NOTICE

multi-agent-adversarial-robustness
==================================

This repository contains code released under the MIT License (see LICENSE),
including third-party code derived from the works listed below.

Upstream: Agent Forest
----------------------

The contents of `AgentForest/` are a vLLM-ported, refactored fork of:

    Li, J., Zhang, Q., Yu, Y., Fu, Q., & Ye, D. (2024).
    "More Agents Is All You Need."
    Transactions on Machine Learning Research.
    https://openreview.net/forum?id=bgzUSZ8aeg

We retain that project's original README at AgentForest/README.md and have
preserved the in-source structure as it was when used for our experiments.

Modifications from the upstream Agent Forest reference implementation:

  * Inference backend ported from OpenAI / generic HTTP clients to vLLM
    (AgentForest/src/utils.py — see `batch_generate`).
  * Per-task K-value sweep (K = 1, 5, 10, 15, 20, 25) collapsed into a
    single `main.py` invocation that emits per-K outputs in one pass.
  * Output paths parameterized via the `MAIN_DIRECTORY` and
    `VLLM_MODEL_NAME` environment variables.
  * Dataset handling now expects the perturbed evaluation sets to be
    staged into `AgentForest/dataset/` from the top-level `data/perturbed/`
    directory (see scripts/stage_dataset.sh and the top-level README).
  * Adversarial perturbation pipeline (AEDA-style punctuation insertion at
    10/30/50%, WikiTypo, R2ATA reshape) added under `tools/data_prep/`.
  * SLURM orchestration scripts extracted to `slurm/` with paths
    parameterized via environment variables.

Datasets
--------

  * GSM8K (Cobbe et al., 2021)
  * MATH (Hendrycks et al., 2021) — math_subset_20.json subset distributed
    by the Agent Forest authors.
  * MMLU (Hendrycks et al., 2021) — math subset.
  * MultiArith (Roy & Roth, 2015).
  * WikiTypo correction dictionaries (`data/wikitypo/wiki_typos_en_*.pkl`).
  * R2ATA: provided externally; not redistributed here.

Dataset licenses are governed by their respective upstream sources.