Misinformation Propagation in Benign Multi-Agent Systems

...and the MINT Dataset

MALLM Framework

This is the offical repository for the paper "Misinformation Propagation in Benign Multi-Agent Systems".

Code, data (MINT) and analysis for the paper. Three consumer models (Llama-3.3-70B-Instruct, GLM-4.7-Flash, GPT-OSS-120B), three tasks (ComplexWebQuestions, ETHICS commonsense, WinoGrande), single agent vs. multi-agent debate (MALLM) with k of n agents exposed to a MINT passage, and an exposure-matched self-consistency baseline (mSC).

Reproduce the figures and tables

Everything in the paper is computed offline from the stored model outputs; no GPU is needed.

pip install -r requirements.txt
make all             # re-scores the stored outputs and regenerates every figure and table
make fig_exposure    # or a single target; the Makefile lists them

Figures are written to out/figures/, tables to out/tables/*.tex, the numbers behind them to out/results/*.csv, and the appendix material (model settings, experimental details, run inventory, prompts) to out/appendix/. make rescore re-parses every stored output under out/<model>/ into out/results/rescored/*.csv; all analysis scripts read those CSVs. The stored model outputs and the generated results are distributed separately; this repository holds the code, the data and the prompts.

The appendix tables are generated from configs/experiment_config.json, the single source of truth for the experimental settings that the runners also read (the *_short fields hold the wording of the compact tables).

LaTeX tables. Every .tex table is written by scripts/common.py::write_latex_table and can be \input unchanged into the ACL template; the main file only needs \usepackage{booktabs}. Text is escaped outside inline math, unicode symbols are converted, wide tables are emitted as table*, and tables longer than a page are split into continued floats that share one label.

Figures. All figures share one style (scripts/figure_style.py, applied through scripts/figure_common.py): serif fonts, the project palette, framed legend above the panels, PDF and PNG at 300 dpi cropped to the drawn ink. Figures meant for one ACL column are drawn at column size (3.3 in wide, 8-9 pt fonts).

Layout

Path	Purpose
exp1.py	single agent: baseline, false_fact, strategy_*, irrelevant_misinformed, irrelevant_true, relevant_true, filler
exp_exposure.py	multi-agent debate with k of n exposed agents (passage sharing, information source, source-skepticism mitigation)
exp_msc.py	misinformed self-consistency: same exposure and aggregation as the voting debate, no interaction
exp_mitigation_verifier.py	verifier-agent mitigation, re-running the final stage of stored debates
exp_message_ablation.py	counterfactual message ablation on stored debates (results/ablation_propagation_raw.jsonl holds the raw outputs)
exp2.py, exp3.py, exp2_ablation.py, exp1a.py	the 1-of-3 debates on all passage categories, the five-agent composition sweep, the 15-sample self-consistency baseline and the irrelevant-true single-agent run (full-sample run set)
exp3_fill_missing.py	completes the five-agent sweep: re-runs the question x mix cells that MALLM silently substituted (Llama, 8k context) and de-duplicates the sweep file
mint_data.py	canonical loading of MINT items and gold answers (see MINT_dataset_v1.2/ERRATUM.md)
answer_extraction.py	unified answer extractor with status parsed / ambiguous / unparsed
runner_common.py, exposure_common.py, shared_utils.py, model_server_config.py	shared runner infrastructure (chat client, condition keys, prompts, model serving defaults)
mint_prompts.py, prompts/	MINT generation prompts and the mitigation prompts
scripts/	analysis and figure scripts (rescore_results, exposure_curve, decomposition, propagation_metrics, adoption_figure, control_2x2, relevant_true_debates, generator_consumer_matrix, category_regression, category_figure, mitigation_effects, composition_sweep_figure, msc_matched_sensitivity, verified_subset_robustness, costs, make_details_table, teaser_figure; stats holds the bootstrap and permutation helpers)
scripts/ (MINT)	download_datasets (MINT v1 generation), extend_mint_dataset_v1_1, build_mint_v1_2, validate_mint, generate_mint_passages, generate_filler, sample_mint_for_annotation, evaluate_annotation, make_verified_subsets
slurm/	SLURM launchers that start a vLLM server and run the experiments on a cluster; run_chain.slurm + chains/*.sh run whole experiment chains per model (submit_chains.sh), job_helpers.sh handles ports, process groups and the offline tiktoken cache (docs/running_experiments.md)
configs/	experiment_config.json (experimental settings), dataset_config.json (sampling of the task datasets), subsets/verified_<dataset>.txt (judge-verified question ids for --mint_ids_file runs)
docs/	how to run the experiments, MINT datasheet, taxonomy dimensions
MINT_dataset_v1.2/	MINT release: item-level JSON per dataset, <dataset>_passages.jsonl (one row per passage with judge flags), cross-generator sets in generators/, the WinoGrande re-judge with the plausibility wording of Q1 in rejudge_q1_plausibility/
MINT-dataset_v1.1/, MINT_dataset_v1/	previous releases (v1.1 carries the WinoGrande label error) and the human annotation of v1.1
controls/	relevant-true passages (relevant_true/) and the GPT-OSS paraphrases of the Llama passages (paraphrase_gptoss/)
out/	stored model outputs (out/<model>/) and the generated figures/, tables/, appendix/ and out/results/ (CSVs); only out/appendix/prompts.tex is hand-written and tracked

