Zhengze Zhou

Yuanda Xu, Hejian Sang, Zhengze Zhou et al.

On-policy knowledge distillation (OPD) trains a student on its own rollouts under token-level supervision from a teacher. Not all token positions matter equally, but existing views of token importance are incomplete. We ask a direct question: which tokens carry the most useful learning signal in OPD? Our answer is that informative tokens come from two regions: positions with high student entropy, and positions with low student entropy plus high teacher--student divergence, where the student is overconfident and wrong. Empirically, student entropy is a strong first-order proxy: retaining $50\%$ of tokens with entropy-based sampling matches or exceeds all-token training while reducing peak memory by up to $47\%$. But entropy alone misses a second important region. When we isolate low-entropy, high-divergence tokens, training on fewer than $10\%$ of all tokens nearly matches full-token baselines, showing that overconfident tokens carry dense corrective signal despite being nearly invisible to entropy-only rules. We organize these findings with TIP (Token Importance in on-Policy distillation), a two-axis taxonomy over student entropy and teacher--student divergence, and give a theoretical explanation for why entropy is useful yet structurally incomplete. This view motivates type-aware token selection rules that combine uncertainty and disagreement. We validate this picture across three teacher--student pairs spanning Qwen3, Llama, and Qwen2.5 on MATH-500 and AIME 2024/2025, and on the DeepPlanning benchmark for long-horizon agentic planning, where Q3-only training on $<$$20\%$ of tokens surpasses full-token OPD. Our experiments are implemented by extending the OPD repository https://github.com/HJSang/OPSD_OnPolicyDistillation, which supports memory-efficient distillation of larger models under limited GPU budgets.

Xiaojian Zhang, Xiang Yan, Zhengze Zhou et al.

The growing significance of ridesourcing services in recent years suggests a need to examine the key determinants of ridesourcing demand. However, little is known regarding the nonlinear effects and spatial heterogeneity of ridesourcing demand determinants. This study applies an explainable-machine-learning-based analytical framework to identify the key factors that shape ridesourcing demand and to explore their nonlinear associations across various spatial contexts (airport, downtown, and neighborhood). We use the ridesourcing-trip data in Chicago for empirical analysis. The results reveal that the importance of built environment varies across spatial contexts, and it collectively contributes the largest importance in predicting ridesourcing demand for airport trips. Additionally, the nonlinear effects of built environment on ridesourcing demand show strong spatial variations. Ridesourcing demand is usually most responsive to the built environment changes for downtown trips, followed by neighborhood trips and airport trips. These findings offer transportation professionals nuanced insights for managing ridesourcing services.

Yuanda Xu, Hejian Sang, Zhengze Zhou et al.

Standard LLM distillation wastes compute on two fronts: problems the student has already mastered (near-zero gradients) and problems far beyond its reach (incoherent gradients that erode existing capabilities). We show that this waste is not merely intuitive but structurally inevitable: the gradient signal-to-noise ratio in distillation provably vanishes at both pass-rate extremes. This theoretical observation leads to Paced, a framework that concentrates distillation on the zone of proximal development -- the frontier of a student model's competence -- via a principled pass-rate weight $w(p) = p^α(1 - p)^β$ derived from the boundary-vanishing structure of distillation gradients. Key results: (1) Theory: We prove that the Beta kernel $w(p) = p^α(1-p)^β$ is a leading-order weight family arising from the SNR structure of distillation, and that it is minimax-robust -- under bounded multiplicative misspecification, worst-case efficiency loss is only $O(δ^2)$. (2)Distillation: On distillation from a larger teacher to a smaller student model with forward KL, Paced achieves significant gain over the base model, while keeping benchmark forgetting at a low level. (3)Self-distillation: On instruction-tuned models with reverse KL, gains are exceeding baselines as well. (4)Two-stage synergy: A forward-KL-then-reverse-KL schedule yields the strongest results in our setting, reaching substantial improvements on standard reasoning benchmarks -- supporting a mode-coverage-then-consolidation interpretation of the distillation process. All configurations require only student rollouts to estimate pass rates, need no architectural changes, and are compatible with any KL direction.

Jelena Markovic-Voronov, Kayhan Behdin, Yuanda Xu et al.

