---
language:
- en
pipeline_tag: text-generation
---
# Under Experiment

> GOALS: SOTA math reasoning for sub-400M parameter LLM

- Base Model: [LiquidAI/LFM2-350M-Math](https://huggingface.co/LiquidAI/LFM2-350M-Math)
- Dataset: [kreasof-ai/RSA-Distillation-Mix](https://huggingface.co/datasets/kreasof-ai/RSA-Distillation-Mix)

## Benchmark

Note: we use thinking token forcing because this model occasionally output response directly without thinking tag.

Standard Decoding:
- AIME 2025: 28.3% (+1.2% from `LiquidAI/LFM2-350M-Math`) -> Link for benchmark output: https://docs.google.com/spreadsheets/d/1Gr0AFT08tWQ8ocPK3TEwfwz1h336Pfe4TB_IQxKDt3A/
- HMMT 2025: TBA
- BRUMO 2025: TBA
- CMIMC 2025: TBA

Recursive Self-Aggregation:
- AIME 2025: TBA
- HMMT 2025: TBA
- BRUMO 2025: TBA
- CMIMC 2025: TBA