AINeutralarXiv – CS AI · Jun 257/10
🧠Researchers discovered that language models forget learned rules midway through training despite continued evidence in data—a phenomenon called 'natural ungrokking.' The survival of rules depends predictably on how often they appear in training data, and attempts to restore forgotten rules through data manipulation fail despite successfully destroying them, revealing asymmetric control over model knowledge.
AIBearisharXiv – CS AI · Jun 57/10
🧠Researchers propose a bilayer SIR epidemic model to analyze how synthetic data contamination spreads across AI systems when models train on each other's outputs. Through theoretical analysis, simulations, and GPT-2 experiments, they demonstrate that cross-contamination can sustain itself (R₀ > 1) and identify detection-based filtering as the most effective intervention strategy.
AIBullisharXiv – CS AI · May 77/10
🧠Researchers develop a theoretical framework explaining how reinforcement learning with verifiable rewards (RLVR) enables long-horizon reasoning in large language models through an implicit curriculum effect. The analysis reveals that mixed-difficulty training naturally progresses from easy to hard problems without explicit scheduling, with learning dynamics determined by the smoothness of the difficulty spectrum.
AINeutralarXiv – CS AI · Mar 67/10
🧠Researchers introduce Non-Classical Network (NCnet), a classical neural architecture that exhibits quantum-like statistical behaviors through gradient competitions between neurons. The study reveals that multi-task neural networks can develop non-local correlations without explicit communication, providing new insights into deep learning training dynamics.
AINeutralarXiv – CS AI · Jun 236/10
🧠Researchers investigate emergent misalignment (EM) in AI models, where narrow fine-tuning causes broad but uneven misalignment across evaluations. Through analysis of training dynamics, model priors, and data, they find that model architecture priors partially predict misalignment outcomes, learning schedules show limited influence on alignment improvement, and activation patterns between training and evaluation reveal significant overlap that correlates with misalignment propagation.
AINeutralarXiv – CS AI · Jun 236/10
🧠Researchers demonstrate that visual shortcuts in vision-language models trained with reinforcement learning emerge sharply and can be controlled through regularization strength. The study reveals a critical intervention window where penalties applied early prevent shortcut formation, but the same penalties become less effective after the model has consolidated these shortcuts.
AINeutralarXiv – CS AI · Jun 116/10
🧠Researchers introduce Velocity-Regularized Adam (VRAdam), a physics-inspired optimizer that improves deep neural network training by adding velocity-based regularization to prevent oscillations and instability. VRAdam demonstrates superior performance compared to standard optimizers like AdamW across multiple benchmarks including image classification, language modeling, and generative modeling tasks.
AINeutralarXiv – CS AI · Jun 86/10
🧠Researchers characterize the training dynamics of on-policy distillation (OPD), a technique used to improve large language model reasoning, revealing it operates in a distinct geometric regime compared to supervised fine-tuning and reinforcement learning. The study shows OPD exhibits 'subspace locking,' where cumulative updates rapidly converge to a narrow low-dimensional channel that is functionally sufficient for performance, suggesting OPD has unique training dynamics rather than existing as a simple intermediate between other training approaches.
AINeutralarXiv – CS AI · Jun 56/10
🧠Researchers propose a continuous-time mathematical model for analyzing gradient descent dynamics in the Edge of Stability regime, where large learning rates cause oscillations in neural network training. The model introduces an effective free energy framework that combines risk with a curvature-related term, enabling better prediction of training dynamics in wide two-layer networks and validated on matrix factorization and CIFAR-10 tasks.
AINeutralarXiv – CS AI · Jun 56/10
🧠Researchers introduce Double Preconditioning (DoPr), a new optimization technique that improves neural network performance during real-world deployment by combining gradient-wise and activation-wise preconditioning. The method addresses test-time feedback—the gap between training metrics and actual task performance in autoregressive models—without requiring improvements in traditional validation loss metrics.
AINeutralarXiv – CS AI · Jun 26/10
🧠Researchers tracked how attention-head circuits form during training across three 1B-parameter language models, revealing that induction circuits and attention-sink circuits emerge as separate phenomena separated by an order of magnitude in training tokens. The study identifies architectural properties (zero BOS-heads in early layers) and demonstrates that circuit identification requires only 0.3-2% of total training data, offering insights into mechanistic interpretability of transformer models.
AINeutralarXiv – CS AI · Jun 26/10
🧠Researchers discovered that continuous-time RNNs trained with noise injected inside activation functions paradoxically perform best when noise remains present at test time, contradicting conventional assumptions about noise removal. This phenomenon stems from noise-induced shifts in neural network dynamics that become computationally integrated into learned representations, revealing that networks can overfit to training noise itself rather than just input-output mappings.
AIBullisharXiv – CS AI · Jun 16/10
🧠Researchers propose a novel method called Signed Entropy Integral (SEI) to detect mislabeled images in training datasets by analyzing how prediction entropy changes during model training. The technique shows that correctly labeled samples exhibit consistent entropy decrease while mislabeled ones maintain high entropy, achieving state-of-the-art performance on medical imaging datasets.
AINeutralarXiv – CS AI · Jun 16/10
🧠Researchers demonstrate that modestly-sized open-source language models can understand rare paired-focus constructions (like "let alone" and "much less"), challenging assumptions that only the largest LLMs grasp complex constructional semantics. The study reveals that semantic understanding of these constructions emerges later in training than syntactic knowledge and correlates with world knowledge acquisition.
AINeutralarXiv – CS AI · May 296/10
🧠Researchers identify a consistent three-regime structure in scientific machine learning (SciML) models, demonstrating that neural networks exhibit distinct failure modes and training behaviors depending on hyperparameter settings. The study reveals that optimization methods are regime-specific with no universal solution, providing a diagnostic framework to improve model robustness across physics-informed neural networks, neural operators, and neural ODEs.
AINeutralarXiv – CS AI · May 276/10
🧠Researchers propose a representation-readout decomposition framework that explains anomalous neural network training phenomena like grokking and double descent by analyzing two competing learning processes: representation learning in encoders and readout calibration in classifiers. The framework provides task-agnostic diagnostics that reveal these phenomena stem from fluctuations in relative learning speeds rather than mysterious delays, challenging existing lazy-to-rich learning theories.
AINeutralarXiv – CS AI · May 76/10
🧠Researchers identify a critical training window where Transformer models decide between memorization and reasoning, finding that applying weight decay during a specific 25% training phase matches full-training performance on compositional tasks. The discovery reveals sharp boundaries in this decision point, with timing shifts of just 100 optimization steps causing dramatic accuracy swings from chance performance to robust reasoning.
AINeutralarXiv – CS AI · Apr 156/10
🧠Researchers investigate on-policy distillation (OPD) dynamics in large language model training, identifying two critical success conditions: compatible thinking patterns between student and teacher models, and genuine new capabilities from the teacher. The study reveals that successful OPD relies on token-level alignment and proposes recovery strategies for failing distillation scenarios.
