Unlocking the Black Box of Latent Reasoning: An Interpretability-Guided Approach to Intervention
Fuente:
arXiv
Gespeichert in:
| Hauptverfasser: | , , , , , , , |
|---|---|
| Format: | Preprint |
| Veröffentlicht: |
2026
|
| Schlagworte: | |
| Online-Zugang: | |
| Tags: |
Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
|
| _version_ | 1866913177585844224 |
|---|---|
| author | Chang, Shuochen Bai, Tong Zhang, Xiaofeng Ma, Qianli Liu, Qingyang Liao, Zhaohe Miao, Yibo Niu, Li |
| author_facet | Chang, Shuochen Bai, Tong Zhang, Xiaofeng Ma, Qianli Liu, Qingyang Liao, Zhaohe Miao, Yibo Niu, Li |
| contents | Latent reasoning enables Large Language Models (LLMs) to perform multi-step inference within continuous hidden states, offering efficiency gains over explicit Chain-of-Thought (CoT). However, the opacity of these continuous thought vectors hinders their reliability and controllability. This paper bridges the gap between mechanistic interpretability and actionable control. We first present a systematic analysis using structural, causal, and geometric probes, revealing that latent vectors encode compressed, faithful representations of reasoning steps, with early vectors acting as critical causal hubs. Building on this, we operationalize these interpretability insights into a suite of training-free, decode-time interventions that refine the latent reasoning process by imposing the identified geometric and semantic priors. Extensive experiments across multiple model scales and diverse task domains demonstrate that our approaches consistently improve reasoning accuracy. Our interpretability-guided interventions consistently unlock latent capabilities and improve reasoning accuracy without any parameter updates. |
| format | Preprint |
| id |
arxiv_https___arxiv_org_abs_2606_01243 |
| institution | arXiv |
| publishDate | 2026 |
| record_format | arxiv |
| spellingShingle | Unlocking the Black Box of Latent Reasoning: An Interpretability-Guided Approach to Intervention Chang, Shuochen Bai, Tong Zhang, Xiaofeng Ma, Qianli Liu, Qingyang Liao, Zhaohe Miao, Yibo Niu, Li Computation and Language Machine Learning Latent reasoning enables Large Language Models (LLMs) to perform multi-step inference within continuous hidden states, offering efficiency gains over explicit Chain-of-Thought (CoT). However, the opacity of these continuous thought vectors hinders their reliability and controllability. This paper bridges the gap between mechanistic interpretability and actionable control. We first present a systematic analysis using structural, causal, and geometric probes, revealing that latent vectors encode compressed, faithful representations of reasoning steps, with early vectors acting as critical causal hubs. Building on this, we operationalize these interpretability insights into a suite of training-free, decode-time interventions that refine the latent reasoning process by imposing the identified geometric and semantic priors. Extensive experiments across multiple model scales and diverse task domains demonstrate that our approaches consistently improve reasoning accuracy. Our interpretability-guided interventions consistently unlock latent capabilities and improve reasoning accuracy without any parameter updates. |
| title | Unlocking the Black Box of Latent Reasoning: An Interpretability-Guided Approach to Intervention |
| topic | Computation and Language Machine Learning |
| url | https://arxiv.org/abs/2606.01243 |