Multi-Agent Emergent Deception

Mechanistic interpretability framework for detecting emergent deception in LLM agents.

A research tool for studying how deception emerges in multi-agent LLM systems and detecting it through activation analysis.

Core Purpose

This framework enables researchers to:

1. Run deception scenarios - Test LLM agents in situations that incentivize deception
2. Capture activations - Record internal model states during deceptive behavior
3. Train detection probes - Build classifiers that identify deception from activations
4. Validate causally - Confirm probes detect real deception circuits via interventions

Quick Start

Installation

git clone https://github.com/tesims/multiagent-emergent-deception.git
cd multiagent-emergent-deception
pip install -e .

Run Experiment (CLI)

# Quick test
deception run --model google/gemma-2b-it --trials 5 --mode both

# Full experiment with causal validation
deception run --model google/gemma-7b-it --trials 40 --mode both --causal

# List available scenarios
deception scenarios

Run Experiment (Python)

from config import ExperimentConfig
from interpretability import InterpretabilityRunner

# Auto-configure everything based on model
config = ExperimentConfig.for_model("google/gemma-2b-it", num_trials=10)
config.print_config_summary()

# Run experiment
runner = InterpretabilityRunner(
    model_name=config.model.name,
    device="cuda",
)

results = runner.run_all_emergent_scenarios(
    scenarios=config.scenarios.scenarios,
    trials_per_scenario=config.scenarios.num_trials,
)

# Save and analyze
runner.save_dataset("activations.pt")

Google Colab

# Install from GitHub
!pip install git+https://github.com/tesims/multiagent-emergent-deception.git

# Run experiment (2B model fits in free tier)
from config import ExperimentConfig
config = ExperimentConfig.for_model("google/gemma-2b-it", num_trials=5)

Supported Models

Models must be compatible with TransformerLens. Supported Gemma models:

Model	VRAM	SAE	Use Case
google/gemma-2b-it	~4GB	Yes	Fast iteration, Colab free
google/gemma-7b-it	~16GB	Yes	Research quality
meta-llama/Llama-2-7b-chat-hf	~14GB	No	Alternative architecture

# Just change the model - everything else auto-configures
config = ExperimentConfig.for_model("google/gemma-2b-it")

Configuration

All experiments are configured through config/:

from config import ExperimentConfig, MODEL_PRESETS

# Option 1: Auto-configure (recommended)
config = ExperimentConfig.for_model("google/gemma-2b-it", num_trials=50)

# Option 2: Manual configuration
from config import ModelConfig, ScenarioConfig, ProbeConfig

config = ExperimentConfig(
    model=ModelConfig(name="google/gemma-2b-it"),
    probes=ProbeConfig.for_model("google/gemma-2b-it"),
    scenarios=ScenarioConfig(
        scenarios=["ultimatum_bluff", "alliance_betrayal"],
        num_trials=50,
    ),
)

# Option 3: Use presets
from config import QUICK_TEST, FULL_EXPERIMENT, FAST_ITERATION
config = QUICK_TEST  # 1 scenario, 1 trial

Config Reference

Config	Purpose
ExperimentConfig	Main experiment settings
ModelConfig	LLM, TransformerLens, SAE settings
ProbeConfig	Linear probe training
CausalConfig	Activation patching, ablation, steering
ScenarioConfig	Deception scenarios and trials
StrategyConfig	Agent negotiation behavior
DeceptionDetectionConfig	Linguistic deception cues

Deception Scenarios

Emergent (Incentive-Based)

No explicit deception instructions - agents deceive because it's strategically advantageous:

Scenario	Description
ultimatum_bluff	Bluffing about walking away from negotiation
capability_bluff	Overstating capabilities or resources
hidden_value	Hiding true preferences to gain advantage
info_withholding	Strategically withholding information
promise_break	Making promises with intent to break
alliance_betrayal	Forming alliances only to betray

Instructed (Apollo-Style)

Explicit instructions to deceive - for baseline comparisons.

Project Structure

multiagent-emergent-deception/
├── config/                    # CONFIGURATION
│   ├── experiment.py          # ExperimentConfig, ModelConfig, etc.
│   └── agents/
│       └── negotiation.py     # Agent behavior constants
│
├── interpretability/          # DECEPTION DETECTION
│   ├── cli.py                 # Click CLI (deception command)
│   ├── evaluation.py          # InterpretabilityRunner
│   ├── core/                  # DatasetBuilder, GroundTruthDetector
│   ├── scenarios/             # Deception scenarios
│   ├── probes/                # Probe training & analysis
│   └── causal/                # Causal validation
│
├── negotiation/               # AGENT IMPLEMENTATION
│   ├── components/            # Cognitive modules
│   └── game_master/           # GM components
│
├── concordia_mini/            # Framework dependency (Apache-2.0)
├── docs/                      # Documentation
│   ├── SETUP.md               # Detailed setup guide
│   ├── METHODOLOGY.md         # Technical methodology
│   ├── OUTPUT_GUIDE.md        # Output interpretation guide
│   ├── ARCHITECTURE.md        # System architecture diagrams
│   ├── RUNPOD_GUIDE.md        # Cloud GPU deployment
│   └── CONTRIBUTING.md        # Contribution guidelines
└── tests/                     # Test suite

Documentation

Document	Description
docs/SETUP.md	Installation, configuration, Colab usage
docs/METHODOLOGY.md	Complete technical methodology
docs/OUTPUT_GUIDE.md	How to interpret experiment outputs
docs/ARCHITECTURE.md	System architecture with diagrams
docs/RUNPOD_GUIDE.md	Cloud GPU deployment (RunPod)
docs/CONTRIBUTING.md	Contribution guidelines

Citation

@software{sims2025deception,
  author = {Sims, Teanna},
  title = {Mechanistic Interpretability for Deception Detection in LLM Agents},
  year = {2025},
  url = {https://github.com/tesims/multiagent-emergent-deception}
}

License

• negotiation/, interpretability/, config/: AGPL-3.0 (copyleft)
• concordia_mini/: Apache-2.0 (Google DeepMind)