Implementation for the research paper "Enhancing LLM Reasoning via Critique Models with Test-Time and Training-Time Supervision".

[ACL 2024] Code for the paper "ALaRM: Align Language Models via Hierarchical Rewards Modeling"

Source code for 'Understanding impacts of human feedback via influence functions'

Experiments for the Neural Interactive Proofs paper

LLM supervising other LLMs

PyINE is a research framework for scalable elicitation and oversight of LLM reasoning, built on instrumented Python programs as a verifiable execution substrate.

Adversarial Deliberation Trees with Mechanistic Verification for scalable LLM oversight

Intent Engineering: the engineering discipline for the human side of AI-led systems | 意图工程:AI 主导系统中面向「人」这一侧的工程学科

Evaluation infrastructure for AI systems beyond direct human supervision

GG Tank Watch - frozen public-information archive of a resolved May 2026 chemical emergency. Conduit-only design; responsible-AI safety patterns enforced in code and tests.

Can we detect a deceptive AI agent? A scalable-oversight study across 3 deception types and 3 detector conditions.

Nexus is a turn-based environment for training LLM agents to negotiate scarce compute resources (GPU, CPU, memory, bandwidth) under budget constraints and deadline pressure. Agents bid, trade, form coalitions, and strategize with hidden information. Designed for scalable AI oversight as well as multi-agent management research.

Reproducible Claude safety pilot testing whether a weak judge defers to corrupted expert-style critiques.

Promotion gate for AI agents: leakage audits, PASS/HOLD/BLOCK verdicts with exact statistics, and a LIVE public MCP endpoint at whetstone.cyberelf.link/mcp that grades agents on exams they provably never trained on. AGPL-3.0.

Executable audits for latent-state claims in AI alignment and high-stakes evaluation.

A 5-layer adversarial-audit engine (dependency-free Python core) plus the two papers that formalise it — managing epistemic circularity, the survivor gate, and three scientific ledgers. It never reports "validated" on internal grounds; only a human does.

Lean 4 theorem prover with an LLM Prover/Critic loop and SHA-256 theorem locking to catch specification gaming.

Reproducible LLM-as-judge grader for med-student clinical notes: rubric scoring + evidence-cited feedback, with a QWK/Pearson/MAE human-agreement harness (validated on ACI-Bench). Rutgers Health Hack 3rd/200+.

A checked prototype for separating machine validity from licensed inference

A multi-agent real estate valuation engine secured by scalable AI oversight and strict guardrails.