🧠 Adaptive LLM Novelty Suppression Framework

Research prototype implementing the full MEA framework: information-theoretic decay estimation, composite novelty/bias/quality scoring, sector-dynamic thresholds, and TD-driven adaptive decoding control.

📝 Input Configuration

Prompt

Baseline LLM Output (Optional)

Evaluation Mode

Single model Compare two local models

Model Selection

Model Source

Model Family

Model

Custom Hugging Face model id (optional — overrides the selection above)

Runs any hub model served by HF Inference Providers using the HF_TOKEN secret. λ̂ decay is scored under the frozen reference proxy; perplexity is unavailable for API models.

Generations

1 5

Max New Tokens

40 350

Temperature (τ₀)

0.2 1.2

Top-p (p₀)

0.7 0.98

Top-k (k₀)

10 200

Compute Perplexity (slower)

Estimate information decay λ̂ (Eq i-v; needed for the decay gap)

Save Results to File

Batch Prompt Set — used only by Run Batch Evaluation

quick_test: 3 prompts · full_evaluation: 8 prompts · bias_assessment: 2 bias-risk prompts. Batch mode compares Models A vs B from the accordion above; results appear in the Batch Evaluation tab.

Results

Model Comparison

Run an experiment to generate the metrics report analysis.