The question
Can a stronger model use feedback about a smaller model’s mistakes to teach it more effectively than a static curriculum? The objective is a specialized student whose improvement persists in learned adapter weights, not just in a longer prompt.
What led here
The completed two-round adaptive study found 44.79% complete-answer accuracy for static teaching and 46.88% for feedback-guided teaching. Its +2.08-point mean difference was inconsistent across three lineages. The new follow-up keeps the route task while improving feedback granularity and measuring the starting parent on the same fresh final pool.
Read the completed adaptive apprenticeship report →
A controlled loop, not an uncontrolled swarm
- Measure the starting students. Three fixed parent adapters receive the same fresh final evaluation before additional training, as well as practice and monitoring measurements. These parents are already-trained checkpoints, not untouched models.
- Compare teaching policies. GPT-5.6 Sol prepares practice-only lessons for the Qwen2.5-7B-Instruct student. Adaptive teaching can use practice feedback to choose problems and explanations; static teaching does not condition on that feedback.
- Keep replay fixed. Both teaching policies use the same original replay source. Three parent lineages create six matched branches, each followed across two training rounds.
- Match exposure and measure learning. Each round receives 144 new and 48 replay problems, with 384 separate answer/explanation updates. Problem counts and updates are matched; actual tokens are recorded rather than assumed equal.
- Evaluate the completed loop. The planned final primary comparison is adaptive versus static complete-answer correctness. Cheapest-route accuracy is secondary, and retention is evaluated at the final stage. Held-out evaluation data never enter the teacher’s feedback.
What is ready, and what is not
An initialization-only check passed on 3 October: the exact saved parent adapter loaded and matched its tensors and tokenizer. No scientific training or answer generation has begun for this follow-up. Current-source adversarial review and separate launch approval remain outstanding. Preparation is not a learning result.
What this can establish
The study can test the combined value of feedback-guided problem selection and explanation writing under a fixed teaching pipeline. It does not isolate explanation quality alone. Both teachers receive common task guidance; only the adaptive teacher receives its own student’s practice responses and granular error diagnostics.
Three fixed parent seeds and one synthetic route family do not establish general intelligence gains, a population-wide effect, or recursive self-improvement. Complete-answer scoring in this study is not the prior report’s order-sensitive post-hoc diagnostic.
Research engineering
The implementation separates practice feedback from held-out evaluation, verifies generated answers, preserves parent and adapter provenance, records learning checkpoints, and retains failures. These controls make a negative or positive result interpretable without changing the experiment after seeing the outcome.