CLaM: Bridging Explicit and Implicit Chain-of-Thought via Controllable Latent Mediation

ACL ARR 2026 January Submission10877 Authors

Keywords: Chain-of-Thought, Latent Space Embedding, Multi-hop Reasoning

Abstract: Chain-of-Thought (CoT) improves multi-step reasoning in LLMs, but it highlights a tension between efficiency and controllability: explicit CoT exposes a rationale channel that can be edited at inference time to steer outputs, whereas implicit/latent CoT internalizes reasoning and operates under an answer-only interface, making targeted intervention difficult. We characterize this gap with a counterfactual framework that distinguishes input-level perturbations from mediator-level interventions, and show empirically that explicit and implicit systems can appear similar under input counterfactuals yet diverge sharply when direct control over intermediate reasoning is required. Motivated by this boundary, we propose **CLaM** (**C**ontrollable **La**tent **M**ediation), which restores an intervention handle for implicit reasoning without emitting rationales: an extractor maps structured intermediate facts into a small set of latent mediator embedding that condition an answer-only student model. Across multiple backbones and editing settings, CLaM enables robust counterfactual interventions and reliable propagation from edited intermediates to final answers, improving controllability while preserving the efficiency of latent reasoning. Our data and code will be available at \url{https://github.com/XXX}.

Paper Type: Long

Research Area: Interpretability and Analysis of Models for NLP

Research Area Keywords: Interpretability and Analysis of Models for NLP, Question Answering

Contribution Types: Model analysis & interpretability

Languages Studied: English

Submission Number: 10877