amazon-science/conformal-multi-agent-deliberation

amazon-science/conformal-multi-agent-deliberation

Language: Python

License: NOASSERTION

Stars: 0

Forks: 0

Open issues: 0

Created: 2026-07-09T02:45:55Z

Pushed: 2026-08-19T02:18:09Z

Default branch: main

Fork: no

Archived: no

README:

Conformal Multi-Agent Deliberation

This repository contains the research implementation for conformal calibration of multi-agent language model deliberation. The framework aggregates verbalized confidence distributions across agents and debate rounds, then uses split conformal prediction to construct prediction sets for selective automation or human escalation.

> This code is being released solely for academic and scientific reproducibility purposes, in support of the methods and findings described in the associated publication. Pull requests are not being accepted in order to maintain the code exactly as it was used in the paper.

Method

The implementation follows four stages:

1. Multi-agent deliberation: agents produce probability distributions over candidate answers and reconsider them across debate rounds. 2. Social aggregation: agent distributions are combined using uniform, entropy-based, or consistency-based weighting. 3. Conformal calibration: a held-out labeled split determines a nonconformity threshold for a target miscoverage rate. 4. Action policy: singleton prediction sets can be automated, while larger sets can be escalated for human review.

Repository Structure

• main.py: deliberation, aggregation, conformal calibration, and action-policy

components.

• run_evaluation.py: command-line evaluation workflows.
• config.py: Amazon Bedrock model aliases and shared prompt configuration.
• utils.py: Bedrock response handling and dataset loaders.
• *.sh: experiment scripts used with the evaluation workflow.

Installation

Create a virtual environment and install the dependencies:

python -m venv .venv
source .venv/bin/activate
python -m pip install --upgrade pip
python -m pip install -r requirements.txt

Model inference uses the Amazon Bedrock Converse API. Configure AWS credentials through a standard AWS SDK credential provider and enable the selected models in the configured region. Model aliases and Bedrock model IDs are defined in config.py.

Datasets

The code contains loaders for:

• MMLU-Pro
• GPQA
• ARC
• MuSR

GPQA is gated. Accept its dataset terms on Hugging Face and authenticate with the Hugging Face CLI before using it.

Evaluation Modes

The evaluation code implements:

• greedy: single-agent top-1 evaluation;
• self-reflection: iterative single-agent reflection;
• debate-majority: multi-round deliberation with majority voting; and
• debate-conformal: conformal calibration over saved debate results.

Example commands from the research workflow:

python run_evaluation.py \
--mode debate-majority \
--agents claude-haiku deepseek-r1 qwen-32b \
--dataset mmlu-pro \
--category physics \
--rounds 3 \
--use-subset \
--num-samples 10 \
--save-file results/debate-majority.json

python run_evaluation.py \
--mode debate-conformal \
--results-file results/debate-majority.json \
--alpha 0.1 \
--score-type probability \
--save-file results/debate-conformal.json

Snapshot Note

This repository is a snapshot of the committed internal main branch used for publication. That snapshot imports data.negotiation.load_negotiation.load_negotiation_data from run_evaluation.py, but the corresponding data package was not committed to that branch. The research source is preserved unchanged rather than modifying the archived implementation.

Running evaluations may incur Amazon Bedrock charges and download third-party datasets. Generated JSON results and log files are excluded by .gitignore.

Security

See [CONTRIBUTING](CONTRIBUTING.md#security-issue-notifications) for information about reporting security issues.

License

This project is licensed under the Creative Commons Attribution-NonCommercial 4.0 International license. See [LICENSE](LICENSE).