The verification threshold trades hallucination reduction against over-refusal, with the default tau of 0.20 providing the best balance¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? A strict threshold of 0.10 minimizes hallucination but rejects too many acceptable responses; a relaxed threshold of 0.30 improves coverage but allows more unsupported claims; the default 0.20 provides the best balance. → S M Asif Hossain 2026

Evidence¶

S M Asif Hossain 2026¶

S M Asif Hossain, Ruksat Khan Shayoni, M. F. Mridha, and Jungpil Shin. (2026). EduGuard: A Safe RAG-Based LLM Tutor for Programming Education. arXiv preprint arXiv:2607.15738. https://arxiv.org/abs/2607.15738 doi:10.48550/arxiv.2607.15738

q2 · i? · design · r2

Verifier threshold sensitivity analysis on BILearn-CS reported in Table 6 across five threshold values. The article states the threshold "controls the trade-off between hallucination reduction and over-refusal"; Table 6 shows hallucination 4.9 with over-refusal 9.7 at tau 0.20.

"A strict threshold of 0.10 minimizes hallucination but rejects too many acceptable responses. A relaxed threshold of 0.30 improves coverage but allows more unsupported claims. The default 𝜏=0.20 provides the best balance."

Discussion¶

Related Claims¶

• On the external CS50-Forum set, EduGuard remains strongest with 86.7% correctness and 6.8% hallucination, though all systems lose grounding — related
• Classification threshold trades sensitivity against specificity: 0.30 yields 91.3% sensitivity, 0.60 yields 89.7% specificity — related
• AI hallucination can be turned into a pedagogical resource by making verification and collaborative fact evaluation integral to AI-supported learning — related
• Source-anchored instruction pairs reduce hallucination: they achieve mean Teacher Fidelity 1.4 points higher than non-anchored pairs, and models trained on anchored subsets produce 23% fewer factual hallucinations — related
• Diversity regularization λ = 0.30 was selected from a sensitivity analysis as balancing recommendation relevance and content diversity — related
• Under illustrative property thresholds, no prompting strategy passes all measured AIAA properties in every model-blueprint cell; few-shot anchoring passes only 2 of 9 cells — related
• Students describe concrete verification behaviors for GenAI explanations, treating evaluation as an iterative filter rather than a single accept-or-reject judgment — related