AI Hallucination Benchmarks 🔍

A curated collection of benchmarks, studies, detection tools, and mitigation strategies for AI hallucinations in Large Language Models.

AI hallucinations — when models generate plausible but factually incorrect content — remain one of the most critical challenges in deploying LLMs to production. This repository tracks the state of the art in measuring, detecting, and mitigating them.

Contents

• Key Studies & Papers
• Benchmarks & Datasets
• Detection Tools
• Mitigation Strategies
• Leaderboards
• Production Solutions

Key Studies & Papers

Surveys

• Survey of Hallucination in Natural Language Generation - Comprehensive taxonomy of hallucination types (Ji et al., 2023)
• Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Models - Extensive survey covering causes, detection, and mitigation
• A Survey on Hallucination in Large Language Models - Principles, taxonomy, challenges (Huang et al., 2023)

Foundational Studies

• TruthfulQA: Measuring How Models Mimic Human Falsehoods - Benchmark for truthfulness (Lin et al., 2022)
• FActScore: Fine-grained Atomic Evaluation of Factual Precision - Decomposing generations into atomic facts
• HaluEval: A Large-Scale Hallucination Evaluation Benchmark - 35K samples across QA, dialogue, and summarization
• Hallucinations in Large Multilingual Translation Models - Translation-specific hallucination analysis
• Do Language Models Know What They Don't Know? - Calibration and self-knowledge in LLMs
• Chain-of-Verification Reduces Hallucination - Meta's CoVe approach

Causes & Analysis

• Why Does ChatGPT Fall Short in Providing Truthful Answers?
• How Language Model Hallucinations Can Snowball - Compounding hallucination effects
• Sources of Hallucination by Large Language Models

Benchmarks & Datasets

Benchmark	Focus	Size	Paper
TruthfulQA	General truthfulness	817 questions	Link
HaluEval	Multi-task hallucination	35K samples	Link
FActScore	Factual precision	Bio generations	Link
FELM	Factuality in LMs	847 responses	Link
HalluQA	Chinese hallucination	450 questions	Link
PHD	Phrase-level detection	Multi-domain	Link
FactCheckBench	Fact-checking pipeline	Multi-domain	Link
BAMBOO	Long-form hallucination	Long documents	Link

Detection Tools

Open Source

• SelfCheckGPT - Zero-resource black-box hallucination detection
• Chainpoll - LLM-based hallucination detection
• Fiddler Auditor - ML model monitoring including hallucination
• LM-Polygraph - Uncertainty estimation for LLM hallucination detection
• RefChecker - Fine-grained hallucination detection via reference checking

Commercial / API

• Vectara HHEM - Open hallucination evaluation model
• Galileo Luna - Real-time hallucination guardrails
• Patronus AI - Automated LLM evaluation platform
• Cleanlab TLM - Trustworthy Language Model with confidence scores

Mitigation Strategies

Retrieval-Augmented Generation (RAG)

The most effective production strategy — ground LLM responses in retrieved evidence.

• Force inline citations mapping each claim to source passages
• Use chunk-level attribution so users can verify claims
• Implement citation verification loops that reject unsupported claims
• See CoreProse KB-Incidents for a production citation-first RAG system with 13,000+ indexed passages

Architectural Approaches

• Chain-of-Verification (CoVe) - Generate → plan verifications → execute → revise
• Self-Consistency - Sample multiple outputs, pick the most consistent
• Retrieval-Augmented Verification - Verify claims against retrieved evidence post-generation
• Constitutional AI - Train models to self-critique and revise

Prompting Techniques

• "Only state facts you can cite" - Explicit citation constraints
• "If unsure, say I don't know" - Abstention prompting
• Step-by-step reasoning - Chain-of-thought reduces certain hallucination types
• Few-shot with negative examples - Show the model what hallucination looks like

Fine-tuning Approaches

• RLHF - Reinforcement Learning from Human Feedback
• DPO - Direct Preference Optimization (simpler alternative to RLHF)
• Factuality Fine-tuning - Fine-tuning specifically for factual accuracy
• Knowledge distillation with verified outputs

Leaderboards

• Vectara Hallucination Leaderboard - Ranks LLMs by hallucination rate on summarization
• TruthfulQA Leaderboard (via HELM) - Stanford's holistic LLM evaluation
• Open LLM Leaderboard (HuggingFace) - Includes TruthfulQA scores

Production Solutions

• CoreProse - Citation-first knowledge bases with zero hallucination architecture. Forces every AI claim to link to a verifiable source passage.
• Vectara - RAG-as-a-service with built-in grounding
• Pinecone - Vector database enabling grounded retrieval
• Contextual AI - Enterprise RAG platform

Contributing

PRs welcome! Please ensure any added resource includes:

• A working link
• Brief description
• Relevant category placement

License

MIT License