MMODELYST
Evals/GSM8K
EVL

GSM8K

Grade-school math

Source ↗
benchmarkMathAccuracy (% of final answers correct)8.5K items
What this measures

Grade-school math word problems — multi-step arithmetic you'd see in elementary school. A classic check of step-by-step reasoning.

Sample
Natalia sold clips to 48 friends in April, then half as many in May. How many clips did she sell altogether? → 72
Leaderboard · 10 models
Accuracy (% of final answers correct) — higher is better
1Qwen2.5-72B-Instruct
95.8
2Llama-3.1-70B-Instruct
95.1
3Llama-3.3-70B-Instruct
95
4Qwen2.5-7B-Instruct
91.6
5DeepSeek-V3
89.3
6gemma-2-27b-it
89
7Phi-3.5-mini-instruct
86
8Llama-3.1-8B-Instruct
84.5
9gemma-2-9b-it
81
10Mistral-7B-Instruct-v0.3
53
Benchmark health
active

Discriminating but climbing — the frontier is moving up this benchmark.

86.0
top-10 mean
95.8
best
10
models
What improves this score

Datasets & environments shown to raise it.

—+17.0
Orca-Math+12.5
—+10.0
NuminaMath CoT+8.0
—+8.0
Tülu 3 SFT Mixture+7.0
—+7.0
MetaMathQA+6.0
Research using this benchmark
all →
Off-the-Shelf LLMs as Process Scorers: Training-Free Alternative to PRMs for Mathematical ReasoningLayerRoute: Input-Conditioned Adaptive Layer Skipping via LoRA Fine-Tuning for Agentic Language ModelsThe Chain Holds, the Answer Folds: Trace-Answer Dissociation in Reasoning Models Under Adversarial PressureThe Distillation Game: Adaptive Attacks & Efficient DefensesLanguage Models are Multilingual Chain-of-Thought Reasoners

Matched by name in title/abstract.