Hallucination and Grounding Detection
Small classifiers that check whether a generated claim is supported by its source documents, for retrieval-augmented generation and summarisation.
Tier Grades
Tier A: Frontier AI labs
feasible
Labs and vendors (Bespoke Labs, Vectara, IBM) release grounding checkers.
Tier B: Indonesian AI labs
feasible
MiniCheck trains on synthetic data from a stronger model, a recipe that does not need frontier compute.
Tier C: Institution units such as ITB
feasible
MiniCheck came from a 3-author team.
Tier D: Organisations such as SEMAI (PoT, PoC, PoV)
feasible
A grounding checker at 400 times lower cost than GPT-4 (MiniCheck) is a practical guard in RAG proofs of value.
Where this comes from
Tiers: a) frontier AI labs, b) Indonesian AI labs, c) institution units such as ITB, d) organisations such as SEMAI delivering proof of technology, concept and value. Each rationale cites the anchors below.
Anchors
- ComputeMiniCheck: small fact-checkers with GPT-4-level performance at 400 times lower cost; parameter count is not in the abstract.
- DataSynthetic training data generated with GPT-4 by building realistic factual errors; evaluation on the unified LLM-AggreFact benchmark.
- Team size (proxy)Proxy only (author count of an exemplar paper, not a headcount): 3 authors on MiniCheck.
- Time horizonMonths.
- Shelf-life signalsGrounding checks folded into guard models (IBM Granite Guardian RAG groundedness), LLM-AggreFact as the shared leaderboard
- Pipeline roleTier A and vendors release checkers and benchmarks; Tier C builds Indonesian grounding test sets; Tier D adds checkers to RAG proofs of value.
Where this comes from
Team size is a proxy (author counts of exemplar papers), not a headcount.
Who and What
Links
R&D TopicHallucination and Grounding Detection
Models
Bespoke-MiniCheck-7BBespoke Labs, guard
HHEM-2.1-OpenVectara, guard
Granite Guardian 4.1-8BIBM, guard
Bespoke-MiniCheck-7BBespoke Labs, guardHHEM-2.1-OpenVectara, guardGranite Guardian 4.1-8BIBM, guardUse Casesnone recorded
Who Runs Itnone recorded
Papers
- TruthfulQA: Measuring How Models Mimic Human FalsehoodsStephanie Lin et al., ACL 2022 (per arXiv comment), 2021 founding
- Why Language Models HallucinateAdam Tauman Kalai et al., arXiv preprint, 2025 state of the art
- SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection for Generative Large Language ModelsPotsawee Manakul et al., EMNLP 2023 (per arXiv comment), 2023 state of the art
- Survey of Hallucination in Natural Language GenerationZiwei Ji et al., ACM Computing Surveys (2022), 2022 survey
Where this comes from
The papers to read first: where the direction began, its current best, and a survey when one exists.
Sources
- arxiv.org/abs/2404.10774 https://arxiv.org/abs/2404.10774
To cite this page: Project AURION, AI Use-Cases Research & Innovation, "Hallucination and Grounding Detection", https://aurion.semai.id/topics/hallucination-and-grounding-detection/, records of October 2026.