# Tina > LoRA-based RL post-training of a 1.5B model achieves competitive reasoning at 260x lower cost than prior SOTA - Authors: Hunter - Published: 2025-05-02 - Source: Bevaya Labs, --- # Tina: Tiny Reasoning Models via LoRA **Paper**: [Tina: Tiny Reasoning Models via LoRA](https://arxiv.org/abs/2504.15777) **Presenter**: Hunter **Note Taker**: Hunter **Date**: May 2, 2025 --- ## Overview Tina demonstrates that strong reasoning ability can be developed in a 1.5B-parameter model via RL post-training with LoRA adapters, at dramatically lower cost than existing approaches. Post-training cost is under $9, representing an estimated 260× cost reduction compared to prior state-of-the-art reasoning models. The resulting model achieves 43.33% Pass@1 on AIME24, with over 20% improvement in reasoning performance relative to the base model. All code, training logs, and checkpoints are released. --- ## Novelty - LoRA applied during RL post-training (not just SFT), keeping most base model weights frozen while adapting for reasoning. - Demonstrates that RL reward shaping rapidly adapts the model to structured reasoning formats without full fine-tuning. - Very low compute budget: post-training completes for under $9, making the recipe accessible to resource-constrained teams. - Consistent hyperparameters validated across multiple reasoning benchmarks. --- ## Learnings - LoRA adapters are sufficient for RL-based reasoning acquisition in small models; full fine-tuning is not required. - RL reward shaping quickly instills structured reasoning behavior (e.g., chain-of-thought format) without large datasets. - A 1.5B model with RL+LoRA can reach competitive reasoning performance, narrowing the gap with much larger models. - The open-source release (weights, logs, code) enables direct replication and adaptation. --- ## Notes for Application - LoRA-based RL is a practical path for adding reasoning capability to small deployed models without high compute overhead. - The under-$9 post-training cost suggests this approach is viable for rapid iteration or domain-specific fine-tuning on limited budgets. - For tasks requiring structured output (extraction, form parsing), the same RL+LoRA recipe could reward well-formed outputs without full fine-tuning. - The 1.5B scale is potentially deployable in latency-sensitive production settings while retaining meaningful reasoning ability.