LLMs Have a Heart of Stone: Demystifying the Soft Thinking Ability of Large Reasoning Models

2025-08-05

AI Summary

Current large language models (LLMs) often degenerate into greedy decoding over the most salient components of soft inputs when performing “soft reasoning” in continuous concept spaces, thereby suppressing reasoning-path diversity and undermining the expressive potential of soft tokens. This work identifies this mechanistic limitation and proposes injecting controllable stochasticity to unlock soft reasoning capabilities: we explicitly model uncertainty during generation by integrating Dirichlet resampling with the Gumbel-Softmax reparameterization. Using probing techniques to analyze internal representations, we validate the approach across eight reasoning benchmarks. Results demonstrate that Gumbel-Softmax significantly enhances soft reasoning performance, shifting LLMs from deterministic decoding toward more exploratory, continuous-space inference. Our method establishes a novel paradigm for soft abstraction modeling—grounded in principled uncertainty quantification—while preserving architectural compatibility and computational tractability.

Technology Category

• Reasoning under Uncertainty: Stochastic Optimization
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Search and Retrieval-Augmented AI: Large language models for search

Abstract

Human cognition naturally engages with abstract and fluid concepts, whereas existing reasoning models often rely on generating discrete tokens, potentially constraining their expressive capabilities. Recent advancements aim to address this limitation by enabling large language models (LLMs) to generate soft, abstract tokens, thus facilitating reasoning within a continuous concept space. This paper explores the `Soft Thinking' capabilities of various LLMs by examining the models' internal behavior using a suite of probing techniques. Contrary to the common belief that Soft Thinking enables the simultaneous exploration of diverse reasoning paths, our findings reveal that LLMs predominantly rely on the most influential component of the soft inputs during subsequent decoding steps. This reliance hinders the exploration of different reasoning paths and reduces vanilla Soft Thinking to a form of greedy decoding, obscuring the advantage of transmitting more information through Soft Tokens. To tackle this issue, we explore sampling strategies to introduce emph{randomness}, employing methods such as Dirichlet resampling and the Gumbel-Softmax trick. Our experiments demonstrate that incorporating randomness can alleviate the limitations of vanilla approaches and unleash the potential of Soft Thinking. Notably, the Gumbel-Softmax trick provides adequate randomness with controlled smoothness, resulting in superior performance across eight reasoning benchmarks.

Problem

• LLMs lack fluid abstract reasoning like human cognition
• Soft Thinking in LLMs reduces to greedy decoding
• Randomness strategies improve Soft Thinking performance in reasoning

Innovation

• Generating soft abstract tokens for reasoning
• Using probing techniques to analyze model behavior
• Employing Gumbel-Softmax for controlled randomness