LED: A New Latent Space Decoding Method to Restore Exploration Capability for Large Reasoning Models

This article introduces Latent Exploration Decoding (LED), an innovative method aimed at solving the problem of excessive conservatism in large reasoning models after post-training. By introducing exploratory noise into the model's latent representation space, LED restores the model's exploration capability while maintaining reasoning quality. The related research has been accepted by ICML 2026. Keywords: Reasoning models, latent space decoding, exploration capability, post-training optimization, Transformer, ICML 2026.

Large language models trained via reinforcement learning perform well in tasks like mathematical reasoning and code generation, but they have the side effect of excessive conservatism: they tend to choose the most confident path even if it means missing better solutions. They easily converge prematurely in complex reasoning tasks, limiting their performance in open-ended tasks.

Latent Exploration Decoding (LED) does not add randomness at the word level; instead, it adds exploratory noise in the latent representation space. The advantage is that it encourages the model to explore different reasoning paths while maintaining the fluency and coherence of the text. By controlling the noise intensity and distribution, a fine balance between exploration and exploitation is achieved.

1. Noise injection position: Middle layers of the Transformer (encoding high-level semantics with minimal disturbance to grammatical details); 2. Adaptive noise: Dynamically adjust distribution and scale according to the task, and adjust parameters based on the confidence of the decoding state; 3. Fallback mechanism: When exploration leads to quality degradation, fall back to a conservative decoding strategy to ensure reliability.

Although specific experimental data have not been fully disclosed, LED has shown significant improvements on multiple reasoning benchmarks, especially outperforming standard decoding in tasks requiring creative thinking or multi-path exploration. Moderate exploration can help the model find better solutions and even improve output quality.

1. Challenges the assumption of post-training model fixation; interventions in the decoding stage can improve behavior; 2. Emphasizes the importance of the latent representation space, inspiring research related to representation engineering; 3. Provides a dimension of controllability: by adjusting noise parameters to balance exploration and reliability, supporting customized applications.

Limitations: Optimal parameters depend on tasks and models; general settings remain to be solved; increases computational overhead, which needs to be considered in latency-sensitive scenarios. Future directions: Fine-grained noise injection strategies (e.g., attention-based adaptive noise), combination with other decoding techniques, applications in specific fields (scientific discovery, drug design).

LED provides an innovative and practical solution to restore the exploration capability of reasoning models, alleviating excessive conservatism through controlled randomness in the latent space. It not only has practical value but also opens up new paths for understanding and manipulating the internal behavior of large models, and will play an important role in balancing model reliability and creativity.