SuperThoughts Speeds Up LLM Chain-of-Thought Reasoning

Long Chain-of-Thought (CoT) reasoning is a powerful technique for enhancing the problem-solving capabilities of large language models (LLMs), but its sequential token generation process makes it computationally expensive and slow. While some research has explored reasoning in continuous latent spaces to bypass discrete token generation, these methods often face challenges with training stability and struggle to scale to more complex, long-horizon tasks due to a lack of sufficient supervision signals. To address these limitations, researchers propose SuperThoughts. This innovative method compresses pairs of consecutive CoT tokens into single latent representations. A lightweight Multi-Token Prediction (MTP) module then decodes two tokens per step, effectively doubling the inference throughput. Crucially, this approach preserves discrete token supervision during training, ensuring stability and scalability. The method was fine-tuned on various Qwen2.5-Math-Instruct models (1.5B, 7B, 14B) and evaluated across challenging benchmarks like MATH500, AMC, OlympiadBench, and GPQA-Diamond. By incorporating a confidence-based adaptive mechanism that reverts to standard decoding when uncertainty is high, SuperThoughts achieved approximately 20-30% reduction in CoT length. This was accomplished while largely maintaining accuracy, with only a minimal 1-2 point accuracy drop on most tasks, demonstrating a significant improvement in efficiency without substantial performance degradation.