List your company
ResearchFormerly WizWand

Mathematical Reasoning on Olympiad (ACC, LEN)

46.1AccuracyReLIFT
-1.4810410.8717323.224535.57727Jun 9, 2025Aug 6, 2025Oct 3, 2025Dec 1, 2025Jan 28, 2026Mar 27, 2026May 25, 2026
Updated 4mo ago

Evaluation results

MethodLinks
ReLIFTBackbone=Qwen2.5-7B2025.0646.13,649
SFTBackbone=Qwen2.5-7B2025.0644.36,234
RLBackbone=Qwen2.5-7B2025.0642.21,438
KAEBackbone=Qwen2.5-Math-7B, Post-training Dataset=MATH (levels 3-5, 5k+), Training Steps=200, Group Size=G = 8, Temperature=0.62026.0439.7—
ReLIFTBackbone=Qwen2.5-Math-1.5B2025.0639.62,274
GRPOBackbone=Qwen2.5-Math-7B, Post-training Dataset=MATH (levels 3-5, 5k+), Training Steps=200, Group Size=G = 8, Temperature=0.62026.0438.52—
Dr. GRPOBackbone=Qwen2.5-Math-7B, Post-training Dataset=MATH (levels 3-5, 5k+), Training Steps=200, Group Size=G = 8, Temperature=0.62026.0437.48—
Qwen-7B-InstructBackbone=Qwen2.5-7B2025.0637.31,550
RLBackbone=Qwen2.5-Math-1.5B2025.0637.21,695
GPGBackbone=Qwen2.5-Math-7B, Post-training Dataset=MATH (levels 3-5, 5k+), Training Steps=200, Group Size=G = 8, Temperature=0.62026.0436.89—
Qwen-Math-1.5B-InstructBackbone=Qwen2.5-Math-1.5B2025.0635.72,834
SFTBackbone=Qwen2.5-Math-1.5B2025.0633.87,231
Qwen-7BBackbone=Qwen2.5-7B2025.0626.51,068
Qwen-Math-1.5BBackbone=Qwen2.5-Math-1.5B2025.0618.51,619
LLaMa-8B-InstructBackbone=LLaMA-3.1-8B2025.0613.81,063
ReLIFTBackbone=LLaMA-3.1-8B2025.06111,050
SFTBackbone=LLaMA-3.1-8B2025.068.61,997
RLBackbone=LLaMA-3.1-8B2025.067.3790
Qwen3-8B (Ours)Backbone=Qwen3-8B2026.040.673—
Qwen3-4B (Ours)Backbone=Qwen3-4B2026.040.671—
Disagreement-Guided Strategy RoutingBackbone=Qwen3-32B2026.040.628—
JETBackbone=DS-Distill-Llama-8B2025.090.5915,822
RL-PLUSBase Model=Qwen2.5-Math-7B, Training Method=RL-PLUS2025.070.588—
RL-PLUSBase Model=Qwen2.5-Math-7B2025.070.588—
DS-Llama-8B (Ours)Backbone=DS-Llama-8B2026.040.585—
SCoPBackbone=Qwen3-32B2026.040.577—
LUFFYBase Model=Qwen2.5-Math-7B2025.070.572—
GTModel=Qwen3-8B-Base2026.050.557—
BoNBackbone=Qwen3-32B2026.040.556—
ReLIFTBase Model=Qwen2.5-Math-7B2025.070.548—
CAREModel=Qwen3-8B-Base2026.050.5434—
MajorityBackbone=Qwen3-32B2026.040.543—
DVBackbone=Qwen3-32B2026.040.54—
BaseBackbone=DS-Distill-Llama-8B2025.090.5248,193
DAPOBackbone=DS-Distill-Llama-8B2025.090.5246,463
All-Pairs+Backbone=qwen3-4b-base, nll coefficient=1.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.519—
MPO+Backbone=qwen3-4b-base, nll coefficient=1.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.519—
GTModel=Qwen3-4B-Base2026.050.5117—
All-Pairs+Backbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.51—
Softmax+Backbone=qwen3-4b-base, nll coefficient=1.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.508—
BaselineBackbone=Qwen3-32B2026.040.505—
SFT+GRPOBase Model=Qwen2.5-Math-7B2025.070.504—
Margin+Backbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.502—
MPO+Backbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.502—
Qwen2.5-7B-Math + SFT + RL-TCERBase Model=Qwen2.5-7B-Math, Training Strategy=SFT + RL-TCER2026.040.496—
DAPOBase Model=Qwen2.5-Math-7B2025.070.496—
DPOBackbone=DS-Distill-Llama-8B2025.090.4956,868
