Researchers trained a small 86M-parameter language model on Indonesian arithmetic using pedagogically-grounded Chain-of-Thought supervision based on the GASING method, achieving over 80% accuracy on held-out problems. The model developed both procedural reasoning and mental-arithmetic capabilities without reinforcement learning, demonstrating that human teaching methods can guide efficient AI training for mathematical reasoning.
This research addresses a fundamental challenge in AI development: teaching language models reliable arithmetic reasoning through methods inspired by human pedagogy rather than brute-force scaling. The team operationalized the GASING method, an Indonesian teaching approach that solves arithmetic left-to-right, into natural-language supervision signals that align with how transformer models generate tokens causally. By training only on next-token prediction without reinforcement learning, they achieved competitive performance with much larger models.
The pedagogical approach to AI training represents a meaningful shift in how researchers conceptualize model development. Rather than increasing parameters or data volume, the work shows that structuring training data according to human learning principles yields more efficient models. The mechanistic analysis revealing three distinct learning phases—where models first learn procedures, then develop associative mental arithmetic—provides insight into how reasoning capabilities emerge from language modeling objectives.
For the broader AI industry, this suggests that domain-specific, methodologically grounded training can produce capable models at lower computational cost. This has implications for resource-constrained settings and developing regions where computational budgets are limited. The use of a syllabic-agglutinative TOBA tokenizer for Indonesian also demonstrates the importance of linguistic considerations in model design.
Future work should explore whether these pedagogical principles generalize across languages, mathematical domains, and reasoning tasks. Testing on more complex arithmetic and abstract reasoning benchmarks would establish whether the efficiency gains hold beyond basic computation. The approach may influence how AI systems are trained for specialized domains where human pedagogical methods have proven effective over centuries.
- →Small 86M-parameter models trained with pedagogically-grounded supervision reach 80%+ accuracy, matching much larger models on arithmetic tasks.
- →The GASING method, an Indonesian teaching pedagogy, successfully guided language model training without reinforcement learning or reward optimization.
- →Mechanistic analysis shows models develop both explicit procedural reasoning and implicit mental-arithmetic capabilities through standard next-token prediction.
- →Training data structure based on human learning principles can be more efficient than scaling parameters, with implications for resource-constrained AI development.
- →The approach demonstrates that domain expertise in pedagogy deserves consideration in AI training methodology alongside computational scaling.