Three platforms, one emerging idea. Pastest now embeds Tutor Mode, on-demand AI support, inside its question bank. Lecturio has introduced an explicitly Socratic AI Tutor within its Q-bank sessions, positioned as real-time guidance grounded in its medical library. And iatroX places its Socratic Tutor inside question sessions, wired to adaptive selection and study planning (all as marketed, last checked August 2026). Independent companies converging on the same product decision is usually the most reliable signal in a category: the question bank is becoming a place where teaching happens, not just testing. This article compares the three by pedagogical job rather than by feature count, because "which is best" is the wrong question and "what does each actually do when I'm stuck" is the right one.
The pedagogical jobs, compared
Explaining: all three do it; the differences are in grounding and context, Lecturio's tutor draws on its library, iatroX's corrections cite clinical guidance, Pastest's rides its explanation heritage. Prompting and hinting: Lecturio's positioning leans here, graded hints that keep the learner working before revealing; Pastest's on-demand model answers when summoned; iatroX withholds by default. Questioning, the Socratic core: this is where the three genuinely diverge, Pastest's Tutor Mode is assistance-shaped, help while you answer; Lecturio and iatroX both claim the Socratic word, with iatroX's implementation built around interrogating the wrong answer, what did you weight, what doesn't fit, until the misconception is named. Misconception identification: iatroX logs it as a first-class object feeding what happens next; the others surface it conversationally. Connecting onward: Lecturio links into its course library, Pastest into its learning content, iatroX into targeted practice. Selecting the next question: adaptive selection is iatroX's spine, performance-driven; Pastest remains analytics-led; Lecturio's bank connects to its curriculum pathways. Planning future study: iatroX's AI Study Planner owns this job explicitly; elsewhere it remains largely the learner's. Mobile continuity: all three ship apps; the test is whether a ten-minute phone session feeds the same learner model as the desktop mock, examined across the market at /blog/best-ai-study-apps-medical-students-2026.
Why the tutor belongs inside the Q-bank
The convergence has a simple logic: context. A tutor inside the question knows what you are attempting, what you answered, what the distractors were and, on the better implementations, what you have got wrong for six weeks, which is precisely the context a generic chatbot lacks and cannot be pasted into a prompt at any sustainable cost. The question bank turns out to be the ideal host for AI tutoring not because questions are sacred but because they generate the two things tutoring runs on: a committed attempt, and an error signal. From post-question explanation to in-question teaching is a small interface change and a large pedagogical one.
Does this improve reasoning or just exam scores?
The honest answer: it depends entirely on what the AI does at the moment of error. Assistance-shaped help improves throughput and comfort, real goods, with the known risk of comfort masquerading as competence. Interrogation-shaped tutoring targets the reasoning itself, and the mechanisms it deploys, forced production, feedback on the learner's own attempt, spaced retesting, are the best-evidenced in learning science, though direct trial evidence for AI tutors specifically remains early, as our standing review at /blog/do-ai-tutors-improve-medical-education-evidence keeps saying. The candidate-level implication: the same three platforms can produce very different learning depending on whether you summon answers or submit to questioning, which is within your control on all of them.
What candidates should now expect, and ask
The category's floor has moved. In 2026 a serious question bank should be expected to: teach at the moment of error, not just explain after it; know your history, weaknesses persisting across sessions; do something with the misconception, a follow-up, a retest, a plan change, rather than filing it under analytics; and travel, the phone session updating the same model as the desktop one. The demo question that separates the three philosophies in ninety seconds: get a question deliberately wrong and watch what the system does next. Assistance offers an explanation; a tutor asks you a question; a learning system schedules your future. The full generational map of where this is heading sits at /blog/how-ai-is-changing-the-medical-question-bank.
Frequently asked questions
Is Pastest's approach worse for being assistance-shaped?
Different, not worse: for past-paper mileage with friction removed, it is exactly right, and its bank remains formidable; the methodology trade-offs are examined properly at /blog/pastest-tutor-mode-vs-iatrox-socratic-tutor. The mistake is only in expecting assistance to do interrogation's job.
Where do AMBOSS and Geeky Medics fit in this convergence?
Same direction, different hosts: AMBOSS attaches its copilot to a knowledge library rather than primarily to questions, and Geeky Medics attaches tutoring to simulation; both are covered in their own deep-dives. The three platforms here share the specific bet that the question is the teaching moment.
Do I need more than one of these?
Usually no: pick by your exam's centre of gravity and commit, past-paper fidelity, library integration, or the adaptive Socratic loop. Running two banks is occasionally rational for MRCP; running two tutors never is.