LLM Auditing Framework

AuditLM

Trust, Verify, and Attribute Every AI Generation

A set of auditing techniques to trace LLM reasoning steps and provide trust and transparency on AI-generated content.

Choose Your Lens

Due to resource constraints, we are currently unable to make these tools publicly accessible. To request access, contact us.

ProveRAG

ProveRAG is a provenance verification tool that decomposes responses into claims, checks against retrieved evidence, and scores citation reliance.

LLM-as-a-Judge Provenance
Launch ProveRAG

LEA

LLM Embedding-based Attribution (LEA) traces every generated token back to its sources with Linear Independence to verify source attribution.

Token-level Embeddings
Launch LEA

TAD

Topological Attribution Distance (TAD) is a segment-level technique that reveals RAG influence on LLM output geometry to trace AI reasonings.

Segment-level Output Geometry
Launch TAD

The Science Behind AuditLM

ProveRAG

The paper introducing claim decomposition and evidence-grounded provenance scoring, using an LLM-as-a-judge to verify how faithfully each response relies on its retrieved citations.

Read the paper

LEA

The LLM Embedding-based Attribution (LEA) paper, detailing how linear independence between token and source embeddings traces every generated token back to the evidence that shaped it.

Read the paper

TAD

The Topolocial Attribution Distance (TAD) paper is a segment-level trajectory attribution, revealing how retrieved context bends the geometry of an LLM's output to trace its reasoning.

Read the paper

How It Works

About Us

Reza Fayyazi

Lead Researcher

Reza Fayyazi

Reza Fayyazi is a PhD Candidate at Rochester Institute of Technology (RIT). He is working at the intersection of Large Language Models (LLMs), Mathematics, and Trustworthy AI. His work focuses on making the reasoning of LLMs auditable by developing explainable techniques to verify provenance, trace reasoning processes, and attribute AI-generated responses back to their sources. AuditLM brings these research techniques together into a single framework so that anyone can trust, verify, and attribute AI generations.

View full profile