Causal chain of thought improves reasoning by isolating context
Causal Chain of Thought: Mitigating Attention Leakage via Context Isolation
Machine LearningArtificial Intelligence
Summary
Chain-of-thought prompting helps language models explain how they reach answers step-by-step, but those steps don’t always actually cause the answer. The authors introduce CausalCoT, which breaks problems into small questions linked like a flowchart and solves each with only the information that directly influences it. This stops the model from shortcutting and making explanations that just sound plausible. Their approach keeps accuracy high while making the reasoning steps much more honest and lets users fix parts and see how corrections flow through the rest.
What this means in practice
- •For natural language processing engineers: Improve trustworthiness and diagnosability of language model reasoning by generating causally faithful explanations for each reasoning step.
- •For machine learning system maintainers: Enable targeted corrections in modular reasoning components with predictable propagation of fixes through subsequent inference steps.
Authors
Gabriele Dominici, Teague McMillan, Fazl Barez, Marc Langheinrich, Martin Gjoreski
Abstract
Chain-of-thought (CoT) prompting elicits step-by-step rationales that appear to explain model predictions, but there is no guarantee these steps causally drive the final answer. In standard autoregressive generation, dense attention allows each reasoning step to attend to the full context, enabling the model to derive answers through direct lookup while generating plausible-sounding but post-hoc justifications. We introduce CausalCoT, an inference framework that enforces causal transparency by decomposing problems into a directed acyclic graph of atomic sub-questions, evaluating each node in an isolated context containing only its direct causal parents. This Markovian structure removes attention pathways to non-parent content, making downstream predictions sensitive to local reasoning rather than global context leakage. Across a diverse suite of reasoning benchmarks, CausalCoT preserves competitive predictive performance while improving explanation faithfulness, increasing causal sensitivity of intermediate reasoning steps by an average of 12 and up to 21 percentage points over the strongest baseline. The structured dependency graph additionally enables test-time interventions in which local corrections propagate predictably through downstream reasoning.