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

Claim · All claims
Evidence · 1 study · causal r2 · q2
Currency · no dated study recorded · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Starting from the Base R1 candidate pool, TEI@4 raises Explore from 20.3% to 31.1% (+10.8pp) and lowers General by −11.7pp, while Verify is roughly preserved (16.0→16.9%). → schoenfeld-tei-1

Evidence¶

schoenfeld-tei-1¶

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 doi:10.48550/arxiv.2605.30666

q2 · i? · causal · r2

Schoenfeld phase analysis of thinking-trace paragraphs (Table 9, GPT-4o-mini paragraph classifier at temperature 0): "TEI@4raises Explore by +10.8pp (20.3%→ 31.1%)". Aligned GRPO instead drops Verify to 5.6% and raises General to 73.2%. No effect size printed.

"Starting from the Base R1 candidate pool,TEI@4raises Explore by +10.8pp (20.3%→ 31.1%) and lowers General by−11.7pp, reaching a deeper deliberation density without changing the model."

Discussion¶

Related Claims¶

• TEI@N inference-time selection raises the improvement rate on pre-incorrect math scenarios from 59.0% to 81.9% at N=8 over Greedy on a frozen DeepSeek-R1-8B base — related
• On the same frozen base, TEI@4 beats both DTR@4 and Greedy on Δ Solve Rate, leak rate, and composite R_total at no training cost — related
• DTR (Deep-Thinking Ratio) does not predict tutoring quality despite predicting math accuracy — related