Use inference-time selection with a small interpretable index as a pedagogical alignment lever between full RL retraining and greedy decoding¶

Strategy · All strategies
Evidence · 1 claim (1 for) · 1 study (1 causal), q2 · 0 of 1 report an effect size · 1 claim rests on one study

Description¶

For practitioners, the article recommends keeping the model frozen and selecting among N stochastic candidate responses with an interpretable index rather than retraining: "a useful pedagogical alignment lever sits between full RL retraining and unprincipled greedy decoding: inference-time selection with a tiny interpretable index retains the base model’s mathematical depth and exposes four interpretable weights for behavioural tuning". TEI@N generates N candidates, scores each with regex features plus a one-pass DTR forward hook, and returns the best-scoring response at cost linear in conversation length.

Design Implications¶

Context¶

Requirements¶

• A frozen reasoning LLM exposing explicit thinking traces; N stochastic candidate generations per turn

Constraints¶

• The selector helps most when the candidate pool is diverse and the base distribution is well-balanced; full RL alignment is described as a sharper but more dangerous instrument

Target Learners¶

• students tutored by LLM math tutors

Target Learning Goals¶

• improving learning outcomes of LLM tutoring without retraining the model

Affordances¶

• Tutoring Effectiveness Index Tei

Claims¶

• TEI@4 selection shifts thinking-trace mass toward deeper deliberation, raising Schoenfeld Explore share by +10.8pp without changing the model [+M]

Related Strategies¶

Examples¶

-

Key Sources¶

• Shim Jaechang and Unggi Lee. 2026. The Tutoring Effectiveness Index: Predicting LLM Math Tutor Quality from Four Conversation Signals. In Proceedings of The 35th ACM International Conference on Information and Knowledge Management (CIKM '26). https://arxiv.org/abs/2605.30666