Building behavioral auditing and alignment tools for LLMs. Current focus: rho-guided alignment -- using teacher-forced confidence probes to steer fine-tuning away from behavioral damage.

Core finding: Standard SFT inverts toxicity discrimination in Qwen 7B (+0.145 baseline to -0.003 post-SFT, 5 seeds, p<0.001). Adding a contrastive confidence loss during training repairs this (d=10.8 on toxicity, d=13.7 on bias, p<0.0001). Contrastive-only training erodes refusal (d=-8.4) but the full method preserves it. The margin gamma=0.1 is structurally necessary.

rho-eval (v2.1.1) -- Drop-in behavioral audit for any LLM. 926 probes across 6 dimensions (factual, toxicity, sycophancy, bias, reasoning, refusal), no internet required. Plugin architecture for custom behaviors. Apple Silicon MLX acceleration. Rho-guided SFT alignment.

pip install rho-eval
rho-eval Qwen/Qwen2.5-7B-Instruct --behaviors all --format table

# Works with PyTorch models -- or MLX models with zero code changes
import mlx_lm
from rho_eval import audit

model, tokenizer = mlx_lm.load("mlx-community/Qwen2.5-7B-Instruct-4bit")
report = audit(model=model, tokenizer=tokenizer, behaviors="all")
# ~5x faster inference, ~10x faster training on Apple Silicon

v2.1.1 -- Rho-Guided SFT (paper):

• SFT toxicity inversion -- Standard fine-tuning inverts toxicity discrimination (5 seeds, p<0.001); contrastive confidence loss repairs it (d=10.8 toxicity, d=13.7 bias, p<0.0001)
• Variance collapse -- Factual sigma drops 63% from SFT-only to rho-guided, making training more reliable across seeds
• Refusal buffer -- Contrastive-only training erodes refusal (d=-8.4, p=0.0005); the SFT component preserves it. Never train contrastive-only if refusal matters
• Margin necessity -- gamma=0 causes bias to go negative; gamma=0.1 prevents over-optimization past the natural separation boundary
• OOD transfer -- In-distribution contrastive training transfers to unseen clinical, social, and logic domains (+5pp accuracy)
• MLX-native training -- Full alignment pipeline runs on Apple Silicon. 7B model trains in ~10 min per condition on M3 Ultra

Earlier findings -- Comparative Anatomy of Behavioral Representations:

1. General-purpose behavioral diagnostic toolkit -- Done (rho-eval v2.0.0)
2. Mechanistic interpretability of behavioral subspaces -- Done (SVD subspace analysis, Grassmann angles, Universal Kill Zone discovery)
3. Rho-guided alignment / fine-tuning -- Done (v2.1.1 -- paper, SFT inversion discovery, contrastive repair, dose-response across Qwen + Llama)
4. Cross-architecture and scale validation -- in progress (5-seed ablation complete, refusal buffer discovered, safety stress test done, 70B planned)
5. Hybrid weight + activation control framework
6. Open behavioral benchmark suite (Fidelity-Bench 2.0)