Analyzing reasoning dynamics and self-reflection strategies in large language models.
This repository is part of an ongoing research study on Agentic AI, focusing on how language models can debate, self-evaluate, and improve their reasoning without explicit fine-tuning.
We explore Training-Free Group Relative Policy Optimization (GRPO) as one of the mechanisms enabling self-improvement through reflection, and compare it to multi-agent debate frameworks where multiple models collaboratively reason to reach consensus.
| Concept | Description |
|---|---|
| Multi-Agent Debate | Multiple agents (LLMs) argue for/against different reasoning paths to reach stable, explainable outcomes. |
| Training-Free GRPO | A single-agent variant where the model learns contextually via group-relative feedback, not gradient updates. |
| Self-Reflection Loops | The model summarizes its best reasoning and reuses it as "experience memory" for subsequent prompts. |
| Reward-Based Comparison | Candidates are scored relative to ground truth or heuristic feedback — simulating reinforcement signals. |
- Single-Agent GRPO Framework (training-free)
→ Generates multiple candidates, scores them, and learns via textual reflection. - Evaluation Metrics for Agent Reasoning
→ Tracks how reasoning consistency, factual accuracy, and label precision evolve. - Audit Traces for Interpretability
→ Stores reasoning steps, rewards, and extracted "experience rules" for transparency. - HPC-Ready SLURM Pipelines
→ Fully compatible with UMass Unity GPU nodes for large-scale model experiments.
Agentic AI represents the next leap beyond static LLMs — systems that:
- Argue, reflect, and revise their own reasoning.
- Adapt to feedback without retraining.
- Exhibit goal-directed behavior and emergent cooperation.
The GRPO layer acts as a control experiment — testing how far a single agent can self-improve without external debate.