BenchmarkHMMT 2025

HMMT 2025

A collection of problems from the 2025 Harvard-MIT Mathematics Tournament, commonly used as an evaluation benchmark for the mathematical reasoning capabilities of large language models.

Timeline (last 90 days)

Oct 30

Pruning-based weakening raises the accuracy of Qwen3.5-4B on HMMT 2025 from 57.7% to 70.5%.

Unverified50%

All Facts (1)

Unverified

Pruning-based weakening raises the accuracy of Qwen3.5-4B on HMMT 2025 from 57.7% to 70.5%.

50%

Source Articles

Reverse Self-Distillation: Weakening a Model Can Actually Improve Its ReasoningOct 30