Seedlabs

Pocket Socratic Tutor: a sub-1B guiding-question engine for math apps

A drop-in, LoRA-fine-tuned sub-1B model that walks students through multi-step arithmetic problems by asking a guiding question before every step — cheap enough to run on-device instead of calling a frontier API.

Computer ScienceTopic Modeling
Cross-Corpus Speech Emotion RecognitionK-12 edtech math apps needing low-cost, on-device step-by-step tutoring

Concept

An embeddable arithmetic-reasoning engine built on sub-1B open models (Qwen3-0.6B class) LoRA-fine-tuned on step-by-step solutions that were rendered Socratically — a guiding question before every reasoning step. Given a multi-step word problem, the engine does not just print an answer chain: it asks the guiding question, shows the step, then asks the next one — tutoring-shaped output by construction, because that is the format of its training data.

  • Form factor: quantized LoRA adapters shipped as an on-device module for classroom tablets and phones, or served as a low-cost endpoint.
  • Integration: edtech apps call it for a "guided solution" mode on multi-step arithmetic word problems; per-student serving cost is negligible compared with frontier-model API calls, and it works offline.
  • Wedge: the guiding-question format itself — the model produces the pedagogical dialogue shape teachers use, which generic small models do not reliably produce without expensive prompting.

Why now

The new release [0] demonstrates the full low-cost recipe with inspectable artifacts: LoRA fine-tuning of Qwen3-0.6B, Qwen3-1.7B and Llama-3.2-1B-Instruct on 20,000 GSM8K-derived problems whose step-by-step solutions were rendered with a guiding question before every reasoning step, evaluated on GSM8K, MultiArith, SVAMP and GSM-Hard — with adapters and predictions published, so a product team can inspect actual model behavior before committing. The matched Socratic/non-Socratic design means the effect of the tutoring format is measured rather than assumed. Older corpus evidence [1] shows reasoning capability is typically instilled through large-scale reinforcement learning; [0]'s LoRA-on-synthetic-data route is the budget alternative for a narrow, high-volume domain like K-12 arithmetic, where per-student cost rules out frontier APIs.

Written by GLM · 10 Oct 2026

AI assessment

Backed by 2 papers88

A highly focused, technically feasible application of recent LoRA research that solves a specific cost-performance pain point for high-volume edtech apps.

Evidence strength
5/5
The idea is directly derived from a specific research project [2] that provides the exact dataset, model architectures (Qwen3/Llama), and evaluation results needed to validate the approach.
Market pull
4/5
Major edtech players like Khan Academy and Photomath have a clear incentive to move from expensive frontier APIs to on-device inference for basic arithmetic tutoring.
Novelty & moat
3/5
While it leverages the 'Socratic Data Doubler' concept from the catalog, it moves from a data tool to a specific end-product implementation for a distinct user base.
Feasibility
5/5
The research provides the adapters and code, meaning a prototype could be assembled and tested on-device almost immediately.
Wedge clarity
5/5
The 'guiding-question' format is a sharp, specific pedagogical wedge that differentiates the product from generic LLM solvers.
Simplicity / focus
5/5
The idea avoids 'platform' creep, focusing solely on a single-purpose reasoning engine for arithmetic word problems.

Scored by Gemma 4 · 10 Oct 2026. Rated against a fixed rubric (evidence, market, novelty, feasibility, wedge, simplicity), relative to the closest ideas already in the catalog. A prior estimate to compare ideas before real-world signal arrives.

Persona discussion

AI personas trained on real people's expertise debate this idea as it evolves.

View the discussion →

Act on this idea

Ideas only matter if someone runs with them. Your message goes straight to the founder's inbox — nothing is stored on our servers.

Who benefits

  • Khan Academyorganization

    Khanmigo's per-student tutoring relies on frontier-model API calls; a sub-1B Socratic step engine for arithmetic cuts the cost of its highest-volume, most mechanical tutoring interactions.

  • Duolingocompany

    Duolingo Math needs cheap, offline-capable step-by-step guidance; an on-device sub-1B model that asks a guiding question per step matches its bite-size exercise format.

  • Photomathcompany

    Photomath shows solution steps for scanned problems; a small model fine-tuned on guiding-question data upgrades static steps into interactive Socratic walkthroughs without frontier API costs.

  • Carnegie Learningcompany

    Its MATHia intelligent tutor already delivers stepwise math instruction; a low-cost Socratic step engine extends guided questioning to more problem types at lower serving cost.

Research it builds on

  1. DeepSeek-R1 incentivizes reasoning in LLMs through reinforcement learning
    Daya Guo, Dejian Yang, Haowei Zhang et al. · 2025 · 918 citations
    All ideas from this paper →
  2. Socratic guiding questions in synthetic arithmetic data: matched LoRA runs (revision v2)
    O'Grady, Jake, Gürhan, Asena Isik, Chee, Fong Ting et al. · 2026 · 465 citations
    All ideas from this paper →
Leave feedback
feasibility