The paper "Towards Responsible Development of Generative AI for Education: An Evaluation-Driven Approach" addresses the challenge of providing equitable and universal access to quality education through the development of generative AI (gen AI) tools. The authors highlight the potential of gen AI to offer personalized tutoring and teaching assistance, but note that this potential has not yet been fully realized due to difficulties in verbalizing pedagogical intuitions and the lack of good evaluation practices. They present a participatory and multidisciplinary approach, involving learners, educators, policymakers, and researchers, to translate high-level principles from learning science into practical educational benchmarks. The paper introduces LearnLM-Tutor, a text-based gen AI tutor fine-tuned for 1:1 conversational tutoring, and demonstrates its superior performance over a prompt-tuned version of Gemini 1.0 in various pedagogical dimensions. The authors also develop a comprehensive suite of seven pedagogical benchmarks, including quantitative, qualitative, automatic, and human evaluations, to assess the performance of conversational AI tutors. The paper discusses the limitations, safety, ethical, and policy implications of their work, emphasizing the need for a shared framework across learning science, EdTech, and AI for Education to guide the responsible development of gen AI in education.The paper "Towards Responsible Development of Generative AI for Education: An Evaluation-Driven Approach" addresses the challenge of providing equitable and universal access to quality education through the development of generative AI (gen AI) tools. The authors highlight the potential of gen AI to offer personalized tutoring and teaching assistance, but note that this potential has not yet been fully realized due to difficulties in verbalizing pedagogical intuitions and the lack of good evaluation practices. They present a participatory and multidisciplinary approach, involving learners, educators, policymakers, and researchers, to translate high-level principles from learning science into practical educational benchmarks. The paper introduces LearnLM-Tutor, a text-based gen AI tutor fine-tuned for 1:1 conversational tutoring, and demonstrates its superior performance over a prompt-tuned version of Gemini 1.0 in various pedagogical dimensions. The authors also develop a comprehensive suite of seven pedagogical benchmarks, including quantitative, qualitative, automatic, and human evaluations, to assess the performance of conversational AI tutors. The paper discusses the limitations, safety, ethical, and policy implications of their work, emphasizing the need for a shared framework across learning science, EdTech, and AI for Education to guide the responsible development of gen AI in education.