List your company
ResearchFormerly WizWand

Code Generation on HumanEval v1 (test)

86.6AccuracyPrefillShare
34.648.161.675.1Feb 12, 2026Mar 1, 2026Mar 18, 2026Apr 4, 2026Apr 21, 2026May 8, 2026May 26, 2026
Updated 5mo ago

Evaluation results

MethodLinks
PrefillShareBackbone=Qwen3-8B-Base, KV Sharing=Supported2026.0286.6—
Full-FTBackbone=Qwen3-8B-Base, KV Sharing=Not Supported2026.0283.5—
TIPSetting=Qwen3-8B (GRPO) -> Qwen3-4B, Supervised-token budget=10%2026.0580.49—
TA-OPDSetting=Qwen3-8B (GRPO) -> Qwen3-4B, Supervised-token budget=10%2026.0579.88—
TA-OPDSetting=Qwen3-14B -> Qwen3-4B, Supervised-token budget=10%2026.0579.57—
EntropySetting=Qwen3-14B -> Qwen3-4B, Supervised-token budget=10%2026.0579.27—
Pure OPDSetting=Qwen3-8B (GRPO) -> Qwen3-4B, Supervised-token budget=10%2026.0578.36—
TA-OPD+Ent.Setting=Qwen3-14B -> Qwen3-4B, Supervised-token budget=10%2026.0578.35—
TA-OPD+Ent.Setting=Qwen3-8B (GRPO) -> Qwen3-4B, Supervised-token budget=10%2026.0578.05—
Pure OPDSetting=Qwen3-14B -> Qwen3-4B, Supervised-token budget=10%2026.0578.05—
TIPSetting=Qwen3-14B -> Qwen3-4B, Supervised-token budget=10%2026.0578.05—
BaseSetting=Qwen3-8B (GRPO) -> Qwen3-4B, Supervised-token budget=10%2026.0577.75—
BaseSetting=Qwen3-14B -> Qwen3-4B, Supervised-token budget=10%2026.0577.75—
EntropySetting=Qwen3-8B (GRPO) -> Qwen3-4B, Supervised-token budget=10%2026.0577.65—
Qwen3-8B-BaseKV Sharing=Inherent2026.0268.3—
TA-OPDSetting=DeepSeek-R1-Distill-Qwen-14B -> Qwen2.5-3B, Supervised-token budget=10%2026.0566.77—
BaseSetting=DeepSeek-R1-Distill-Qwen-14B -> Qwen2.5-3B, Supervised-token budget=10%2026.0564.63—
TA-OPD+Ent.Setting=DeepSeek-R1-Distill-Qwen-14B -> Qwen2.5-3B, Supervised-token budget=10%2026.0562.81—
TIPSetting=Qwen3-4B -> Qwen3-1.7B, Supervised-token budget=10%2026.0562.5—
Pure OPDSetting=DeepSeek-R1-Distill-Qwen-14B -> Qwen2.5-3B, Supervised-token budget=10%2026.0562.2—
TA-OPD+Ent.Setting=Qwen3-4B -> Qwen3-1.7B, Supervised-token budget=10%2026.0561.59—
SFT-GTSFT Training Domain=Code SFT (Ling-Coder-SFT)2026.0561.55—
TA-OPDSetting=Qwen3-4B -> Qwen3-1.7B, Supervised-token budget=10%2026.0560.98—
BaseSetting=Qwen3-4B -> Qwen3-1.7B, Supervised-token budget=10%2026.0560.97—
TABOMSFT Training Domain=Code SFT (Ling-Coder-SFT)2026.0560.36—
EntropySetting=Qwen3-4B -> Qwen3-1.7B, Supervised-token budget=10%2026.0559.45—
Pure OPDSetting=Qwen3-4B -> Qwen3-1.7B, Supervised-token budget=10%2026.0559.15—
TABOMSFT Training Domain=Math SFT (MixChain-Z-PRM12K)2026.0558.54—
TIPSetting=DeepSeek-R1-Distill-Qwen-14B -> Qwen2.5-3B, Supervised-token budget=10%2026.0558.54—
SFT-SDSFT Training Domain=Math SFT (MixChain-Z-PRM12K)2026.0557.92—
EntropySetting=DeepSeek-R1-Distill-Qwen-14B -> Qwen2.5-3B, Supervised-token budget=10%2026.0554.88—
SFT-SDSFT Training Domain=Code SFT (Ling-Coder-SFT)2026.0553.66—
No-SFTSFT Training Domain=None (Base Model)2026.0552.66—
PrefillShareBackbone=LLaMA3.1-8B, KV Sharing=Supported2026.0248.8—
Full-FTBackbone=LLaMA3.1-8B, KV Sharing=Not Supported2026.0248.2—
SFT-GTSFT Training Domain=Math SFT (MixChain-Z-PRM12K)2026.0546.34—
LLaMA3.1-8BKV Sharing=Inherent2026.0236.6—
Forward KLFamily=KL Divergence Family2026.05—38.41
Forward KL + OursFamily=KL Divergence Family, Strategy=Bridge-Garden2026.05—41.46
Hard KDTeacher=DeepSeek-Coder-6.7B, Student=DeepSeek-Coder-1.3B, Category=Reference & Baselines2026.05—35.37
Reverse KLFamily=KL Divergence Family2026.05—39.02
Reverse KL + OursFamily=KL Divergence Family, Strategy=Bridge-Garden2026.05—37.8
Skew FKLFamily=Distance & Skewed Family2026.05—42.07
Skew FKL + OursFamily=Distance & Skewed Family, Strategy=Bridge-Garden2026.05—42.41
Skew RKLFamily=Distance & Skewed Family2026.05—42.07
Skew RKL + OursFamily=Distance & Skewed Family, Strategy=Bridge-Garden2026.05—41.24
Student (No Distill)Model=DeepSeek-Coder-1.3B, Category=Reference & Baselines2026.05—33.54
TeacherModel=DeepSeek-Coder-6.7B, Category=Reference & Baselines2026.05—76.83
Total VariationFamily=Distance & Skewed Family2026.05—39.63
Total Variation + OursFamily=Distance & Skewed Family, Strategy=Bridge-Garden2026.05—40.02
α-β divergenceFamily=Unified Divergence Family2026.05—41.46
α-β divergence + OursFamily=Unified Divergence Family, Strategy=Bridge-Garden2026.05—43.9