LearnLM: Pedagogical AI Tutor

• LearnLM is a pedagogically fine-tuned large language model that leverages structured system instructions to mimic expert human tutoring.
• It employs a multi-step fine-tuning pipeline including supervised fine-tuning, reward modeling, and RLHF to enhance tutoring quality and learner performance.
• Supported by rigorous evaluations and classroom RCTs, LearnLM demonstrates significant gains in pedagogical alignment, adaptive tutoring, and safety in real-world deployments.

LearnLM is a pedagogically fine-tuned LLM family based on Google’s Gemini architecture, designed to function as an interactive, adaptive tutor across academic domains. The LearnLM paradigm reframes tutoring as "pedagogical instruction following"—embedding explicit system-level instructions that specify pedagogical attributes for each model turn—enabling dynamic adaptation to learner needs without hard-coding any single pedagogic theory. Multiple instantiations of LearnLM have demonstrated substantial and statistically robust gains in tutoring quality, pedagogical alignment, and learner performance compared to contemporaneous flagship models, as validated through scenario-based, expert-driven evaluations and randomized controlled classroom trials (Team et al., 2024, Team et al., 29 Dec 2025, Jurenka et al., 2024).

1. Pedagogical Instruction Following: Conceptual Foundation

LearnLM introduces "pedagogical instruction following," a shift from traditional instruction tuning that exposes models to generic prompts and outputs. In this paradigm, every training and evaluation datum begins with a richly structured System Instruction describing desired pedagogical attributes: tutor persona, scaffolding strategies, pacing heuristics, reflection prompts, constraints on answer revelation, and more. Conditioning on these meta-instructions enables LearnLM to model pedagogical adaptivity reminiscent of expert human tutoring—multi-turn guidance, remediation of misconceptions, and motivational tailoring—without prescribing a canonical instructional approach. This modularity permits developers and educators to specify pedagogical behaviors suited to individual learners and contexts (Team et al., 2024).

2. Architectural Features and Fine-Tuning Pipeline

LearnLM retains the transformer architecture and parameterization of its Gemini base model; no architectural modifications or new attention mechanisms are introduced in reported versions. The model’s pedagogical capabilities arise from the post-training integration of instructional data using three principal methods:

• Supervised Fine-Tuning (SFT): Each training conversation is prepended with a system-level pedagogical instruction, and the dataset is a mixture: $$D_{\text{total}} = \lambda \cdot D_{\text{gem}} \cup (1-\lambda) \cdot D_{\text{ped}}$$, where $D_{\text{gem}}$ is Gemini's baseline data and $D_{\text{ped}}$ contains curated multi-turn pedagogical dialogues. The mixing ratio $\lambda$ is chosen to preserve general-purpose capabilities while ensuring pedagogical signal (e.g., $\lambda \approx 0.9$; exact values not disclosed) (Team et al., 2024).
• Reward Modeling (RM): Human experts rate pairs of model outputs against scenario-specific pedagogical instructions. A reward model $R_\phi(S, \text{context}, r)$ is trained to match these graded or binary human preferences via cross-entropy minimization:

$$L_\text{RM} = - \sum_{(r^+, r^-)} \log \sigma(R_\phi(r^+) - R_\phi(r^-))$$

(Team et al., 2024).

• Reinforcement Learning from Human Feedback (RLHF): Proximal Policy Optimization is employed to maximize expected reward over sampled continuations, subject to KL-penalty constraints:

$$L_\text{RL}(\theta) = -\mathbb{E}_{r \sim p_\theta} [ R_\phi(S, \text{context}, r) ]$$

(Team et al., 2024).

Reported extensions and variants—including LearnLM-Tutor and Gemini 2.0 Flash-based models—follow analogous SFT pipelines, with specific fine-tuning datasets drawn from curriculum-aligned problems, annotated misconceptions, synthetic tutor-learner roleplays, and safety-oriented demonstrations (Team et al., 29 Dec 2025, Jurenka et al., 2024).

3. Datasets and Data Preparation

LearnLM’s pedagogical fine-tuning utilizes multi-source datasets encompassing:

• Structured System Instructions: Each conversation begins with a system-level specification of learner persona, goals, and pedagogical constraints (e.g., "do not reveal answers directly," "use Socratic questioning") (Team et al., 2024).
• Human Tutoring Transcripts: Real 1:1 tutoring sessions curated for quality and topic relevance (Jurenka et al., 2024).
• Synthetic Role-plays: Gemini-powered agent dialogues simulating tutor-learner exchanges, manually filtered for pedagogic fidelity.
• GSM8k Conversions: Mathematical word-problems restructured into stepwise Socratic dialogues.
• Golden Scripts: Expert-written, rubric-based multi-turn tutorials grounded in authentic lesson materials.
• Safety Fine-Tuning Data: Off-policy queries paired with safe rejections or redirections.
• Curriculum-aligned Problems: For classroom deployments, datasets are engineered to include math exercises, common misconceptions, pedagogical corrections, and human-approved dialog exemplars (Team et al., 29 Dec 2025).

