SFT and GRPO alignment gains reflect learned proficiency-conditioned generation rather than memorization or output collapse¶
Claim · All claims
Evidence · 1 study · designr2·q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly
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? For non-source-based prompts, neither SFT nor GRPO produces any near-duplicate of a training essay (0.00% with ROUGE-L>0.7), and 8-gram overlap is negligible and comparable to held-out real essays. → Heejin Do 2026
Evidence¶
Heejin Do 2026¶
Heejin Do, Jakub Kontak, Mrinmaya Sachan. (2026). SWIM: Student Writing Simulation via Proficiency-Conditioned Generation. arXiv preprint arXiv:2609.03215. https://arxiv.org/abs/2609.03215 doi:10.48550/arxiv.2609.03215
q2 · i? · design · r2
Memorization and diversity analysis of SFT and GRPO generations, checking ROUGE-L near-duplicates, 8-gram overlap, and within-profile repetition across 132 groups. The paper reports "neither method produces any near-duplicate of a training essay (0.00% with ROUGE-L>0.7 )".
"For non-source-based prompts, neither method produces any near-duplicate of a training essay (0.00% with ROUGE-L>0.7 ), and their 8-gram overlap is negligible"
Discussion¶
Related Claims¶
- Data scarcity and long-form generation remain bottlenecks: on Prompt 8 prompting outperforms SFT and GRPO — related
- GRPO with the Proficiency Alignment Reward improves alignment over SFT across every writing trait and essay prompt — related
- Zero-shot prompting collapses proficiency variation toward an idealized high-scoring student population, while SFT and GRPO match the human score and length distributions — related
- Essays within each tool group showed consistent homogeneity across named entities, n-grams, and topic ontology — related
- Models reproduce high-proficiency writing more faithfully than authentic low-proficiency linguistic form, with distinct failure modes by method — related
- GRPO-with-PAR gains over SFT transfer to independent evaluators never used during policy optimization — related
Page metadata
| Field | Value |
|---|---|
type |
claim |
id |
no-memorization-or-generation-collapse-in-trained-simulators |
title |
SFT and GRPO alignment gains reflect learned proficiency-conditioned generation rather than memorization or output collapse |
description |
SFT and GRPO alignment gains reflect learned proficiency-conditioned generation rather than memorization or output collapse |
status |
review |
evidence_strength |
moderate |
newest_evidence |
2026 |
recheck_by |
2029 |
generated |
by process:wiki-ingest on 2026-10-11 |
trust |
unverified — no one has confirmed this page's content |
sources |
1 — heejin-do-2026 |
Frontmatter is the page's machine-readable record — see the Schema & Guide.