Towards Robust Causal Deep Learning

Abstract

Deep learning is increasingly used not only for prediction, but also for causal inference, sequential decision-making, and the monitoring of language-model reasoning. In these settings, predictive accuracy alone is not enough: the learned computation must align with the causal, decision-theoretic, or counterfactual object being estimated. This dissertation develops methods for strengthening the reliability of deep learning in three settings where standard tools can silently fail: causal effect estimation from observational data, finite-budget Monte Carlo Tree Search, and chain-of-thought reasoning in large language models. The common thesis is that robustness comes from imposing the right structural inductive bias. Each chapter identifies a mismatch between a flexible learning architecture and the statistical object it is asked to estimate, then introduces a principled correction that uses known structure rather than ignoring it. The first chapter introduces the DAG-aware Transformer, a neural architecture for causal effect estimation that embeds a supplied causal directed acyclic graph as a hard structural constraint on attention. Standard attention permits all variables to exchange information, but causal variables are not exchangeable tokens: treatments, outcomes, confounders, and proxies play asymmetric roles in identification. The proposed architecture masks attention scores according to the causal graph and omits layer normalization, which can distort heterogeneous causal variables by forcing them onto a common scale. This produces representations whose information flow respects the causal pathways assumed by the estimand. The architecture supports a unified family of estimators, including outcome-regression, propensity-score, augmented inverse-probability-weighted, and proximal bridge-function estimators. Across the LaLonde, ACIC, and Demand benchmarks, the DAG-aware Transformer improves over classical nonparametric baselines, modular multilayer perceptrons, and causally agnostic graph and Transformer architectures. Ablations further show that robustness depends on the credibility of the supplied graph: a correctly specified DAG improves estimation, while a misspecified DAG can become a hard wrong constraint. The second chapter develops Doubly Robust Monte Carlo Tree Search (DR-MCTS), a modification of MCTS that improves finite-budget search by replacing the standard mean backup with an adaptive hybrid of Monte Carlo rollouts and doubly robust off-policy correction. Standard MCTS backs up returns generated by the behavior policy induced by PUCT exploration and rollout sampling, while the search decision depends on values under a target policy induced by the current tree. This behavior-target mismatch is especially costly when each rollout requires calls to a large language model, tests, tool use, or environment interaction. DR-MCTS estimates the behavior policy, target policy, value functions, and residual corrections from tree statistics, avoiding the need to train a separate value model. A finite-sample analysis shows that when the hybrid backup reduces per-node mean squared error relative to the standard MCTS backup, the number of simulations sufficient for best-arm identification is reduced proportionally. Gridworld and closed-form contextual-bandit experiments validate this theorem-regime behavior. HumanEval experiments show that DR-MCTS can solve code-generation tasks earlier at pass@1 parity when integrated into an LLM tree-search agent, while WebShop probes a harder sparse-reward, long-horizon regime and clarifies when the adaptive hybrid should lean toward or away from the doubly robust correction. The third chapter turns to the integrity of chain-of-thought reasoning in language models. Chain-of-thought can make model behavior appear interpretable, but a visible reasoning trace is useful for monitoring only if it is both human-readable and counterfactually connected to the answer. This chapter introduces three lightweight, task-agnostic health metrics: Necessity, which tests whether the answer depends on the presence of the reasoning trace; Paraphrasability, which tests whether semantically equivalent reasoning preserves the answer; and Substantivity, which tests whether replacing the reasoning with irrelevant content changes the answer. These metrics are validated against deliberately constructed model organisms exhibiting healthy reasoning, post-hoc rationalization, encoded reasoning, and internalized reasoning. Across supervised fine-tuning checkpoints and multiple reasoning datasets, the metrics distinguish different failure modes and provide a practical diagnostic suite for assessing the monitorability of language-model reasoning. Together, the three chapters argue for a structural view of robust causal deep learning. A causal graph can discipline attention; an off-policy correction can make tree search more sample-efficient; and counterfactual interventions can diagnose whether language-model reasoning traces are faithful enough to monitor. Across causal identification, sequential search, and reasoning in language models, the dissertation shows that deep learning systems become more reliable when their internal computation is constrained by the structure of the question they are meant to answer.