CausalCoT
A framework that improves the faithfulness of chain-of-thought reasoning by decomposing problems into a directed acyclic graph and isolating context, so that each subproblem can only see its direct causal parent nodes.
Timeline (last 90 days)
CausalCoT is a framework that decomposes a problem into a directed acyclic graph (DAG) of atomic subproblems, and when solving each node allows the model to see only its direct causal parents.
CausalCoT uses context isolation to cut off attention pathways to content of non-parent nodes, making downstream predictions sensitive to local reasoning rather than relying on global information leakage.
In CausalCoT, each node's value is conditioned only on its direct parents and is conditionally independent of earlier ancestors given the parents, which is equivalent to computing that node's conditional probability.
Experiments in the paper show that while maintaining competitive predictive performance, CausalCoT improves the causal sensitivity of intermediate reasoning steps by an average of 12 percentage points over the strongest baseline, with a maximum improvement of 21 percentage points.
CausalCoT's structured dependency graph supports test-time intervention: after a local reasoning step is corrected, its effect propagates predictably to downstream nodes along the causal graph.
CausalCoT presupposes that a problem can be reliably decomposed into a DAG structure; for tasks with ambiguous causal structure or high coupling, the accuracy of graph construction may become a bottleneck.
CausalCoT adopts an approach of solving each node in isolation, which incurs additional overhead in reasoning efficiency.
All Facts (7)
CausalCoT is a framework that decomposes a problem into a directed acyclic graph (DAG) of atomic subproblems, and when solving each node allows the model to see only its direct causal parents.
50%UnverifiedCausalCoT uses context isolation to cut off attention pathways to content of non-parent nodes, making downstream predictions sensitive to local reasoning rather than relying on global information leakage.
50%UnverifiedIn CausalCoT, each node's value is conditioned only on its direct parents and is conditionally independent of earlier ancestors given the parents, which is equivalent to computing that node's conditional probability.
50%UnverifiedExperiments in the paper show that while maintaining competitive predictive performance, CausalCoT improves the causal sensitivity of intermediate reasoning steps by an average of 12 percentage points over the strongest baseline, with a maximum improvement of 21 percentage points.
50%UnverifiedCausalCoT's structured dependency graph supports test-time intervention: after a local reasoning step is corrected, its effect propagates predictably to downstream nodes along the causal graph.
50%UnverifiedCausalCoT presupposes that a problem can be reliably decomposed into a DAG structure; for tasks with ambiguous causal structure or high coupling, the accuracy of graph construction may become a bottleneck.
50%UnverifiedCausalCoT adopts an approach of solving each node in isolation, which incurs additional overhead in reasoning efficiency.
50%