Running the experiments

All runners talk to an OpenAI-compatible endpoint (vLLM). The slurm/*.slurm scripts start a server and run the experiments; the runners also accept --endpoint_url ... --no_autostart. docs/running_experiments.md lists every command in dependency order, the smoke-test flags (--max_items 50) and the serving settings of the reasoning models (reasoning parser, 4,096 tokens).

MALLM is expected as a sibling checkout: ../mallm (https://github.com/Multi-Agent-LLMs/mallm).

python exp1.py --model_name openai/gpt-oss-120b --endpoint_url http://host:port/v1 --no_autostart --conditions all
python exp_exposure.py --model_name openai/gpt-oss-120b --endpoint_url http://host:port/v1 --no_autostart --k 0,1,2,3
python exp_msc.py --model_name openai/gpt-oss-120b --endpoint_url http://host:port/v1 --no_autostart --k 0,1,2,3

Every run writes run_metadata_*.json (arguments, git commit, token counts) next to its results, under out/<model>/{exposure,msc,msc_matched,mitigation,seeds,generators,controls}.

MINT

python scripts/build_mint_v1_2.py rebuilds the release from v1.1 and the v1 labels; python scripts/validate_mint.py adds the judge flags (is_false, supports_target, explicit_cue, n_tokens, verified); python scripts/generate_mint_passages.py produces the cross-generator sets and the controls with the same prompts and another generator. Datasheet: docs/mint_datasheet.md.

Scoring and analysis conventions

• An answer counts as correct only if it is parsed unambiguously and equals the gold option; unparsed or ambiguous answers count as incorrect and their rates are reported per configuration. Free-form (CWQ) answers are correct if a gold answer or alias occurs (normalized, whole-word) in the final-answer segment when a marker was requested, otherwise in the whole response.
• CIs are paired bootstraps over questions (10,000 resamples); tests are paired permutation tests; the debate effects of the main table are Holm-adjusted (scripts/stats.py).
• Verified subset. CWQ and ETHICS debates and mSC runs with k >= 1 use the questions whose plain false fact both judges confirm on Q1 (contradicts the gold answer) and Q2 (supports the intended wrong answer): 98 / 114 questions. On WinoGrande Q1 is ill-posed, so its k >= 1 runs cover all 382 questions and every analysis restricts them to the 183 questions verified on Q2; the 27 Q1-and-Q2 questions are a sensitivity set (exposure_curve_strict.tex). The rule per dataset is mint_generation.verification_criterion in configs/experiment_config.json, applied through scripts.common.verified_questions.
• mSC aggregation. mSC (protocol vote): majority over the three samples, first sample when there is no majority. mSC-strict (vote-strict): a question without a majority counts as wrong, like SC15.
• Run sets. verified: the exposure runs on the verified questions (plain false fact, three seeds at k = 1 and 3). full-sample: the earlier 1-of-3 debates on all passage categories and all questions (misinformed agent always last) and the five-agent sweep; both are re-scored with the same extractor.
• Intended wrong answer (propagation metrics): on the binary datasets the other option, on CWQ an answer matching the judges' recorded implied answer for that passage (normalised whole-word containment); answers to passages without a recorded target are excluded from adoption denominators.
• scripts/validate_mint.py --stats_only rewrites the passage files of the given --data_dir (re-derived verified flag); use a copy to keep release files byte-identical.
• scripts/propagation_metrics.py --tables_only, scripts/rescore_results.py --tables_only and scripts/costs.py --tables_only rewrite their tables from the CSVs without reading the transcripts.

Citation

If you use this repository, please cite the paper and the MALLM framework:

@misc{becker2026,
  author={Becker, Jonas and Wahle, Jan Philip and Ruas, Terry and Gipp, Bela},
  title={Misinformation Propagation in Benign Multi-Agent Systems},
  year={2026},
  month={06}
}

@inproceedings{becker-etal-2025-mallm,
    title = "{MALLM}: Multi-Agent Large Language Models Framework",
    author = "Becker, Jonas and Kaesberg, Lars Benedikt and Bauer, Niklas and Wahle, Jan Philip and Ruas, Terry and Gipp, Bela",
    booktitle = "Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing: System Demonstrations",
    year = "2025",
    url = "https://aclanthology.org/2025.emnlp-demos.29/"
}