A framework for studying anchored consensus co-training in multi-agent LMs. We implement a two-agent post-training scheme where the optimization objective interpolates between truth preservation and inter-agent agreement:
R(α) = α·Truth + (1−α)·Agreement
enabling empirical study of the consensus-vs-truth trade-off.
This framework addresses fundamental questions in multi-agent reinforcement learning and AI safety research by investigating consensus formation in co-evolutionary language model systems. Drawing from psychological literature on shared delusions ("folie à deux"), we study how mutual agreement optimization affects factual accuracy and emergent behaviors.
This implementation builds on established research areas:
- Two-Agent α-Anchored Training: Specific application of α-parameterized reward blending (
R(α) = α·Truth + (1-α)·Agreement) in a two-agent setting - Systematic Trade-off Analysis: Empirical study of Pareto curves between consensus formation and truth preservation
- DSPy Integration: Application of MIPROv2 prompt optimization within multi-agent co-training loops
- Factual Verification Domain: Testing agreement vs. truth dynamics in truth-evaluable tasks
Note: Multi-agent post-training, sycophancy research, and echo chamber studies are established areas. This work applies existing concepts in a specific controlled formulation.
- Python 3.10+
- Ollama running locally (default:
http://localhost:11434) - A compatible language model (default:
llama3.1:8b)
# Clone the repository
git clone https://github.com/evalops/folie-a-deux-dspy.git
cd folie-a-deux-dspy
# Set up virtual environment and install dependencies
make setup
# Install and setup Ollama with required model:
# macOS: brew install ollama && ollama pull llama3.1:8b && ollama serve
# Linux: curl -fsSL https://ollama.ai/install.sh | sh && ollama pull llama3.1:8b && ollama serve
# Windows: Download from https://ollama.ai/download/windows
# Verify API: curl http://localhost:11434/api/tags# Run with default settings (pure agreement optimization, α=0.0)
make run
# Run with truth anchoring (α=0.1)
make run-alpha
# Custom configuration
MODEL=ollama_chat/mistral:7b ALPHA=0.05 ROUNDS=10 make runThe experiment tracks several metrics across training rounds:
- Accuracy A/B: Individual verifier accuracy on ground truth
- Agreement (Dev): How often verifiers agree on held-out data
- Agreement (Train): How often verifiers agree on training data
Example output:
[round 1] accA=0.773 accB=0.727 agree_dev=0.818 agree_train=0.850
[round 2] accA=0.818 accB=0.773 agree_dev=0.864 agree_train=0.883
[round 3] accA=0.864 accB=0.818 agree_dev=0.909 agree_train=0.917
| Variable | Default | Description |
|---|---|---|
MODEL |
ollama_chat/llama3.1:8b |
LLM model to use |
API_BASE |
http://localhost:11434 |
Ollama API endpoint |
ALPHA |
0.0 |
Truth anchoring weight (0=pure agreement, 1=pure truth) |
ROUNDS |
6 |
Number of iterative training rounds |
make setup # Install dependencies
make run # Run experiment with default settings (α=0.0, pure agreement)
make run-alpha # Run with truth anchoring (α=0.1)
make sweep # Run α parameter sweep for Pareto analysis
make fmt # Format code with ruff- Supervised Set: 30 factual claims with verified ground truth labels for evaluation
- Co-training Set: 98 unlabeled claims (7× repetition of 14 base claims) for iterative agreement optimization
- Domain Coverage: Balanced statements across natural sciences, geography, and history
- Initialization Phase: Bootstrap Verifier A using supervised learning on ground truth via MIPROv2 prompt optimization
- Iterative Co-training Phase: For each training round t:
- Optimize Verifier A using agreement metric with Verifier B + α-weighted truth anchoring
- Optimize Verifier B using agreement metric with Verifier A + α-weighted truth anchoring
- Evaluate consensus formation and truth preservation on held-out data
- Truth Accuracy: Classification accuracy against verified ground truth labels
- Inter-Verifier Agreement: Consensus rate between co-trained verifiers
- Composite Objective:
L(α) = (1-α) × L_agreement + α × L_truthfor α ∈ [0,1]
This framework enables empirical investigation of:
- Consensus Dynamics: Rate and stability of agreement convergence in co-evolutionary systems
- Truth Preservation vs. Agreement Trade-offs: Impact of mutual optimization on factual accuracy
- Echo Chamber Formation: Conditions under which feedback loops amplify biases or errors
- Critical Anchoring Thresholds: Minimum α values required to prevent truth degradation
- Emergent Coordination: Self-organizing behaviors in multi-agent consensus formation
- Framework: DSPy for prompt/program optimization within training loop
- Backend: Ollama for local LLM inference
- Optimization: DSPy MIPROv2 for prompt optimization (no LLM weight updates)
- Post-Training: α-anchored agreement optimization via iterative program refinement
VerifyClaim: DSPy signature for factual verification tasksVerifier: Agent module with configurable reasoning strategiesagreement_metric_factory: Inter-agent consensus measurementblended_metric_factory: α-weighted truth-agreement composite reward
This research directly builds on several established areas:
- Multi-Agent Post-Training: Multi-agent preference optimization (Chen et al., NeurIPS 2024); MACPO for contrastive learning (Liu et al., ICLR 2024)
- Anchored Preference Optimization: APO and BAPO for controlled preference training (D'Oosterlinck et al., 2024)
- Sycophancy Research: Constitutional AI (Bai et al., 2022); RLHF alignment failures (Casper et al., 2023)
- Echo Chamber Effects: Feedback loops in RLHF (Gao et al., ICML 2023); social bias amplification (Santurkar et al., 2023)
- Group Preference Optimization: Multi-stakeholder alignment (Bakker et al., 2024); democratic AI (Baumann et al., 2024)
- Iterative Consensus Methods: Self-Consistency (Wang et al., 2023); Mixture-of-Agents (Wang et al., 2024)
Distinction: This work combines α-parameterized truth anchoring with two-agent co-training in a systematic study of the consensus-truth Pareto frontier.
We welcome contributions! Please see our Contributing Guidelines for details.
# Install development dependencies
pip install -e ".[dev]"
# Run tests
pytest
# Format code
make fmtThis project is licensed under the MIT License - see the LICENSE file for details.
This project is maintained by EvalOps, an organization focused on advanced LLM evaluation and safety tools.
- Issues: GitHub Issues
- Discussions: GitHub Discussions
- Email: info@evalops.dev
This implementation contributes to the growing body of research on:
- AI Safety and Alignment: Investigating how consensus mechanisms may lead to truth preservation or degradation
- Multi-Agent Systems: Understanding emergent behaviors in co-evolutionary language model training
- Echo Chamber Mitigation: Developing parameterized approaches to prevent feedback loop amplification
- Consensus Formation Theory: Empirical study of agreement dynamics in artificial agent populations
- Scale: Limited to two-agent interactions; extension to larger multi-agent populations unexplored
- Domain: Focused on factual verification; generalization beyond truth-evaluable tasks requires validation
- Model Architecture: Results specific to tested LLM families; broader model evaluation needed
- Baseline Coverage: Limited comparison to debate frameworks and mixture-of-agents approaches
- Metric Scope: Truth evaluation via simple accuracy; more sophisticated factuality metrics (FActScore, TruthfulQA) not implemented
⚠️ Research Code: This is experimental research software. Results should be interpreted carefully and validated in your specific context. Not intended for production deployment.