FG-0516 Multi-Paradigm Learning Strategy
Description
Apply multiple learning paradigms to the same experience to maximize knowledge extraction. For each engagement: SFT extracts technique descriptions for supervised fine-tuning, RL captures decision sequences as reward-scored tuples for reinforcement learning, and unsupervised methods cluster findings for anomaly detection and pattern discovery. Three paradigms, one engagement, three data products. The same experience yields fundamentally different training signals depending on which lens you apply.
Implementation
Before each engagement, set up extraction for all three paradigms. SFT: define the pair format and extract technique descriptions from writeups. RL: instrument the agent loop to capture state-action-reward-state tuples with a defined reward function. Unsupervised: collect raw findings for post-engagement clustering and embedding analysis. Run all three extractions concurrently. After the engagement, process each paradigm's output through its own pipeline. Track which paradigm produces the most downstream improvement per engagement.
Success Indicators
- Each engagement produces structured output for at least two of three paradigms
- SFT pairs, RL tuples, and clustering outputs all feed distinct downstream pipelines
- The team can articulate which paradigm is producing the most value for their current needs
- Paradigm-specific metrics improve over time as each pipeline matures
Failure Modes
- Extraction overhead for three paradigms exceeding the engagement's value, making the cure worse than the disease
- One paradigm dominating while others are neglected, collapsing back to single-paradigm extraction
- RL reward function misaligned with actual goals, producing tuples that train the wrong behavior
- Unsupervised paradigm producing clusters without actionable insight, consuming resources for noise
Sub-Methods 3
Extract question-answer training pairs from engagement writeups and lab results. Each technique described becomes a supervised fine-tuning example. Format consistency matters: define the pair schema once, use it everywhere, validate before ingestion.
Instrument the agent execution loop to capture state-action-reward-next_state tuples. Define the reward function before the engagement: positive rewards for progress (service discovery, finding, access gained), negative rewards for waste (redundant scans, detection, scope violations). The tuples train decision-making in addition to knowledge.
After an engagement, cluster raw findings using embeddings, PCA, or other dimensionality reduction. Look for anomalies that do not fit known patterns, clusters that suggest new technique categories, and outliers that warrant deeper investigation. The unsupervised paradigm finds what you did not know to look for.