Multimodal Large Language Models (MLLMs) frequently suffer from unfaithfulness, generating reasoning chains that drift from visual evidence or contradict final predictions. We propose Faithful-First Reasoning, Planning, and Acting (RPA) framework in which FaithEvi provides step-wise and chain-level supervision by evaluating the faithfulness of intermediate reasoning, and FaithAct uses these signals to plan and execute faithfulness-aware actions during inference. Experiments across multiple multimodal reasoning benchmarks show that faithful-first RPA improves perceptual faithfulness by up to 24% over prompt-based and tool-augmented reasoning frameworks, without degrading task accuracy. Our analysis shows that treating faithfulness as a guiding principle perceptually faithful reasoning trajectories and mitigates hallucination behavior. This work thereby establishes a unified framework for both evaluating and enforcing faithfulness in multimodal reasoning. Code is at https://github.com/lijunxian111/Faithful-First-RPA.

Educational LLMs such as writing coaches, coding assistants, study companions, and AI tutors are widely marketed as student-centered, usually because they are “personalized” or because they raise performance on the assisted task. We argue that this is an unverified pedagogical claim: a system that helps students complete tasks is not necessarily a system that helps students learn. The distinction is between performance-centered assistance, which optimizes the immediate product, and student-centered learning, which preserves the learner’s cognitive work and validates capability the learner keeps without the system. A 2025 randomized trial shows the gap in practice: an unguarded tutor raised assisted performance while harming unassisted exam scores. We trace this to a units mismatch: “student-centered” is a learner-side claim, yet it is measured almost entirely with tutor-side proxies. We propose five auditable standards (S1–S5) that translate established learning-science principles into requirements for benchmark design, reward shaping, red-teaming, and deployment monitoring; a small illustrative process audit then shows both that the standards can be operationalized and why such audits cannot substitute for learner-side validation. Our position is that LLM-based educational systems should be judged not by answer quality, engagement, satisfaction, or assisted performance alone, but by whether they measurably support learners’ capacity to reason, self-regulate, transfer, and keep learning without the model.