We study the problem of routing queries to large language models (LLMs) under cost, GPU resources, and concurrency constraints. Prior per-query routing methods often fail to control batch-level cost, especially under non-uniform or adversarial batching. To address this, we propose a batch-level, resource-aware routing framework that jointly optimizes model assignment for each batch while respecting cost and model capacity limits. We further introduce a robust variant that accounts for uncertainty in predicted LLM performance, along with an offline instance allocation procedure that balances quality and throughput across multiple models. Experiments on two multi-task LLM benchmarks show that robustness improves accuracy by 1-14% over non-robust counterparts (depending on the performance estimator), batch-level routing outperforms per-query methods by up to 24% under adversarial batching, and optimized instance allocation yields additional gains of up to 3% compared to a non-optimized allocation, all while strictly controlling cost and GPU resource constraints.

Yuanda Xu, Hejian Sang, Zhengze Zhou et al.

Reinforcement Learning with Verifiable Rewards (RLVR) has become the leading paradigm for enhancing reasoning in Large Language Models (LLMs). However, standard RLVR algorithms suffer from a well-documented pathology: while they improve Pass@1 accuracy through sharpened sampling, they simultaneously narrow the model's reasoning boundary and reduce generation diversity. We identify a root cause that existing methods overlook: the uniform penalization of errors. Current approaches -- whether data-filtering methods that select prompts by difficulty, or advantage normalization schemes -- treat all incorrect rollouts within a group identically. We show that this uniformity allows overconfident errors (incorrect reasoning paths that the RL process has spuriously reinforced) to persist and monopolize probability mass, ultimately suppressing valid exploratory trajectories. To address this, we propose the Asymmetric Confidence-aware Error Penalty (ACE). ACE introduces a per-rollout confidence shift metric, c_i = log(pi_theta(y_i|x) / pi_ref(y_i|x)), to dynamically modulate negative advantages. Theoretically, we demonstrate that ACE's gradient can be decomposed into the gradient of a selective regularizer restricted to overconfident errors, plus a well-characterized residual that partially moderates the regularizer's strength. We conduct extensive experiments fine-tuning Qwen2.5-Math-7B, Qwen3-8B-Base, and Llama-3.1-8B-Instruct on the DAPO-Math-17K dataset using GRPO and DAPO within the VERL framework. Evaluated on MATH-500 and AIME 2025, ACE composes seamlessly with existing methods and consistently improves the full Pass@k spectrum across all three model families and benchmarks.

Yuanda Xu, Hejian Sang, Zhengze Zhou et al.

In settings where labeled verifiable training data is the binding constraint, each checked example should be allocated carefully. The standard practice is to use this data directly on the model that will be deployed, for example by running GRPO on the deployment student. We argue that this is often an inefficient allocation because it overlooks a reward-density principle: sparse sequence-level reward should train models where exploration is productive, while dense token-level teacher reward should be used where the aim is to compress behavior into a smaller model. In this view, GRPO-style sparse RL and OPD-style dense teacher supervision are not separate recipes; they are different reward-density regimes. The allocation rule is simple: use scarce labeled training data upstream on the strongest model that can turn it into reward-shaped behavior, then transfer that behavior downstream as dense supervision. We evaluate this rule on verifiable math with Qwen3 and Llama models. At fixed Qwen3-1.7B deployment-student size, an RL-improved 8B teacher distilled through the dense bridge outperforms direct GRPO on the same student, while transfer from the same teacher before RL underperforms. The bridge is important: a forward-KL warmup on teacher rollouts followed by OPD on student rollouts is consistently strongest on MATH before any post-bridge student-side sparse RL, and also gives the best pre-Stage~3 AIME endpoints for the canonical 8B/14B teachers. The bridge also makes later student-side sparse RL effective: GRPO that is weak on a cold student lifts MATH from $75.4\%$ to $78.5\%$ after the bridge and outperforms a matched replay control by $2.8$ points. The operational principal is to avoid using scarce labeled data on the least prepared policy: use sparse reward for teacher-side discovery, dense transfer for student compression, and student-side sparse reward only after the bridge.

Hejian Sang, Yuanda Xu, Zhengze Zhou et al.

