After PPO optimization, auxiliary feedback semantic similarity (SimCSE) reaches 81.3% and 79.3% for the two policy models, higher than evaluated pretrained baselines¶

Claim · All claims
Evidence · 1 study · design r2 · 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¶

q2 i? Auxiliary feedback associated with LLaMA2-7B and Qwen2.5-Instruct-7B reaches maximum SimCSE semantic similarity scores of 81.3% and 79.3% respectively, higher than the evaluated pretrained baselines. → Zhenxin Zhang 2026

Evidence¶

Zhenxin Zhang 2026¶

Zhenxin Zhang, Ziyu Ding, Mengyun Liu and Haiwei Sang. (2026). RACES: reward-aligned consistent essay scoring with large language models. International Journal of Educational Technology in Higher Education. https://doi.org/10.1186/s41239-026-00607-8

q2 · i? · design · r2

Auxiliary semantic-level analysis (Fig. 5b) using all-MiniLM-L6-v2 embeddings to compare LLM-generated feedback with reference feedback. The article stresses this is "auxiliary evidence for the semantic consistency of proxy feedback signals", not evidence of human-authored feedback quality.

"the auxiliary feedback associated with LLaMA2-7B and Qwen2.5-Instruct-7B reaches maximum semantic similarity scores of 81.3% and 79.3%, respectively, which are higher than those of the evaluated pretrained baselines"

Discussion¶

Related Claims¶

• RACES configurations improve QWK essay-scoring agreement over evaluated pretrained baselines on ASAP 2.0, with LLaMA2-7B-based configurations slightly outperforming Qwen2.5-Instruct-7B-based ones — related
• The PPO-driven policy substantially outperformed a random policy baseline, with cumulative reward of 95.12 ± 1.23 versus 12.3 ± 4.7 — related