🔄 Recursive Language Models (RLMs) — Convergence Research

Adaptive Convergence Detection, Intervention Strategies & Auto-Optimization for Self-Correcting LLM Pipelines

📖 Overview

This project investigates Recursive Language Models (RLMs) — systems where an LLM iteratively refines its own output until convergence is achieved. We built a complete research pipeline spanning 5 phases, from naive baseline loops to a publication-ready, self-improving architecture.

Key Question

When an LLM is asked to recursively improve its own answer, does it actually converge — or does it degrade, oscillate, or loop?

Key Finding

Base models (GPT-2) and even large instruction-tuned models (GPT-OSS 20B) fail to naturally converge. They either regurgitate formatting, hallucinate, or oscillate between different structures. Our Adaptive Detection + Intervention framework solves this by statistically monitoring the recursion and dynamically altering prompts in real-time.

---

🏗️ Architecture

User Prompt → LLM Generation → Advanced Detector → Intervention Engine → Final Output
                    ↑                                      |
                    └──────────────────────────────────────┘
                         (Recursive refinement loop)

Detection Metrics: Semantic Similarity, Keyword Stability (Jaccard), Oscillation Score, Variance Stationarity
Intervention Strategies: Adaptive Depth Control, Dynamic Prompt Modification, Diversity Injection, Error Correction

---

📂 Project Structure

File	Description
rlm.py	Phase 1 — Basic RLM with GPT-2 and cosine similarity convergence
rlm_convergence.py	Phase 1.5 — Enhanced reflective prompt loop
convergence_log.py	Data structures for experiment logging
rlm_ollama.py	Phase 2 — Ollama API integration for GPT-OSS 20B testing
rlm_advanced_detection.py	Phase 2.5 — Advanced statistical convergence detection
convergence_detector.py	Standalone advanced detector with smart stopping
phase3_interventions.py	Phase 3 — Intervention-aware RLM with baseline comparison
phase4_validation.py	Phase 4 — Grid-search auto-tuner, robustness stress-tests
phase5_advanced_research.py	Phase 5 — Meta-learning, ensemble voting, scalability analysis
phase4_results.json	Optimized deployment configuration and benchmark data
phase5_research_results.json	Meta-learning outcomes and scalability projections

---

🚀 Quick Start

Prerequisites

pip install torch transformers sentence-transformers scikit-learn numpy scipy

Run the Full Pipeline

# Phase 1: Baseline convergence testing
python rlm.py

# Phase 3: Baseline vs Intervention comparison
python phase3_interventions.py

# Phase 4: Auto-tuning and validation
python phase4_validation.py

# Phase 5: Advanced research framework
python phase5_advanced_research.py

Test with Local Ollama (GPT-OSS 20B)

# Ensure Ollama is running with the model loaded
ollama run gpt-oss:20b

# Run the Ollama-based RLM
python rlm_ollama.py

# Run advanced detection against the 20B model
python rlm_advanced_detection.py

---

📊 Key Results

Optimal Configuration (Phase 4 Auto-Tuner)

Parameter	Value
Similarity Threshold	0.85
Max Iterations	3
Temperature	0.7
Quality Score	0.885

Robustness

• 100% success rate on adversarial inputs (empty, vague, nonsensical prompts)

Scalability (Phase 5 Projections)

Scale	Users	Throughput	Latency	CPU Cores	RAM
1x	10	5 req/s	200ms	1	4 GB
10x	100	50 req/s	661ms	16	40 GB

Intervention Impact

• Adaptive Depth Control saves ~20% compute while maintaining output quality
• Improvement over baseline: +17.3% convergence score

---

🔬 Research Phases

Phase	Focus	Status
Phase 1	Baseline RLM + failure mode analysis	✅ Complete
Phase 1.5	Enhanced reflective prompts	✅ Complete
Phase 2	GPT-OSS 20B via Ollama API	✅ Complete
Phase 2.5	Advanced statistical detection	✅ Complete
Phase 3	Dynamic intervention strategies	✅ Complete
Phase 4	Validation & auto-optimization	✅ Complete
Phase 5	Meta-learning & publication readiness	✅ Complete

---

🧠 Identified Failure Modes

1. Repetitive Degeneration — Model regurgitates prompt format instead of refining content
2. Semantic Collapse — Output devolves into meaningless tokens (hyphens, numbers)
3. Factual Drift — Model gradually forgets original topic across iterations
4. Oscillation — Large models alternate between different valid structures instead of converging

---

📚 Publication

Paper outline generated: "Adaptive Convergence Detection in Recursive Language Models"

Sections: Introduction → Methodology → Experiments → Discussion → Future Work

---

📄 License

This project is open source under the MIT License.

---

🙏 Acknowledgments

• Models Used: GPT-2 (HuggingFace), GPT-OSS 20B (Ollama)
• Similarity: all-MiniLM-L6-v2 (Sentence-Transformers)
• Frameworks: PyTorch, Transformers, scikit-learn, SciPy