This repository hosts LACONIC-DeepSeek-R1-Distill-1.5B-1500, a LACONIC-trained variant of DeepSeek-R1-Distill-Qwen-1.5B.

LACONIC is a length-aware reinforcement learning method for making LLM responses substantially shorter while preserving task performance. During training, it combines task reward with an adaptive length-based cost so that the model learns to stay near a target response budget. This checkpoint targets a budget of 1500 tokens.

In practice, LACONIC is designed to reduce response length with minimal deployment overhead: the released model uses the usual decoding stack and does not require special inference-time control logic.