STABLE-OPDBackbone=Qwen2.5-Math-7B, Teacher=OpenThinkerV32026.040.493—
DPOBackbone=qwen3-4b-base, nll coefficient=1.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.49—
DPOBackbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.481—
OracleModel=Qwen3-8B-Base2026.050.4806—
Margin+Backbone=qwen3-4b-base, nll coefficient=1.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.479—
OPDBackbone=Qwen2.5-Math-7B, Teacher=OpenThinkerV32026.040.475—
Softmax+Backbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.474—
Re-ScheduleBackbone=Qwen3-4B-Base2025.100.474—
SFTBackbone=DS-Distill-Llama-8B2025.090.4736,249
OPENREASONER-ZEROBackbone=Qwen2.5-Math-7B2026.040.471—
OpenReasonerBase Model=Qwen2.5-Math-7B2025.070.471—
GRPOBackbone=Qwen2.5-Math-7B2026.040.468—
Qwen2.5-7B-Math + SFT + RL-EndoRBase Model=Qwen2.5-7B-Math, Training Strategy=SFT + RL-EndoR2026.040.468—
GRPOBase Model=Qwen2.5-Math-7B, Training Method=GRPO2025.070.468—
GRPOBase Model=Qwen2.5-Math-7B2025.070.468—
RFTBackbone=qwen3-4b-base, nll coefficient=0.0, Setting=Online, Checkpoint selection=Highest checkpoint2026.040.467—
EntropyModel=Qwen3-8B-Base2026.050.4628—
TAPOBase Model=Qwen2.5-Math-7B2025.070.462—
CAREModel=Qwen3-4B-Base2026.050.4612—
ProbModel=Qwen3-8B-Base2026.050.4591—
RandomModel=Qwen3-8B-Base2026.050.4571—
TTRLModel=Qwen3-8B-Base2026.050.4547—
RL-PLUSBase Model=Qwen2.5-Math-1.5B, Training Method=RL-PLUS2025.070.452—
OracleModel=Qwen3-4B-Base2026.050.4495—
SFTBackbone=Qwen2.5-Math-7B2026.040.437—
Qwen2.5-7B-Math + SFTBase Model=Qwen2.5-7B-Math, Training Strategy=SFT2026.040.437—
SFTBase Model=Qwen2.5-Math-7B, Training Method=SFT2025.070.437—
SFTBase Model=Qwen2.5-Math-7B2025.070.437—
OAT-ZEROBackbone=Qwen2.5-Math-7B2026.040.434—
OatBase Model=Qwen2.5-Math-7B2025.070.434—
GRPOBackbone=Qwen3-4B-Base2025.100.434—
Re-Schedule_sigmoidBackbone=Qwen2.5-7B, Weight Function=sigmoid2025.100.433—
RandomModel=Qwen3-4B-Base2026.050.4254—
ProbModel=Qwen3-4B-Base2026.050.4229—
Re-Schedule_linearBackbone=Qwen2.5-7B, Weight Function=linear2025.100.421—
TTRLModel=Qwen3-4B-Base2026.050.4201—
ACCBackbone=Qwen3-4B-Base2025.100.42—
EntropyModel=Qwen3-4B-Base2026.050.412—
QWEN2.5-MATH-7B-INSTRUCTBackbone=Qwen2.5-Math-7B2026.040.41—
Qwen2.5-Math-7B-InstructBase Model=Qwen2.5-Math-7B2025.070.41—
GRPOBackbone=Qwen2.5-7B, Category=Classical RLVR Methods2025.100.404—
PRIME-ZEROBackbone=Qwen2.5-Math-7B2026.040.403—
PRIMEBase Model=Qwen2.5-Math-7B2025.070.403—
SimpleRL-ZooBackbone=Qwen2.5-7B, Category=Classical RLVR Methods2025.100.403—
OPOBackbone=Qwen2.5-7B, Category=Classical RLVR Methods2025.100.403—
GRPO w/ SFT LossBase Model=Qwen2.5-Math-7B2025.070.394—
ACC_sigmoidBackbone=Qwen2.5-7B, Category=Scheduling Methods2025.100.394—
LPPOBackbone=Qwen2.5-7B, Category=Scheduling Methods2025.100.367—
STEERBackbone=Qwen2.5-Math-1.5B, Sampling Strategy=avg@12025.100.366—
Seed-GRPOBackbone=Qwen2.5-7B, Category=Scheduling Methods2025.100.363—
OPOBackbone=Qwen2.5-Math-1.5B, Sampling Strategy=avg@12025.100.359—
Llama3.1-8B-Instruct + SFT + RL-TCERBase Model=Llama3.1-8B-Instruct, Training Strategy=SFT + RL-TCER2026.040.353—
Entro. Adv.Backbone=Qwen2.5-Math-1.5B, Sampling Strategy=avg@12025.100.349—