GRPO pedagogical alignment of a reasoning LLM lowers actual student learning while raising judge-rated helpfulness (a multi-dimensional alignment tax)¶
Claim · All claims
Evidence · 1 study · causalr1·q2
Currency · no dated study recorded · not yet checked against its sources · not reviewed by a person
Subclaims¶
Reading the evidence codes
q is the design tier, i the impact, n the sample, and the kind with its r
the rigour judged within that kind: what each code means.
This claim: no standardized effect is recorded in observations/ for it yet, so no benchmark reading is possible.
q2 i? GRPO alignment of DeepSeek-R1-8B drops Δ Solve Rate from +.180 to −.012, cuts thinking length by −93.3%, and reduces OpenLearnLM CK/PK accuracy by −71%/−80% relative, while judge-rated helpfulness rises from .220 to .416. → alignment-tax-1
Evidence¶
alignment-tax-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 · r1
Comparison of base vs. GRPO-aligned DeepSeek-R1-8B across in-domain tutoring scenarios and the OpenLearnLM OOD suite (Table 7): thinking length 1,764→119 words/turn, visible length 194→21, ΔSol −.192, helpful_rate +.196. No standardised effect size printed.
"Pedagogical alignment cuts thinking length by an order of magnitude and visible length by nearly the same amount; the visible mathematical content density drops from25.5%to9 .5%; the student’s post-tutoring solve rate actually falls below the pre-tutoring rate (Δ=−. 012)."
Discussion¶
Related Claims¶
- GRPO alignment collapses Mathematical-Knowledge-for-Teaching utterances and surges Socratic-style Exploratory questions in tutor output — related
- DTR (Deep-Thinking Ratio) does not predict tutoring quality despite predicting math accuracy — related
- In held-out pairwise judge evaluation, combining Edu-QuRater and answer-structure GRPO rewards yields the highest pedagogical-quality win rate (81.08%) while retaining instruction-following above parity (68.24%) against the Qwen3-4B base model — reports the opposite
- Applying GRPO alone to LLMs for tagging yields significantly lower performance than supervised fine-tuning — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
grpo-pedagogical-alignment-tax-lowers-learning |
title |
GRPO pedagogical alignment of a reasoning LLM lowers actual student learning while raising judge-rated helpfulness (a multi-dimensional alignment tax) |
description |
GRPO pedagogical alignment of a reasoning LLM lowers actual student learning while raising judge-rated helpfulness (a multi-dimensional alignment tax) |
status |
review |
evidence_strength |
moderate |
generated |
by process:wiki-ingest on 2026-10-11 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — alignment-tax-1 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.