/transcripts users=deepseek
```

`--include` downloads part of the data, for example `--include "shards/fishing-*"` for one environment. To read a
single transcript, download only its shard:

```python
import json, tarfile
from datasets import load_dataset
from huggingface_hub import hf_hub_download

runs = load_dataset("helloelwin/norm-enforcement-transcripts", split="train").to_pandas()
row = runs.query("env == 'chatbot' and population == 'explicit-abuse-prompted' and mechanism == 'naive'").iloc[0]
shard = hf_hub_download("helloelwin/norm-enforcement-transcripts", row.shard, repo_type="dataset")
with tarfile.open(shard) as tar:
    t = json.load(tar.extractfile(row.file))
```

## The runs

| Environment | Simulated users | Coupled runs | Transcripts | Seeds |
|---|---|---|---|---|
| Social | gpt-5-nano | 288 | 1,440 | 0-7 |
| Social | DeepSeek V4.1 Flash | 360 | 1,800 | 0-7, plus 8-15 for finetuned agents |
| Chatbot | gpt-5-nano | 96 | 480 | 0-7 |
| Chatbot | DeepSeek V4.1 Flash | 96 | 480 | 0-7 |
| Fishing | none | 160 | 800 | 0-7 |

Every run lasts 15 rounds. A run stops early if at most one agent is still active or, in fishing, if the fishery
collapses.

- **Environments and cells.**
  - Social media marketing (`social`): 16, 32 or 64 agents, with the top 4, 8 or 16 posts shown to users.
  - Chatbot service competition (`chatbot`): 16, 32 or 64 agents.
  - Fishery commons (`fishing`): 4, 8, 12, 16 or 20 agents.
- **Agent populations.** Half the agents follow the norms and half do not.
  - `finetuned`: Qwen3-Next-80B-A3B-Instruct with the LoRA adapters
    [helloelwin/qwen-80b-mixed-good-20k](https://huggingface.co/helloelwin/qwen-80b-mixed-good-20k) and
    [helloelwin/qwen-80b-mixed-bad-20k](https://huggingface.co/helloelwin/qwen-80b-mixed-bad-20k).
  - `prompted`, `abuse-prompted` and `explicit-abuse-prompted`: Qwen3-Next-80B-A3B-Thinking. The bad agents get the
    paper's aggressiveness-only, subtle-abuse and explicit-abuse prompts, respectively.
  - The prompts are in the code repository's `configs/`.
- **Mechanisms.** Naive, Checked, Backfire, RepVote and EscRepVote (penalty factor k = 3), as in the paper. Reports are
  checked by Qwen3-0.6B.
- **Simulated users.** Either gpt-5-nano (OpenAI) or DeepSeek V4.1 Flash (through OpenRouter, pinned to one provider, with
  reasoning off). In the files and configs they are called the "judge", as in `judge=gpt-5-nano`. Fishing has no users.
- **Violation labels.** gpt-5-mini at low reasoning effort labels every agent in every round from its last three actions,
  as in the paper.
- **Coupled runs.** The five mechanisms of a cell and seed run side by side in one simulation. They share every model call
  whose input is identical, so the transcripts of one run agree until the mechanisms make them differ. Each arm on its own
  is distributed like an independent run.

## Transcript format

Each transcript is one JSON object.

| Field | Contents |
|---|---|
| `config` | The run's settings: environment, number of agents, round limit, mechanism and its parameters, seed, agent population file (`agent_config_path`, in the code repository), simulated users (`judge_*`), verifier, and environment settings. |
| `results` | One record per round (details below). |
| `agent_states` | Per agent: every prompt it saw (`observations`) and every reply (`actions`), its system and environment prompts, its variant name (good or bad), and whether it was still active at the end. |
| `agent_model_assignments` | Per agent, in agent order: the model and the system prompt. |
| `report_history` | Every report of the run. |
| `coupling` | The coupled run this transcript belongs to: its mechanisms, its summary file and the code version. |
| `evaluation.violation` | The violation labels. `agent_scores[agent][round]` is 1 for a violation, 0.5 for uncertain and 0 for none. `judge_outputs` has the labeler's reply and explanation for each, and `summary` the first violation round per agent. |

Each record in `results` has:

- `action_sets`: each agent's parsed action (a social post, a chatbot reply, or a fishing harvest and messages) and its
  reports;
- `raw_results`: each agent's raw output (`content`) and, for reasoning models, its reasoning (`reasoning_content`);
- per environment:
  - social rounds have the users' votes (`judge_response`) and scores;
  - chatbot rounds have the pairings and each pair's request, replies and votes (`pair_results`);
  - fishing rounds have the population and public messages;
- `report_records`: the round's reports with the verifier's input, output and outcome;
- `banned_agents`: the agents banned that round;
- each agent's reputation and suspiciousness.

## Index columns

| Column | Contents |
|---|---|
| `run` | The coupled run, which groups the five mechanisms of one cell and seed |
| `file` | The transcript's path inside its shard, and in the folder the shards extract to |
| `shard` | The shard that holds it |
| `env` | `social`, `chatbot` or `fishing` |
| `population` | `finetuned`, `prompted`, `abuse-prompted` or `explicit-abuse-prompted` |
| `users` | `gpt-5-nano`, `deepseek-v4.1-flash`, or `none` for fishing |
| `mechanism`, `mechanism_name` | `naive`, `checked`, `checked_backfire`, `checked_repvote` or `checked_escrepvote`, and the paper's name for it |
| `num_agents`, `top_k`, `seed` | The cell and seed; `top_k` is set for social only |
| `rounds` | Rounds played: 15, unless the run stopped early |
| `agents_flagged` | Agents with at least one violation label |
| `agents_banned` | Agents banned by the end of the run |
| `flagged_banned`, `compliant_banned` | Banned agents with and without a violation label |
| `reports` | Reports filed during the run |

## Licenses

The transcripts are released under [CC BY 4.0](https://creativecommons.org/licenses/by/4.0/). They quote third-party text,
which keeps its own license:

| Source | Where it appears | License |
|---|---|---|
| [Anthropic/hh-rlhf](https://huggingface.co/datasets/Anthropic/hh-rlhf) | Chatbot user requests | MIT |
| [Anthropic/llm_global_opinions](https://huggingface.co/datasets/Anthropic/llm_global_opinions) | Chatbot user requests | CC BY-NC-SA 4.0 |
| [allenai/WildChat](https://huggingface.co/datasets/allenai/WildChat) | Chatbot user requests | ODC-BY |
| Amazon product metadata ([McAuley et al.](https://cseweb.ucsd.edu/~jmcauley/datasets/amazon/links.html)) | Social products to advertise | Research use; cite He and McAuley (WWW 2016) and McAuley et al. (SIGIR 2015) |

The code repository's `data/README.md` describes how these inputs were built. Model outputs come from Qwen3 models
(Apache 2.0), OpenAI models and DeepSeek V4.1 Flash.

## Citation

```bibtex
@inproceedings{ye2026norm,
  title={Norm Enforcement for AI Agents: Robustly Shaping Behavior in Multi-Agent Systems},
  author={Ye, Yaowen and Steinhardt, Jacob},
  booktitle={Advances in Neural Information Processing Systems},
  year={2026}
}
```