# Qwen2.5

On evidence-sufficiency judgment, a domain with no decidable correctness reward, a boundary-keeping recipe of supervised fine-tuning followed by GRPO on an LLM-judge reward raised spontaneous correct refusal under action pressure from 25% in the instruction-tuned baseline to 100% (40 per arm; Wilson 95% CI lower bound 91.2%, non-overlapping with the baseline interval [14.2%, 40.2%]), while a scrambled-pairing placebo fell to 5%. The same recipe's automatic false refusal on 281 well-formed claims was 1.1% (Wilson upper bound 3.09%), so the effect is boundary-specific rather than blanket refusal; SFT saturates the face, leaving the GRPO increment not separately identifiable, and ratings come from model sessions, not human experts.

On evidence-sufficiency judgment, a domain with no decidable correctness reward, a boundary-keeping recipe of supervised fine-tuning followed by GRPO on an LLM-judge reward raised spontaneous correct refusal under action pressure from 25% in the instruction-tuned baseline to 100% (40 per arm; Wilson 95% CI lower bound 91.2%, non-overlapping with the baseline interval [14.2%, 40.2%]), while a scrambled-pairing placebo fell to 5%. The same recipe's automatic false refusal on 281 well-formed claims was 1.1% (Wilson upper bound 3.09%), so the effect is boundary-specific rather than blanket refusal; SFT saturates the face, leaving the GRPO increment not separately identifiable, and ratings come from model sessions, not human experts.

On Movies and TV, Video Games, and Books from Amazon Reviews 2023, the study builds controlled hard candidate sets from SASRec's highly ranked negatives (candidate sizes 10 to 100) and compares Jev, a decision-oriented "System One Model" from TypeSafe AI, against SASRec, DCNv2, and pointwise and listwise Qwen2.5 7B Instruct rerankers, finding that Jev maintains strong recommendation effectiveness (NDCG@10, Hit Rate@10, MRR) while its observed serving latency grows substantially more gradually than pointwise Qwen, though it remains far above recommendation-specific models.

On Movies and TV, Video Games, and Books from Amazon Reviews 2023, the study builds controlled hard candidate sets from SASRec's highly ranked negatives (candidate sizes 10 to 100) and compares Jev, a decision-oriented "System One Model" from TypeSafe AI, against SASRec, DCNv2, and pointwise and listwise Qwen2.5 7B Instruct rerankers, finding that Jev maintains strong recommendation effectiveness (NDCG@10, Hit Rate@10, MRR) while its observed serving latency grows substantially more gradually than pointwise Qwen, though it remains far above recommendation-specific models.

The study systematically characterizes the scaling properties of same-family on-policy distillation (OPD) on Qwen2.5 (0.5B–14B) math reasoning, finding a regular early useful-transfer regime in which held-out accuracy rises approximately linearly with the square root of token-level reverse KL from the student initialization, and fitting power laws in student scale, teacher scale, and teacher gold score that predict peak accuracy and transfer rate, with every weak-to-strong student peaking above its own teacher.

The study systematically characterizes the scaling properties of same-family on-policy distillation (OPD) on Qwen2.5 (0.5B–14B) math reasoning, finding a regular early useful-transfer regime in which held-out accuracy rises approximately linearly with the square root of token-level reverse KL from the student initialization, and fitting power laws in student scale, teacher scale, and teacher gold score that predict peak accuracy and transfer rate, with every weak-to-strong student peaking above its own teacher.