learning-rule comparison

This work addresses the fundamental problem of how task structure, learning rules, and input distribution jointly govern learning efficiency in nonlinear perceptrons—moving beyond conventional student–teacher frameworks and linear-output assumptions. We develop a stochastic-process-based flow equation that unifies the dynamical evolution of both supervised learning (SL) and reinforcement learning (RL). For the first time in nonlinear perceptrons, we quantitatively characterize the differential regulatory role of input noise: SL exhibits greater robustness to input noise, whereas RL suffers accelerated forgetting and slower task coverage. These theoretical predictions are empirically validated on MNIST, precisely reproducing learning/forgetting curves and cross-task interference magnitudes. Our framework establishes a novel, testable paradigm for analyzing nonlinear learning dynamics in both biological and artificial neural networks, providing a rigorous quantitative foundation for comparative analysis across learning paradigms.

Current AI research often treats models as static artifacts, overlooking the fundamental influence of training dynamics on critical properties such as capability, bias, robustness, and safety. This work proposes shifting the focus toward the training process itself to establish a science of AI centered on training dynamics. By analyzing the interactions among data, objectives, architectures, and optimizers, the paper develops a theoretical framework that is predictive, intervenable, and design-oriented. Integrating approaches from mechanistic interpretability, fairness, memory mechanisms, and simplicity biases, it uncovers causal links between early-training signals and final model behavior. The study systematically outlines key challenges and open problems, offering both theoretical pathways and practical foundations for extending scaling laws beyond performance to encompass multidimensional model attributes.

This study addresses the challenge of attributing performance changes in adaptive artificial intelligence (AI) medical devices during continuous iteration. To disentangle the effects of intrinsic model improvements from those induced by dynamic external environments, this work proposes a three-dimensional evaluation framework—comprising learning capacity, data-driven potential, and knowledge retention—and, for the first time, decouples these dimensions into independent metrics. Through case studies simulating population distribution shifts, combined with dynamic performance tracking and knowledge retention measurements, the framework demonstrates that adaptive AI systems can balance learning and stability under gradual data evolution, while exhibiting a trade-off between plasticity and stability in abrupt change scenarios. This approach enables fine-grained, regulatory-grade assessment of the evolutionary trajectory of adaptive AI systems.

Precisely encoding nuanced, domain-specific policies into prompt instructions for large language models (LLMs) remains a fundamental challenge. Method: This paper proposes a data–model co-evolution paradigm featuring an iterative, human-feedback-driven closed loop that simultaneously expands the test set dynamically and refines prompt instructions—integrated with structured human–AI collaboration, rationale-based behavioral attribution analysis, and iterative instruction evaluation. Contribution/Results: The approach mechanizes the concretization of ambiguous policies, systematically uncovers edge cases, and strengthens policy verifiability. A user study demonstrates that the framework significantly improves the systematicity and consistency of instruction refinement, while enhancing LLM adherence to localized policies and complex semantic rules.

Modern deep learning training pipelines involve heterogeneous components—such as multi-task heads, distillation objectives, or multimodal encoders—yet lack a unified formalism for modeling inter-component dependencies and enabling holistic optimization. Method: We propose *Learning Diagrams*, the first categorical framework for declaratively specifying training workflows as composable, compiler-ready graphical structures. It enables constraint-driven component composition and automatic synthesis of joint loss functions that enforce predictive consistency across submodels. Contribution/Results: Learning Diagrams unifies diverse paradigms—including few-shot multi-task learning, knowledge distillation, and multimodal learning—under a single abstraction, supporting dynamic in-training and post-hoc reconfiguration. Integrated with PyTorch and Flux.jl, our open-source implementation includes a graph compiler and demonstrates cross-paradigm expressivity and compositional modeling efficacy across canonical benchmarks. Empirical results show substantial improvements in systematicity, modularity, and interpretability of complex model construction.

This study investigates the mechanisms by which successive post-training stages—continued pretraining (CPT), supervised fine-tuning (SFT), and reinforcement learning (RL)—affect the generalization capabilities of biomedical reasoning models. Through systematic training and evaluation of over 100 models across genomic, transcriptomic, and protein-related tasks, the authors employ a controlled ablation approach to dissect the heterogeneous and stage-specific impacts of each phase on in-distribution (ID) and out-of-distribution (OOD) performance. The findings reveal that CPT enhances alignment with biological language structures, SFT improves ID performance at the cost of OOD generalization, and RL can recover OOD robustness when applied following strong SFT. The optimal strategy combines brief SFT, extended RL, and asymmetric stage capacities, effectively balancing ID accuracy and OOD generalization.

This study addresses the observation that the generalization capability of language models during pretraining does not improve monotonically, but instead oscillates frequently between rote memorization and intelligent reasoning. To investigate this, we construct an evaluation suite to identify and define the "mode jumping" phenomenon, modeling it as a circuit competition problem under capacity constraints. We propose a theoretical framework for capacity allocation, wherein data windows govern circuit competition, and integrate intermediate checkpoint selection with pretraining data selection strategies to monitor and control generalization dynamics. Our findings challenge the conventional assumption of stable model maturation by demonstrating that intermediate checkpoints can exhibit superior reasoning and alignment capabilities compared to the final model. Furthermore, we show that strategic data selection effectively stabilizes the generalization process throughout pretraining.

This work addresses the limitations of evaluating language model training efficiency solely through end-point metrics under compute-constrained settings, which often overlooks instability and negative returns during training. The authors propose a multidimensional evaluation paradigm based on training trajectories, conducting repeated-measures experiments within a fixed token budget to track the training dynamics of a 4.26M-parameter Llama-style model on the TinyStories corpus. Combining repeated-measures ANOVA with interval-level telemetry analysis, they observe rapid early convergence—validation loss drops from 8.36 to 2.80 within approximately 4M tokens—followed by non-monotonic degradation, with loss rebounding to 3.90 and no discernible stable training phase. These findings suggest that indiscriminately increasing training tokens may be counterproductive, offering a novel perspective for efficient low-resource training evaluation.

研究通过构建生成事实图（GFG）并结合nanoGPT，统一了神经系统的训练、学习和推理过程，实现对功能支持重组的精确控制与预测。

研究在概念漂移、预算和延迟约束下，通过比较不同模型更新策略（定期重训、错误阈值触发、统计漂移触发）来解决生产机器学习系统性能下降问题。