SpiralThinker: Latent Reasoning through an Iterative Process with Text-Latent Interleaving

TL;DR

SpiralThinker addresses latent reasoning with an iterative latent-update mechanism interleaved with explicit textual reasoning, enabling deeper, coherent internal computation without extra token generation. It introduces a data scheme that alternates text and $N$ latent tokens, a lightweight latent adapter, and a progressive alignment objective to supervise latent dynamics across $K$ iterations. Empirical results on GSM8K-Aug, ProsQA, and StrategyQA show state-of-the-art performance among latent reasoning methods, with ablations confirming the necessity of both iteration and alignment and revealing dataset-specific optima for $N$ and $K$. The approach offers a principled bridge between explicit interpretability and implicit computational depth, enhancing robustness across diverse reasoning tasks while maintaining token efficiency, with potential for dynamic iteration in future work.

Abstract

Recent advances in large reasoning models have been driven by reinforcement learning and test-time scaling, accompanied by growing interest in latent rather than purely textual reasoning. However, existing latent reasoning methods lack mechanisms to ensure stable evolution of latent representations and a systematic way to interleave implicit and explicit reasoning. We introduce SpiralThinker, a unified framework that performs iterative updates over latent representations, enabling extended implicit reasoning without generating additional tokens. A progressive alignment objective combined with structured annotations maintains coherence between latent and textual reasoning. Across mathematical, logical, and commonsense reasoning tasks, SpiralThinker achieves the best overall performance among latent reasoning approaches, consistently surpassing previous methods across all benchmarks. Detailed analyses reveal that both iteration and alignment are indispensable, the numbers of latent tokens and iterations exhibit dataset-specific optima, and appropriate alignment proves critical for an effective iterative process. Overall, SpiralThinker bridges iterative computation and latent reasoning, demonstrating that aligned iterative updates can reliably steer reasoning in the latent space.

Figures (5)

• Figure 1: (a) Explicit reasoning processes textual tokens once. (b) Implicit reasoning processes latent representations once. (c) SpiralThinker interleaves textual and latent reasoning through an iterative process.
• Figure 2: Training process of SpiralThinker. Step indicates a textual step, and <latent> indicates a latent step. Only one <latent> token is illustrated for clarity.
• Figure 3: Accuracy on different datasets as the number of latent tokens varies.
• Figure 4: Accuracy on different datasets as the number of iterations varies.
• Figure 5: The upper part shows the reasoning steps generated by SpiralThinker for a sample problem, while the lower part presents the top three tokens most similar to each latent representation at the first latent step during the iterative process. The top-ranked token is highlighted in red.