Preprocessing includes: deduplication, removal of off-topic content, context window truncation, and up-weighting high-quality scripts. The dataset size is on the order of tens of thousands of multi-turn examples; exact counts and detailed distributions are not publicly specified.

4. Evaluation Protocols, Benchmarks, and Results

The LearnLM family has been subjected to stringent evaluation protocols spanning quantitative, qualitative, automatic, and expert-driven measures:

• Scenario-based, multi-turn evaluations: 49 ecologically valid learning scenarios, role-played by expert tutors, yielded 2,360 paired conversations for side-by-side comparison and 31-item Likert-scale ratings.
• Automatic and Human Benchmarks: Includes standard academic tests (MMLU, MATH, HellaSwag, HumanEval), safety screens (RealToxicityPrompts, BBQ), turn-level factual verification, turn-level and conversation-level expert pedagogy ratings, and LME auto-eval critiques on active engagement, misconception identification, and topicality.
• Classroom RCTs: In a UK secondary-school trial ($N=165$ students), LearnLM supervised by expert tutors supported math learning on the Eedi platform. Immediate remediation rates were LearnLM 93.0%, human tutor 91.2%; misconception resolution rates were LearnLM 95.4%, human tutor 94.9%. Knowledge transfer to novel problems showed LearnLM outperforming human tutors by 5.5 percentage points (66.2% vs 60.7%), posterior probability $$P(\text{LearnLM transfer} > \text{human tutor}) = 93.6\%$$ (Team et al., 29 Dec 2025).

Model Preference Comparison

Model Comparison	Mean Preference Strength (%)	95% CI Excludes 0
LearnLM vs. GPT-4o	31	Yes
LearnLM vs. Claude 3.5	11	Yes
LearnLM vs. Gemini 1.5 Pro	13	Yes

All results demonstrate robust statistical support for LearnLM’s improvements in tutoring quality, adherence to pedagogical instructions, and fidelity to human tutor behavior (Team et al., 2024).

5. Qualitative Impact and Safety Assessment

Human expert and educator feedback consistently highlights LearnLM’s strengths:

• Socratic Dialogue: Tutors praised LearnLM's student-led questioning, noting the model often formulated probing questions they had not considered ("The questions were ones I hadn’t thought of").
• Professional Development: Three tutors reported adopting new Socratic practices learned from the model.
• Minimal Edits Needed: In classroom deployments, 76.4% of messages required zero or minimal edits (Levenshtein $k \leq 2$; median edit length 59 characters) (Team et al., 29 Dec 2025).
• Social-Emotional Adjustments: Most edits introduced personal rapport or moderated tone, with pacing frequently adjusted to prevent frustration.
• Safety: Post-trial audits found zero harmful messages and only 0.1% factual errors. Surveys indicated increased tutor comfort with AI support, while students rated interactive tutoring higher than static hints (3.9/5 vs. 3.6/5) (Team et al., 29 Dec 2025).
• Scalability: Tutors managed more concurrent sessions when assisted by LearnLM; cost per session was reduced.

6. Lessons, Limitations, and Future Directions

LearnLM exemplifies several best practices for responsible AI in education:

• Evaluation-driven development: Multidimensional pedagogical benchmarking, scenario-based human evaluation, and iterative LLM-based auto-critique protocols directly inform model refinement (Jurenka et al., 2024).
• Participatory Design: Co-design with educators, learners, and multidisciplinary stakeholders ensures real-world relevance and pedagogical robustness.
• End-to-End Safety: Safety is maintained via dataset curation, fine-tuning, policy UI guardrails, red-teaming, and ongoing monitoring.
• Generalization and Adaptivity: LearnLM retains Gemini’s general reasoning and multimodal capabilities, with injected pedagogical behaviors (active learning, metacognitive prompting, scaffolding, curiosity stimulation) (Team et al., 2024).
• Limitations: Reported limitations include the need for longitudinal studies to assess persistent learning gains, evaluation in non-STEM domains requiring interpretive or creative skills, and further refinement of automated systems for pacing and emotional nuance (Team et al., 29 Dec 2025). No negative regressions were evident in factual accuracy or academic benchmarks; slight trade-offs in encouraging tone have been observed due to shorter, focused responses (Jurenka et al., 2024).

A plausible implication is that pedagogically fine-tuned generative models such as LearnLM enable scalable, individualized support that matches or exceeds the median performance of expert human tutors in controlled settings, contingent on ongoing supervision and safety protocols.