René F. Kizilcec

Hansol Lee, AJ Alvero, René F. Kizilcec et al.

Algorithmic predictions are inherently uncertain: even models with similar aggregate accuracy can produce different predictions for the same individual, raising concerns that high-stakes decisions may become sensitive to arbitrary modeling choices. In this paper, we define \emph{algorithmic sensitivity} as the extent to which arbitrary modeling choices propagate into human decisions: how much a decision outcome shifts when a more favorable versus less favorable algorithmic prediction is presented to the decision-maker for the same individual. We estimate this in a randomized field experiment ($n=19{,}545$) embedded in a selective U.S. college admissions cycle, in which admissions officers reviewed each application alongside an algorithmic score while we randomly varied whether the score came from one of two similarly accurate prediction models. Although the two models performed similarly in aggregate, they frequently assigned different scores to the same applicant, creating exogenous variation in the score shown. Surprisingly, we find little evidence of algorithmic sensitivity: presenting a more favorable score does not meaningfully increase an applicant's probability of admission on average, even when the models disagree substantially. These findings suggest that, in this expert, high-stakes setting, human decision-making is largely invariant to arbitrary variation in algorithmic predictions, underscoring the role of professional discretion and institutional context in mediating the downstream effects of algorithmic uncertainty.

Zhuqian Zhou, Kirk Vanacore, Bakhtawar Ahtisham et al.

Large-scale sharing of dialogue-based data is instrumental for advancing the science of teaching and learning, yet rigorous de-identification remains a major barrier. In mathematics tutoring transcripts, numeric expressions frequently resemble structured identifiers (e.g., dates or IDs), leading generic Personally Identifiable Information (PII) detection systems to over-redact core instructional content and reduce dataset utility. This work asks how PII can be detected in math tutoring transcripts while preserving their educational utility. To address this challenge, we investigate the "numeric ambiguity" problem and introduce MathEd-PII, the first benchmark dataset for PII detection in math tutoring dialogues, created through a human-in-the-loop LLM workflow that audits upstream redactions and generates privacy-preserving surrogates. The dataset contains 1,000 tutoring sessions (115,620 messages; 769,628 tokens) with validated PII annotations. Using a density-based segmentation method, we show that false PII redactions are disproportionately concentrated in math-dense regions, confirming numeric ambiguity as a key failure mode. We then compare four detection strategies: a Presidio baseline and LLM-based approaches with basic, math-aware, and segment-aware prompting. Math-aware prompting substantially improves performance over the baseline (F1: 0.821 vs. 0.379) while reducing numeric false positives, demonstrating that de-identification must incorporate domain context to preserve analytic utility. This work provides both a new benchmark and evidence that utility-preserving de-identification for tutoring data requires domain-aware modeling.

Bakhtawar Ahtisham, Kirk Vanacore, Zhuqian Zhou et al.

Large Language Models (LLMs) are increasingly deployed to automatically label and analyze educational dialogue at scale, yet current pipelines lack reliable ways to detect when models are wrong. We investigate whether reasoning generated by LLMs can be used to predict the correctness of a model's own predictions. We analyze 30,300 teacher utterances from classroom dialogue, each labeled by multiple state-of-the-art LLMs with an instructional move construct and an accompanying reasoning. Using human-verified ground-truth labels, we frame the task as predicting whether a model's assigned label for a given utterance is correct. We encode LLM reasoning using Term Frequency-Inverse Document Frequency (TF-IDF) and evaluate five supervised classifiers. A Random Forest classifier achieves an F1 score of 0.83 (Recall = 0.854), successfully identifying most incorrect predictions and outperforming baselines. Training specialist detectors for specific instructional move constructs further improves performance on difficult constructs, indicating that error detection benefits from construct-specific linguistic cues. Using the Linguistic Inquiry and Word Count (LIWC) framework, we examine four linguistic markers of correctness: Causation, Differentiation, Tentativeness, and Insight. Correct predictions exhibit grounded causal language (e.g., because, therefore), while incorrect reasoning is substantially more likely to rely on epistemic hedging (e.g., might, could) and performative metacognition (e.g., think, realize). Syntactic complexity does not distinguish correct from incorrect reasoning, and longer reasoning is not more reliable. These findings demonstrate that reasoning-based error detection offers a practical and scalable approach to quality control in automated educational dialogue analysis.

Kirk Vanacore, Danielle R Thomas, Digory Smith et al.

This paper introduces a scalable causal inference framework for estimating the immediate, session-level effects of on-demand human tutoring embedded within adaptive learning systems. Because students seek assistance at moments of difficulty, conventional evaluation is confounded by self-selection and time-varying knowledge states. We address these challenges by integrating principled analytic sample construction with Deep Knowledge Tracing (DKT) to estimate latent mastery, followed by doubly robust estimation using Causal Forests. Applying this framework to over 5,000 middle-school mathematics tutoring sessions, we find that requesting human tutoring increases next-problem correctness by approximately 4 percentage points and accuracy on the subsequent skill encountered by approximately 3 percentage points, suggesting that the effects of tutoring have proximal transfer across knowledge components. This effect is robust to various forms of model specification and potential unmeasured confounders. Notably, these effects exhibit significant heterogeneity across sessions and students, with session-level effect estimates ranging from $-20.25pp$ to $+19.91pp$. Our follow-up analyses suggest that typical behavioral indicators, such as student talk time, do not consistently correlate with high-impact sessions. Furthermore, treatment effects are larger for students with lower prior mastery and slightly smaller for low-SES students. This framework offers a rigorous, practical template for the evaluation and continuous improvement of on-demand human tutoring, with direct applications for emerging AI tutoring systems.