A Two-Tiered Rescue Protocol to Mitigate Difficulty-Based Failures of ChatGPT 5 and Gemini on the German M2 Medical Exam: Evaluation Study (Preprint)
Notice bibliographique
Résumé
BACKGROUND: Large language models (LLMs) have demonstrated expert-level performance on medical licensing examinations, but most benchmarks focus on final accuracy, obscuring model-specific behaviors. Critical gaps remain in understanding model efficiency (latency), the efficacy of tiered "rescue" protocols for error correction, and the systematic correlation between performance and human-rated question difficulty. The German M2 exam, paired with the AMBOSS platform's user-data-driven difficulty ratings, provides a unique opportunity to map AI performance directly against human cognitive load. OBJECTIVE: This study aimed to move beyond singular accuracy scores by (1) evaluating and comparing the baseline (Tier 1) accuracy and response latency of next-generation rapid-response LLMs; (2) analyzing the efficacy of a two-tiered rescue (Tier 2) protocol in correcting initial errors; and (3) correlating model performance with the user-data-driven Amboss difficulty rating. METHODS: We evaluated four LLMs (Gemini 2.5 Flash/Pro and ChatGPT 5 Instant/Thinking) on the complete 316-item German M2 (Fall 2024) medical exam, including all multimodal (image-based) questions. A zero-shot copy-paste prompting strategy was utilized, and outputs were evaluated against ground-truth answers using a strict exact-match criterion. A two-tiered protocol was used: Tier 1 (Flash/Instant) provided baseline responses. If incorrect, a Tier 2 (Pro/Thinking) model was deployed as a "rescue." Performance was analyzed using McNemar's test, Wilcoxon signed-rank test, Fisher's exact test, and logistic regression. RESULTS: Baseline (Tier 1) accuracy was identical at 91.46% (95% CI 87.85-94.06; n = 289/316) for both Gemini 2.5 Flash and ChatGPT 5 Instant, with 27 errors each. However, Gemini Flash (Mean=1.57s) was significantly faster than ChatGPT Instant (Mean = 2.07s; P < .001). Additionally, ChatGPT Instant expended significantly more time on incorrect answers compared to correct ones (P = .002), whereas Gemini Flash showed no such hesitation (P = .814). The Tier 2 rescue rate for ChatGPT 5 Thinking (48.15%, 13/27; 95% CI 30.74-66.01) was higher, though not statistically significant (P = .406), than for Gemini 2.5 Pro (33.33%, 9/27; 95% CI 18.64-52.18). This rescue protocol elevated final accuracy to 94.30% (95% CI 91.18-96.37) for the Gemini system and 95.57% (95% CI 92.70-97.34) for the ChatGPT system (P = .481). A strong, inverse relationship with difficulty was found: for every one-point difficulty increase, the odds of a correct Tier 1 response decreased by 42.1% (OR 0.579, 95% CI 0.425-0.788; P < .001) for Gemini Flash and 47.7% (OR 0.523, 95% CI 0.379-0.720; P < .001) for ChatGPT Instant. This negative correlation persisted even after the rescue (P = .013 and P = .006, respectively). CONCLUSIONS: Expert-level LLM performance on the German M2 exam masks a critical, systematic vulnerability: a significant decrease in accuracy directly correlated with increased question difficulty. A two-tiered "rescue" system is an effective strategy to mitigate these difficulty-based failures and achieve >95% accuracy, rivaling the best-performing, full-capacity models. We conclude that a simple reliance on a single model is insufficient; hierarchical systems that manage query difficulty are essential for safe and effective integration into medical education.
Récupéré en direct depuis OpenAlex et désinversé. Les résumés ne sont pas conservés dans cette base de données : les index inversés représentent 8,6 Go des 9,3 Go de texte de la base, et le serveur dispose de 13 Go libres.
Comment cette classification a été obtenuedéplier
Prédiction machine sur la base complète
Imitation des enseignantsNi prévalence calibrée, ni vérité terrain. Validation humaine à venir. Le volet Gemma est une étiquette directe du modèle pour chaque travail de la base, lue sur la notice réduite au titre. Le volet Codex est un classifieur appris des 10 348 étiquettes directes de Codex et calibré sur les taux pondérés de l'échantillon; les champs sans appui suffisant ne portent aucun appel Codex. Le mode candidate est l'union des deux volets; le consensus est leur intersection. Ces sorties portent le statut machine_predicted_unvalidated et ne sont pas des étiquettes humaines.
Scores du classifieur distillé par catégorie (deux têtes)
| Catégorie | Codex | Gemma |
|---|---|---|
| Métarecherche | 0,007 | 0,038 |
| Méta-épidémiologie (sens strict) | 0,002 | 0,001 |
| Méta-épidémiologie (sens large) | 0,001 | 0,001 |
| Bibliométrie | 0,001 | 0,000 |
| Études des sciences et des technologies | 0,000 | 0,001 |
| Communication savante | 0,001 | 0,001 |
| Science ouverte | 0,002 | 0,002 |
| Intégrité de la recherche | 0,002 | 0,001 |
| Charge utile insuffisante (le modèle a refusé de juger) | 0,003 | 0,002 |
Scores machine (provisoires)
Les deux têtes enseignantes du modèle étudiant, lues sur ce travail. Un score ordonne la base pour la relecture; il n'affirme jamais une catégorie, et le statut de validation accompagne chaque rangée tel quel.
Scores de référence d'un modèle non mature (critères de maturité non atteints, 7 itérations). Un score ordonne; il n'affirme jamais une catégorie.
score_only:v0-immature-baseline · tel quel depuis la passe de notation : score_only signifie que le nombre peut ordonner les travaux, et qu'aucune étiquette de catégorie n'en découleClassification
machine, non validéePrédiction automatique; un appel candidat d’une seule source (Gemma direct ou Codex distillé), pas un consensus.
Le détail, modèle par modèle et score par score, se trouve en fin de page sous « Comment cette classification a été obtenue ».