A toolbox for benchmarking trustworthiness of multimodal large language models (MultiTrust, NeurIPS 2024 Track Datasets and Benchmarks)

Code for ACL 2024 paper "TruthX: Alleviating Hallucinations by Editing Large Language Models in Truthful Space"

Don't make LLMs honest. Make every factual claim auditable. — An LLM Claim Auditing Layer with T1-T7 truth gradients. 98.1% business effectiveness on LiarBench v0.2.

[ICML'2024] Can AI Assistants Know What They Don't Know?

A toolbox for benchmarking Multimodal LLM Agents trustworthiness across truthfulness, controllability, safety and privacy dimensions through 34 interactive tasks

Unofficial implementation of Microsoft’s Claimify Paper: extracts specific, verifiable, decontextualized claims from LLM Q&A to be used for Hallucination, Groundedness, Relevancy and Truthfulness detection

[ACL 2026] Illusions of Confidence? Diagnosing LLM Truthfulness via Neighborhood Consistency

A research project on competing notions of truth in large language models.

Practical anti-trendslop and truthfulness rules for AI tools

Google Chrome Extension to add LLM text reading in a browser

Improving LLM truthfulness via reporting confidence

LMRI (the Gaslighting Index): does an LLM correct a falsehood planted in its own previous answer — and keep the correction under four rounds of social pressure? 46 models, 180 items, ~23k judged rounds. Dataset, harness, transcripts, leaderboards.

Example VerityIndex Reports on Media

Measuring the computational cost of deception in LLMs through logprob analysis

Adaptive, probe-controlled activation steering that cuts LLM hallucination rate by ~10pp (64.6%→58.5%) on TruthfulQA — steers only when a real-time risk probe flags danger, unlike fixed-strength ITI/CAA/TSV baselines reproduced here for comparison.

Investigating honesty, deception and steering in large language models. Replicating and extending the MASK honesty benchmark on frontier models, working toward internal representation analysis and activation steering for honesty.