9517 papers
This paper investigates the phenomenon of subliminal learning in language model distillation, revealing that hidden bias transfer occurs under hard distillation not through global entanglement, but via a small set of critical "divergence tokens" processed in early layers, making the effect both mechanistically specific and fragile to prompt variations.
BeyondBench introduces a contamination-resistant evaluation framework that uses on-the-fly algorithmic problem generation to assess the true reasoning capabilities of 101 language models across 44 tasks, revealing significant performance gaps in complex problem-solving and the critical role of tool usage.
The paper proposes SphereAR, a hyperspherical latent-based autoregressive model that constrains inputs and outputs to a fixed-radius hypersphere to eliminate variance collapse and achieve state-of-the-art image generation performance, surpassing diffusion and masked-generation models at comparable scales.
This paper establishes explicit upper bounds on the quadratic Wasserstein distance between trained single-layer neural networks and their Gaussian process limits, demonstrating that the approximation error decays polynomially with network width while accounting for the influence of architectural parameters and training dynamics.
This paper introduces a novel NVFP4 training framework that combines Random Hadamard transforms, 2D quantization, stochastic rounding, and selective high-precision layers to successfully pretrain a 12-billion-parameter model on 10 trillion tokens with performance comparable to FP8 baselines, thereby demonstrating the viability of stable 4-bit precision training for large language models.
OPPO is a lightweight, model-agnostic framework that accelerates PPO-based RLHF training by 1.8× to 2.8× and improves GPU utilization by overlapping pipeline execution through intra-step streaming and inter-step adaptive overcommitment to mitigate sequential dependencies and long-tail latency.
VidGuard-R1 is a novel AI-generated video detector that leverages Group Relative Policy Optimization (GRPO) and specialized reward models to overcome the limitations of static datasets, achieving state-of-the-art zero-shot performance while providing human-interpretable, physics-grounded explanations for its forensic judgments.
This paper establishes non-asymptotic bounds on the efficiency of conformalized regression trained via SGD, revealing how prediction set length depends on training and calibration set sizes as well as the miscoverage level, and identifying phase transitions that guide optimal data allocation.
This paper demonstrates that simply widening neural network models, combined with suitable softmax temperature calibration, is sufficient to achieve linear mode connectivity without the need for parameter permutations, a phenomenon explained by the layerwise exponentially weighted connectivity (LEWC) property where merged layer outputs act as exponentially weighted sums of the original models' outputs.
This paper introduces XFactor, the first geometry-free self-supervised model that achieves true novel view synthesis by disentangling camera pose from scene content through a novel augmentation scheme, thereby enabling pose transferability across different 3D scenes without relying on explicit 3D inductive biases.
This paper introduces CBF-RL, a framework that integrates Control Barrier Functions directly into the reinforcement learning training process to internalize safety constraints, thereby enabling safe, robust, and filter-free deployment of policies in real-world scenarios like humanoid robot navigation.
This paper introduces B-ODIL, a Bayesian extension of the Optimization of a Discrete Loss (ODIL) method that integrates PDE-based prior knowledge with data likelihood to solve inverse problems with quantified uncertainties, demonstrating its effectiveness through synthetic benchmarks and a clinical application for estimating brain tumor concentration from MRI scans.
The paper introduces Schrödinger Bridge Mamba (SBM), a novel one-step speech enhancement model that synergizes the Schrödinger Bridge training paradigm with the Mamba architecture to achieve superior denoising and dereverberation performance with real-time feasibility, outperforming strong generative and discriminative baselines.
This paper demonstrates that existing alignment-based defenses against control-flow hijacking in multi-agent systems are vulnerable to evasion due to inherent safety-functionality conflicts and limited context visibility, and proposes ControlValve, a new defense mechanism that enforces control-flow integrity and least privilege through permitted control-flow graphs and contextual rules.
This paper theoretically demonstrates that in overparameterized two-layer ReLU networks trained below the edge of stability, data geometry dictates generalization by determining whether gradient descent learns shared patterns (for data hard to shatter) or favors memorization (for data easily shattered).
This paper introduces a principled framework for testing excessive influence in linear least-squares models by deriving exact influence formulas and their corresponding extreme value distributions, thereby enabling rigorous hypothesis tests to distinguish between natural sampling variation and problematic data subsets across various scientific domains.
LLEMA is a unified framework that integrates large language models with chemistry-informed evolutionary rules and memory-based refinement to efficiently discover chemically plausible, stable, and multi-objective optimized materials, outperforming existing generative and LLM-only baselines across diverse application domains.
This paper proposes an auto-adaptive sampling method for Physics Informed Neural Networks (PINNs) that utilizes arbitrary problem-specific heuristics to accurately resolve interfacial regions in Allen-Cahn equations without post-hoc resampling, demonstrating superior performance over residual-adaptive frameworks.
The paper introduces FMint-SDE, a multimodal foundation model based on decoder-only transformers that leverages in-context learning and error correction to provide a general-purpose, efficient, and accurate solution for simulating stochastic differential equations across diverse scientific domains.
MotionStream is a real-time video generation framework that distills a bidirectional teacher model into a causal student using Self Forcing with Distribution Matching Distillation and sliding-window attention with attention sinks, enabling sub-second, infinite-length streaming generation with interactive motion controls on a single GPU.