Each SRJudge component contributes: the Reasoner and Judger add measurable F1 gains over the Selector alone¶

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? Adding the GRPO-trained Reasoner on top of the Selector improves F1 by 2.02%, 0.96%, and 1.92% on S Math, S Bio, and S Phy respectively, and adding the Judger further improves performance on all three datasets. → Zhiwei Yang 2026

Evidence¶

Zhiwei Yang 2026¶

Zhiwei Yang, Jiahua Yang, Huiru Lin, Xing Chen, and Quanlong Guan. (2026). SRJudge: Empowering Large Language Models with Selective Reasoning for Fine-Grained Knowledge Concept Tagging. arXiv preprint. https://arxiv.org/abs/2609.36982 doi:10.48550/arxiv.2609.36982

q2 · i? · design · r2

Ablation study (Table 4) comparing Selector-only, Reasoner-only, S+R, and S+R+J variants with two Judger sizes. The article reports the Reasoner gains above and states that "integrating the Judger consistently enhances performance on all three datasets", with the larger Qwen3-32B Judger (J2) yielding the best F1 scores.

"After incorporating the Reasoner on top of the Selector, the F1 scores are improved by 2.02%, 0.96%, and 1.92% on the S Math, S Bio, and S Phy datasets, respectively."

Discussion¶

Related Claims¶

• A dynamic position reward in RL training improves tagging performance and prevents overfitting to Selector outputs — related
• Applying GRPO alone to LLMs for tagging yields significantly lower performance than supervised fine-tuning — related
• SRJudge improves prediction on hard-to-distinguish samples, with larger gains on S Math than on the higher-grade datasets — related
• SRJudge consistently outperforms state-of-the-art baselines on three knowledge concept tagging datasets — related
• Pruning low-advantage completions during RL improves F1 up to a middle pruning rate while reducing training time — related
• Fine-tuned SLMs retrieve the correct concept within a top-K shortlist but have limited top-1 accuracy — related
• Ablations show removing the strategy selector substantially increases leakage and replacing the NLI verifier with same-model self-check reduces performance — related