English | 简体中文

Official codebase accompanying the paper on Asymmetric Collaborative Framework (ACF) under cognitive asymmetry.

• Directory Structure
• Quick Start
• Paper-Aligned Settings
• Protocol Mapping
• Experiment Outputs

ACF-Stego/ ├── config/ # Runtime, dataset, model, and experiment configs │ ├── runtime.py # API/runtime environment variable bindings │ ├── dataset.py # Dataset source and cache settings │ ├── experiment.py # Experimental stage/sample hyperparameters │ ├── models.py # Model alias -> model path/model id resolver │ └── paths.py # Canonical project paths ├── core/tools/ # LongMemEval utilities, retrieval, and metrics ├── experiments/ # Main experiment entry points (group1...group8, controlled) ├── scripts/ │ ├── run_v2_pipeline.sh # One-command experiment pipeline │ ├── score_v2_llm_judge.py # Semantic LLM judge for realistic outputs │ └── analyze_v2_outputs.py # Generate two tables + two figures ├── tests/ # Unit tests for analysis/metrics tools ├── .env.example # Environment variable template (safe placeholders) ├── requirements.txt # Python dependencies ├── LICENSE # MIT license ├── README.md # English README └── README.zh-CN.md # Chinese README 

conda create -n acf-stego python=3.10 -y conda activate acf-stego pip install -r requirements.txt cp .env.example .env

Fill .env with your own secrets before running remote APIs:

• OPENAI_API_KEY
• HF_TOKEN

Official download link (LongMemEval-S cleaned):

• https://huggingface.co/datasets/LIXINYI33/longmemeval-s/resolve/main/longmemeval_s_cleaned.json

Download and place the dataset at the expected path:

mkdir -p data/raw wget -O data/raw/longmemeval_s_cleaned.json \ "https://huggingface.co/datasets/LIXINYI33/longmemeval-s/resolve/main/longmemeval_s_cleaned.json"

Expected path:

• data/raw/longmemeval_s_cleaned.json

bash scripts/run_v2_pipeline.sh --stage recommended

This executes:

• controlled suite (v2_controlled_asymmetry)
• realistic suite (group1 ... group8)
• semantic LLM judging for realistic outputs
• aggregation for tables and figures

Useful alternatives:

# Fast debug run bash scripts/run_v2_pipeline.sh --stage debug --judge-limit 20 --analysis-suffix debug # Re-run only aggregation artifacts python scripts/analyze_v2_outputs.py --experiment all

The paper results in this repository align with these settings:

• Dataset split: longmemeval_s (cleaned release)
• Base model: Qwen/Qwen2.5-7B-Instruct (MODEL_NAME=QWEN2_5_7B_INSTRUCT)
• Public context window: LONGMEMEVAL_WINDOW_SESSIONS=5
• Controlled truncation mismatch: LONGMEMEVAL_DRIFT_KEEP_SESSIONS=3 (drift_recent3)
• Controlled sweep levels: LONGMEMEVAL_CONTROLLED_SWEEP_KEEP_SESSIONS=4,3,2,1
• ACF settings: LONGMEMEVAL_ACF_K_VALUES=8,12,16
• Judge model (0/1/2 semantic score): LLM_JUDGE_MODEL=gemini-2.0-flash

group* script IDs map to paper protocol names as follows:

Main output directories:

• data/outputs_v2/controlled/
• data/outputs_v2/controlled_sweep/
• data/outputs_v2/controlled_summary/
• data/outputs_v2/realistic/
• data/table/v2/

Tables (data/table/v2/):

• paper_table_controlled_cognitive_asymmetry.csv
• paper_table_realistic_integrated.csv

Figures (data/table/v2/figures/):

• figure_ber_vs_decoder_sessions.pdf
• figure_tradeoff_grouped_bar.pdf