List your company
ResearchFormerly WizWand

Mathematical Reasoning on AIME 24 (test)

97.5AccuracyIterative Critique-and-Routing Controller
-3.922.42548.7575.075Aug 5, 2025Sep 22, 2025Nov 10, 2025Dec 29, 2025Feb 16, 2026Apr 6, 2026May 25, 2026
Updated 4mo ago

Evaluation results

MethodLinks
Iterative Critique-and-Routing ControllerController Backbone=Qwen3-8B-Base2026.0597.5
Agent 1Controller Backbone=Qwen3-4B-Base, Agent 1 Model=Qwen3-30B-A3B-Instruct2026.0583.3
Iterative Critique-and-Routing ControllerController Backbone=Qwen3-4B-Base2026.0583.3
Agent 1Controller Backbone=Qwen3-8B-Base, Agent 1 Model=Qwen3-30B-A3B-Instruct2026.0583.3
RoBERTa RouterController Backbone=Qwen3-8B-Base2026.0583.3
Router-R1Controller Backbone=Qwen3-8B-Base2026.0580
Router-R1Controller Backbone=Qwen3-4B-Base2026.0576.7
RoBERTa RouterController Backbone=Qwen3-4B-Base2026.0566.7
Solver + Verifier (AstraFlow)Backbone=Qwen3-8B, Framework=AstraFlow, Time/iter (s)=77.652026.0547.3
RouterDCController Backbone=Qwen3-8B-Base2026.0546.7
Solver + Verifier (verl)Backbone=Qwen3-8B, Framework=verl, Time/iter (s)=212.642026.0544.6
SolverBackbone=Qwen3-8B2026.0542.9
Controller V2Controller Backbone=Qwen3-4B-Base2026.0533.3
Random RouterController Backbone=Qwen3-4B-Base2026.0527.5
RouterDCController Backbone=Qwen3-4B-Base2026.0527.5
Random RouterController Backbone=Qwen3-8B-Base2026.0526.7
GAC + Token-φVariant=Token-φ2026.0520.8
Controller V2Controller Backbone=Qwen3-8B-Base2026.0520
GAC w/o φVariant=w/o φ2026.0520
HPTCategory=Recent hybrid SFT–RL2026.0518.7
LUFFYCategory=Recent hybrid SFT–RL2026.0518.5
KL-ctrlCategory=Rule-based controllers2026.0518.4
CHORDCategory=SFT–RL mixing baselines2026.0518.2
Nash-MTLCategory=Multi-objective solvers2026.0518.1
CAGradCategory=Multi-objective solvers2026.0518
GradNorm-ctrlCategory=Rule-based controllers2026.0518
SRFTCategory=Recent hybrid SFT–RL2026.0517.9
SFT-best + RLCategory=SFT–RL mixing baselines2026.0517.1
Controller V1Controller Backbone=Qwen3-8B-Base2026.0516.7
DPOCategory=RL-free alignment2026.0516.4
IPOCategory=RL-free alignment2026.0516.1
SFT-best2026.0515.8
Controller V1Controller Backbone=Qwen3-4B-Base2026.0513.3
GRPO (pure RL)Category=SFT–RL mixing baselines2026.0513.2
Qwen2.5-7B-Inst.2026.0511.7
Agent 2Controller Backbone=Qwen3-4B-Base, Agent 2 Model=Qwen2.5-7B-Instruct2026.0510
Agent 2Controller Backbone=Qwen3-8B-Base, Agent 2 Model=Ministral-3-8B-Instruct2026.0510
BaseBackbone=LLaMA 3.1-8B-Instruct, Optimization Setting=test-time optimization, Reward Supervision=noisy rule-based2025.0810
VRPOBackbone=LLaMA 3.1-8B-Instruct, Optimization Setting=test-time optimization, Reward Supervision=noisy rule-based2025.0810
Agent 3Controller Backbone=Qwen3-4B-Base, Agent 3 Model=Qwen2.5-1.5B-Instruct2026.056.7
GRPOBackbone=LLaMA 3.1-8B-Instruct, Optimization Setting=test-time optimization, Reward Supervision=noisy rule-based2025.086.67
PPOBackbone=LLaMA 3.1-8B-Instruct, Optimization Setting=test-time optimization, Reward Supervision=noisy rule-based2025.086.67
Agent 3Controller Backbone=Qwen3-8B-Base, Agent 3 Model=Llama-3-2-1B-Instruct2026.050