Reasoning models think out loud, but much of what they say is noise. We introduce OPSDC (On-Policy Self-Distillation for Reasoning Compression), a method that teaches models to reason more concisely by distilling their own concise behavior back into themselves. The entire approach reduces to one idea: condition the same model on a "be concise" instruction to obtain teacher logits, and minimize per-token reverse KL on the student's own rollouts. No ground-truth answers, no token budgets, no difficulty estimators. Just self-distillation. Yet this simplicity belies surprising sophistication: OPSDC automatically compresses easy problems aggressively while preserving the deliberation needed for hard ones. On Qwen3-8B and Qwen3-14B, we achieve 57-59% token reduction on MATH-500 while improving accuracy by 9-16 points absolute. On AIME 2024, the 14B model gains 10 points with 41% compression. The secret? Much of what reasoning models produce is not just redundant-it is actively harmful, compounding errors with every unnecessary token.

Zhengze Zhou, Giles Hooker, Fei Wang

An increasing number of machine learning models have been deployed in domains with high stakes such as finance and healthcare. Despite their superior performances, many models are black boxes in nature which are hard to explain. There are growing efforts for researchers to develop methods to interpret these black-box models. Post hoc explanations based on perturbations, such as LIME, are widely used approaches to interpret a machine learning model after it has been built. This class of methods has been shown to exhibit large instability, posing serious challenges to the effectiveness of the method itself and harming user trust. In this paper, we propose S-LIME, which utilizes a hypothesis testing framework based on central limit theorem for determining the number of perturbation points needed to guarantee stability of the resulting explanation. Experiments on both simulated and real world data sets are provided to demonstrate the effectiveness of our method.

Zhengze Zhou, Lucas Mentch, Giles Hooker

This paper develops a general framework for analyzing asymptotics of $V$-statistics. Previous literature on limiting distribution mainly focuses on the cases when $n \to \infty$ with fixed kernel size $k$. Under some regularity conditions, we demonstrate asymptotic normality when $k$ grows with $n$ by utilizing existing results for $U$-statistics. The key in our approach lies in a mathematical reduction to $U$-statistics by designing an equivalent kernel for $V$-statistics. We also provide a unified treatment on variance estimation for both $U$- and $V$-statistics by observing connections to existing methods and proposing an empirically more accurate estimator. Ensemble methods such as random forests, where multiple base learners are trained and aggregated for prediction purposes, serve as a running example throughout the paper because they are a natural and flexible application of $V$-statistics.

Xilei Zhao, Zhengze Zhou, Xiang Yan et al.

Machine learning has proved to be very successful for making predictions in travel behavior modeling. However, most machine-learning models have complex model structures and offer little or no explanation as to how they arrive at these predictions. Interpretations about travel behavior models are essential for decision makers to understand travelers' preferences and plan policy interventions accordingly. Therefore, this paper proposes to apply and extend the model distillation approach, a model-agnostic machine-learning interpretation method, to explain how a black-box travel mode choice model makes predictions for the entire population and subpopulations of interest. Model distillation aims at compressing knowledge from a complex model (teacher) into an understandable and interpretable model (student). In particular, the paper integrates model distillation with market segmentation to generate more insights by accounting for heterogeneity. Furthermore, the paper provides a comprehensive comparison of student models with the benchmark model (decision tree) and the teacher model (gradient boosting trees) to quantify the fidelity and accuracy of the students' interpretations.

Zhengze Zhou, Giles Hooker

We propose a modification that corrects for split-improvement variable importance measures in Random Forests and other tree-based methods. These methods have been shown to be biased towards increasing the importance of features with more potential splits. We show that by appropriately incorporating split-improvement as measured on out of sample data, this bias can be corrected yielding better summaries and screening tools.

Yichen Zhou, Zhengze Zhou, Giles Hooker

This paper examines the stability of learned explanations for black-box predictions via model distillation with decision trees. One approach to intelligibility in machine learning is to use an understandable `student' model to mimic the output of an accurate `teacher'. Here, we consider the use of regression trees as a student model, in which nodes of the tree can be used as `explanations' for particular predictions, and the whole structure of the tree can be used as a global representation of the resulting function. However, individual trees are sensitive to the particular data sets used to train them, and an interpretation of a student model may be suspect if small changes in the training data have a large effect on it. In this context, access to outcomes from a teacher helps to stabilize the greedy splitting strategy by generating a much larger corpus of training examples than was originally available. We develop tests to ensure that enough examples are generated at each split so that the same splitting rule would be chosen with high probability were the tree to be re trained. Further, we develop a stopping rule to indicate how deep the tree should be built based on recent results on the variability of Random Forests when these are used as the teacher. We provide concrete examples of these procedures on the CAD-MDD and COMPAS data sets.