Searcharxiv⌕ Search

arXiv subjects

Zhenxing Dou

Publications and source records attributed to Zhenxing Dou.

5 recordsLinked to original sources

Rethinking Supervision Scaling in On-Policy Distillation: From Dense Supervision to Stochastic Horizons

On-policy distillation (OPD) provides dense token-level supervision along student-generated trajectories, yet the learning benefits of longer trajectories may not scale proportionally with their computational cost. We investigate how supervision utility evolves with trajectory length. Controlled experiments reveal diminishing returns from additional supervised tokens, while showing that later positions remain learnable and exhibit substantial cross-position supervision transfer. Gradient analysis further identifies shared update components across response positions and diverse position-specific residuals whose cross-rollout consistency decreases at greater depths. Motivated by these findings, we propose Stochastic Horizons On-Policy Distillation (SH-OPD), which reduces redundant supervision by randomly varying rollout horizons while preserving access to deeper response states. SH-OPD samples a shared horizon for each training batch, enabling efficient parallel generation without per-sequence horizon scheduling overhead. We characterize the resulting reweighting bias and derive a compute-reweighting trade-off under standard optimization assumptions. Experiments on mathematical reasoning and code generation demonstrate that SH-OPD achieves slightly higher mathematical reasoning accuracy at approximately 45% of the per-update cost of standard OPD, while maintaining competitive code generation performance at around 60% of the baseline cost. These results highlight the benefits of efficient supervision allocation across response depths.

cs.LG↗

SRA: Selective Residual Adaptation for LLM Reasoning Enhancement

Knowledge distillation transfers reasoning capabilities from large language models to lightweight students. However, distilled small models still struggle with multi-step reasoning. We observe that the distribution divergences between teacher and student models along reasoning paths are highly sparse, and target tokens almost always fall within the student top-$K$ candidate set. Based on this finding, unlike approaches updating student backbones or requiring runtime teacher computation, we propose Selective Residual Adaptation (SRA), a plug-and-play framework that internalizes localized corrections. Concretely, SRA employs a Discrepancy Gate to identify divergence positions alongside a Candidate Residual Adapter for top-$K$ logit calibration to correct the mismatch. Theoretically, we establish error bounds ensuring strict sequence level performance gains. Across mathematical benchmarks on Qwen3 models from 0.6B to 4B, SRA consistently enhances distilled student models, raising average Pass@16 by up to 4.37 percentage points while preserving general capabilities.

cs.LG↗

Beyond the Black Box: A Survey on the Theory and Mechanism of Large Language Models

The rapid emergence of Large Language Models (LLMs) has precipitated a profound paradigm shift in Artificial Intelligence, delivering monumental engineering successes that increasingly impact modern society. However, a critical paradox persists within the current field: despite the empirical efficacy, our theoretical understanding of LLMs remains disproportionately nascent, forcing these systems to be treated largely as ``black boxes''. To address this theoretical fragmentation, this survey proposes a unified lifecycle-based taxonomy that organizes the research landscape into six distinct stages: Data Preparation, Model Preparation, Training, Alignment, Inference, and Evaluation. Within this framework, we provide a systematic review of the foundational theories and internal mechanisms driving LLM performance. Specifically, we analyze core theoretical issues such as the mathematical justification for data mixtures, the representational limits of various architectures, and the optimization dynamics of alignment algorithms. Moving beyond current best practices, we identify critical frontier challenges, including the theoretical limits of synthetic data self-improvement, the mathematical bounds of safety guarantees, and the mechanistic origins of emergent intelligence. By connecting empirical observations with rigorous scientific inquiry, this work provides a structured roadmap for transitioning LLM development from engineering heuristics toward a principled scientific discipline.

cs.CL↗

PRIME: Physics-Related Intelligent Mixture of Experts for Transistor Characteristics Prediction

In recent years, machine learning has been extensively applied to data prediction during process ramp-up, with a particular focus on transistor characteristics for circuit design and manufacture. However, capturing the nonlinear current response across multiple operating regions remains a challenge for neural networks. To address such challenge, a novel machine learning framework, PRIME (Physics-Related Intelligent Mixture of Experts), is proposed to capture and integrate complex regional characteristics. In essence, our framework incorporates physics-based knowledge with data-driven intelligence. By leveraging a dynamic weighting mechanism in its gating network, PRIME adaptively activates the suitable expert model based on distinct input data features. Extensive evaluations are conducted on various gate-all-around (GAA) structures to examine the effectiveness of PRIME and considerable improvements (60\%-84\%) in prediction accuracy are shown over state-of-the-art models.

cs.LG↗

Optimal subsampling for large scale Elastic-net regression

Datasets with sheer volume have been generated from fields including computer vision, medical imageology, and astronomy whose large-scale and high-dimensional properties hamper the implementation of classical statistical models. To tackle the computational challenges, one of the efficient approaches is subsampling which draws subsamples from the original large datasets according to a carefully-design task-specific probability distribution to form an informative sketch. The computation cost is reduced by applying the original algorithm to the substantially smaller sketch. Previous studies associated with subsampling focused on non-regularized regression from the computational efficiency and theoretical guarantee perspectives, such as ordinary least square regression and logistic regression. In this article, we introduce a randomized algorithm under the subsampling scheme for the Elastic-net regression which gives novel insights into L1-norm regularized regression problem. To effectively conduct consistency analysis, a smooth approximation technique based on alpha absolute function is firstly employed and theoretically verified. The concentration bounds and asymptotic normality for the proposed randomized algorithm are then established under mild conditions. Moreover, an optimal subsampling probability is constructed according to A-optimality. The effectiveness of the proposed algorithm is demonstrated upon synthetic and real data datasets.

math.ST↗