Phase-aware LLM activation steering and linear probing. A memory-efficient, practical implementation of Representation Engineering (RepE) for safety research.

Layer-wise hidden-state probing for early detection of harmful intent in small instruction-tuned language models.

🏛️ Champollion cracked hieroglyphs in 1822. I applied the same logic to LLM internals. 95% accuracy, $0 cost, fully reproducible. Contributors welcome.

EECS E6895 final project measuring reward-gaming behavior in Gemma 2B with shell-game evals, LoRA SFT, and leakage-aware probes.

Open-source agent skills for Claude Code, Codex, GitHub Copilot, and other coding agents: GPT-5.6-style rigor, delegation, linear probes, and self-evolving workflows.

Fine-Tuning Silently Breaks Linear Safety Monitors

Research code for claim-level correctness probes on Llama activations.

Linear classifier probes (Alain & Bengio, 2016) — trains independent logistic regression classifiers on frozen GPT-2 representations to measure how sentiment information emerges across layers.

A research tool for studying how deception emerges in multi-agent LLM systems and detecting it through activation analysis.

Inline linear-probe safety monitoring on production LLM serving: overhead, quantization robustness, and prefix-cache blind spots, measured end to end on vLLM and SGLang.

Detection is not causal control: necessity tests for linear probe directions across 4 chat models, with a picker that alarms when no silencing handle exists.

Research: can LLM internal activations predict reasoning failure before it's visible? Linear probe on DeepSeek-R1 hidden states at 150 tokens achieves AUC 0.612 vs 0.445 baseline (p=0.001) on AIME math problems. Signal emerges when surface-level features carry zero information.

Construct-validity diagnostics for mechanistic interpretability probes: demonstrating why held-out probe accuracy cannot establish semantic representation under confounded stimulus designs.

Testing whether RLHF instills a privileged self-emotion direction in the residual stream, and whether surgical steering on emotion sub-directions can decouple sycophancy from harshness. Open-weight replication on Qwen2.5-32B.