List your company

Mathematical Reasoning on Average GSM8k-Aug, GSM-Hard, SVAMP, MultiArith

79.3AccuracySFT-CoT
6.60425.47744.3563.223Jan 21, 2026Feb 16, 2026Mar 15, 2026Apr 10, 2026May 7, 2026Jun 2, 2026Jun 29, 2026

Evaluation results

MethodLinks
SFT-CoTModel=Qwen3-VL-4B-Instruct2026.0179.3108.40.73——
CoT-SFTBackbone=Qwen3-VL-4B-Instruct2026.0679.3———108.4
CoT-SFTBackbone=Qwen3-4B2026.0679.2———105.1
SFT-CoTModel=Qwen3-VL-2B-Instruct2026.0163.8117.50.54——
DLR (Full)Backbone=Qwen3-VL-4B-Instruct2026.0663.2———5.4
DLR (Full)Backbone=Qwen3-4B2026.0662.9———5.4
DLR (SFT)Backbone=Qwen3-VL-4B-Instruct2026.0657.3———5.2
DLR (SFT)Backbone=Qwen3-4B2026.0656.9———5.3
RoT (Ours)Model=Qwen3-VL-4B-Instruct2026.0155.4321.73——
RoTBackbone=Qwen3-VL-4B-Instruct2026.0655.4———32
CoT2026.0253.621.4———
CoLaRCompression ratio=22026.0248.810———
LSTRCompression ratio=22026.0248.411.4———
SFT-CoTModel=LLaVa-V1.6-Mistral-7B2026.0148.3130.80.37——
Direct-SFTBackbone=Qwen3-4B2026.0648.1———0
SFT-w/o CoTModel=Qwen3-VL-4B-Instruct2026.0147.80———
Direct-SFTBackbone=Qwen3-VL-4B-Instruct2026.0647.8———0
LSTRCompression ratio=2, Ablation=w/o Skip2026.0247.312.9———
CoLaRBackbone=Qwen3-4B2026.0647.3———31.8
ReGuLaRBackbone=LLaMA-3.2-1B-Instruct2026.0145.63.03———
CoLaRBackbone=Qwen3-VL-4B-Instruct2026.0645.1———31.8
LSTRCompression ratio=2, Ablation=w/o Sparse2026.0243.912.7———
LSTRCompression ratio=52026.0243.34.66———
LSTRCompression ratio=5, Ablation=w/o Skip2026.0241.95———
CoLaRCompression ratio=52026.0241.74.57———
CoLaRBackbone=LLaMA-3.2-1B-Instruct, Compression Rate=52026.0141.74.7———
RoT (Ours)Model=Qwen3-VL-2B-Instruct2026.0137321.16——
LSTRCompression ratio=5, Ablation=w/o Sparse2026.0235.64.7———
RoT (Ours)Model=LLaVa-V1.6-Mistral-7B2026.0134.3321.07——
CoconutBackbone=Qwen3-4B2026.0631.6———6
CoconutBackbone=Qwen3-VL-4B-Instruct2026.0630.5———6
SFT-w/o CoTModel=Qwen3-VL-2B-Instruct2026.0128.60———
Coconut2026.0227.66———
CoconutBackbone=LLaMA-3.2-1B-Instruct2026.0126.66———
SFT-w/o CoTModel=LLaVa-V1.6-Mistral-7B2026.0126.60———
iCoTBackbone=Qwen3-4B2026.0625.9———0
iCoT2026.0224.60———
iCoTBackbone=LLaMA-3.2-1B-Instruct2026.0124.60———
Distill2026.0214.36———
CODIBackbone=LLaMA-3.2-1B-Instruct2026.0114.36———
CODIBackbone=Qwen3-4B2026.069.7———6
CODIBackbone=Qwen3-VL-4B-Instruct2026.069.4———6
CoconutRationale Format=GSM8k-Aug2026.05—6—27.59—
CoLaR-2Rationale Format=GSM8k-Aug2026.05—10.04—48.84—
Ours-RLRationale Format=GSM8k-Aug2026.05—11.45—54.42—
Ours-SFTRationale Format=GSM8k-Aug2026.05—11.98—51.27—
SFT-CoTRationale Format=GSM8k-Aug2026.05—21.35—56.98—