awesome-hallucination-detection

Papers and Summaries

Verifiable Rewards Beyond Math and Code: Lightweight Corpus-Grounded Process Supervision for Factual Question Answering

• Metrics: Exact Match (EM); training cost / wall-clock speedup (×) vs. neural-verifier baselines
• Datasets: Five knowledge-intensive / factual QA benchmarks (incl. TriviaQA; 30 model×benchmark cells over six 3B–14B instruction-tuned models)
• Comments: Proposes CorVer (Corpus Verify), a lightweight, plug-in process reward for RL fine-tuning (GRPO) of LLMs on factual QA that replaces neural verifiers (NLI, LLM-as-judge, retrieve-and-grade) with a corpus-grounded signal derived from Wikipedia co-occurrence counts. A small (0.5B) extractor turns each generated sentence into entity/claim spans whose corpus co-occurrence — queried via an Infini-gram index — yields sentence-level credit, which is mapped to token-level advantages and combined with response-level judge and format rewards in a GRPO update. Beats the raw baseline in every one of 30 cells (six instruction-tuned models, 3B–14B; five QA benchmarks), with an average +4.1 EM on TriviaQA, and outperforms four neural-verifier baselines in 18/20 cells at 4.8–8.4× lower training cost. Code at https://github.com/shichengf/CorVer. (arXiv 2026)

REFUTE: Scientific Critique & Epistemic Calibration Benchmark

• Metrics: Critique skill score, Brier calibration, forced-choice flaw accuracy, planted-flaw soundness accuracy, missing-evidence refusal rate
• Datasets: REFUTE (refute_hard_60, refute_120, refute_soundness; 120 critique + 74 soundness vignettes from recent science paper summaries)
• Comments: Open Apache-2.0 Hugging Face benchmark from BGPT measuring whether LLMs critique recent science summaries with calibrated, evidence-grounded judgment. Separates critique skill from uncertainty honesty: strongest critics are often most overconfident when evidence is weak. Includes judge-free Inspect AI and lm-evaluation-harness adapters plus a technical report. Complements factuality/hallucination detection benchmarks by testing overclaim correction and missing-evidence refusal on scientific summaries. (Technical report, Leaderboard) (2026)

QuCo-RAG: Quantifying Uncertainty from the Pre-training Corpus for Dynamic Retrieval-Augmented Generation

• Metrics: Exact Match (EM), token-level F1
• Datasets: HotpotQA, 2WikiMultihopQA, ASQA (long-form), PubMedQA (biomedical)
• Comments: Proposes QuCo-RAG, a dynamic RAG framework that determines when to retrieve by quantifying uncertainty from objective pre-training-corpus statistics rather than ill-calibrated model-internal signals (logits, entropy). Two-stage detection: (1) Pre-Generation Knowledge Assessment — queries entity frequencies in the pre-training corpus and triggers retrieval when entities are low-frequency (long-tail knowledge risk); (2) Runtime Claim Verification — extracts knowledge triplets from each generated sentence and verifies entity co-occurrence in the corpus, where zero co-occurrence triggers retrieval and regeneration. Both stages leverage Infini-gram for millisecond-latency queries over a 4-trillion-token corpus. On OLMo-2 (7B/13B/32B, matched corpus), achieves +5–12 EM over state-of-the-art dynamic-RAG baselines including DRAGIN. Cross-model transfer to Llama-3, Qwen2.5, GPT-4.1, and GPT-5-chat — which have undisclosed pre-training data — yields up to +14 EM by exploiting web-scale corpus overlap. Best performance on ASQA long-form QA and PubMedQA, while internal-signal methods show limitations in either efficiency or effectiveness. Code at https://g