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 -0.086 post-SFT, p<0.001). Adding a contrastive confidence loss during training repairs this while preserving factual gains. The contrastive signal alone (no SFT data) achieves the largest toxicity improvement of any condition tested (d=49.4 vs SFT-only).

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 discovered and fixed — Standard fine-tuning inverts toxicity discrimination; adding a contrastive confidence loss during training repairs it (Qwen 7B: -0.086 SFT-only to +0.804 rho-guided)
• Monotonic dose-response — Behavioral preservation scales linearly with contrastive weight across both Qwen 7B and Llama 3.1-8B
• Contrastive-only condition — Contrastive loss alone (no SFT data) achieves the largest effect (d=49.4), suggesting behavioral probes carry enough signal to steer training without task data
• Refusal robustness — New behavioral dimension; independent from toxicity (refusal unchanged at +0.70 even when toxicity inverts)
• 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 running, refusal robustness, 70B planned)
5. Hybrid weight + activation control framework
6. Open behavioral benchmark suite (Fidelity-Bench 2.0)