2323 papers
This paper employs mechanistic interpretability to reveal that knowledge distillation not only compresses teacher models into smaller students but also fundamentally restructures their internal circuits by reorganizing, discarding, and relying more heavily on fewer components, necessitating new metrics to quantify these internal functional shifts beyond mere output similarity.
This paper introduces Ready2Unlearn, a proactive, model-agnostic training-time optimization approach that leverages meta-learning principles to prepare machine learning models for efficient and principled future unlearning, shifting the focus from reactive post-deployment algorithms to forward-looking readiness.
To address the data scarcity in dexterous manipulation imitation learning, this paper introduces EgoDex, the largest and most diverse dataset of its kind featuring 829 hours of Apple Vision Pro-captured egocentric videos with precise, native 3D hand and finger tracking, alongside established benchmarks for training and evaluating manipulation policies.
FreeKV is a training-free framework that combines speculative retrieval, fine-grained correction, and hybrid CPU-GPU memory management to significantly accelerate KV cache retrieval for large language models, achieving up to a 13× speedup over state-of-the-art methods while maintaining near-lossless accuracy.
This paper introduces the first provably convergent online algorithms for decision-focused learning in dynamic environments by regularizing non-differentiable objectives and employing perturbation techniques to handle non-convexity, thereby establishing static and dynamic regret bounds and demonstrating superior performance over standard benchmarks.
Vid2World is a general framework that transforms pre-trained video diffusion models into interactive world models by implementing causalization techniques and a causal action guidance mechanism to enable high-fidelity, controllable, and autoregressive future prediction across diverse domains.
MAS-ZERO is a novel, self-evolved inference-time framework that automatically designs, critiques, and refines multi-agent system configurations for specific tasks without requiring a validation set, achieving significant performance improvements over manual and existing automatic baselines across reasoning, coding, and agentic benchmarks.
The paper proposes HDLxGraph, a novel framework that integrates Abstract Syntax Trees and Data Flow Graphs into Retrieval Augmented Generation to overcome structural and vocabulary mismatches in Hardware Description Language tasks, while also introducing the HDLSearch benchmark to demonstrate significant improvements in search, debugging, and code completion accuracy over existing baselines.
This paper introduces WikiDBGraph, a large-scale benchmark suite derived from 100,000 real-world relational databases, designed to evaluate and expose the limitations of existing collaborative learning frameworks in handling the complex, unaligned, and interconnected nature of practical data silos.
This paper introduces AgarCL, a research platform based on the non-episodic game Agar.io designed to advance continual reinforcement learning by providing a complex, dynamic environment where standard algorithms and existing continual learning methods face significant challenges beyond the traditional stability-plasticity dilemma.
This paper introduces X-MethaneWet, the first cross-scale global wetland methane benchmark dataset combining physics-based simulations and real-world observations, and demonstrates how deep learning models enhanced by transfer learning can significantly improve methane flux prediction and climate modeling.
This paper establishes a maximum principle and the Hamilton-Jacobi-Bellman equation for optimal control on infinite-dimensional probability distribution spaces, and leverages these theoretical results to develop a scalable deep learning algorithm for solving high-dimensional multi-agent control problems.
The paper introduces VISTA, a novel training-free framework that leverages Vision-Language Models to predict stock prices by jointly analyzing textual data and line charts through zero-shot prompting, achieving significant performance improvements over traditional statistical and text-only baselines.
This paper proposes a novel defense against prompt injection attacks in large language models by augmenting intermediate token representations with layer-specific trainable embeddings to enforce instruction hierarchy, achieving a 1.6x to 9.2x reduction in attack success rates compared to state-of-the-art methods without compromising model utility.
This paper proposes Orthogonal Common Neighbor (OCN), a novel link prediction method that addresses redundancy and over-smoothing in higher-order common neighbors through orthogonalization and normalization, achieving significant performance improvements over state-of-the-art baselines.
The paper introduces ViTaPEs, a transformer-based architecture that employs a novel two-stage positional encoding strategy to effectively fuse visual and tactile modalities, achieving state-of-the-art performance and zero-shot generalization across diverse recognition and robotic grasping tasks without relying on pre-trained vision-language models.
The paper introduces LoFT, a novel parameter-efficient fine-tuning method that aligns optimizer dynamics (momentum and variance) with full fine-tuning within a low-rank subspace, thereby eliminating the need for hyperparameter tuning and achieving performance comparable to full fine-tuning without increasing inference costs.
This paper introduces Progressive Neural Collapse (ProNC), a novel continual learning framework that overcomes the limitations of fixed global ETF targets by progressively expanding the simplex equiangular tight frame with new class prototypes, thereby effectively mitigating catastrophic forgetting while maintaining flexibility and efficiency.
This paper proposes a novel, lightweight, and plug-and-play adaptive correction method that utilizes a learnable operator to enforce strict conservation laws in neural operators, thereby significantly improving their accuracy, stability, and flexibility compared to existing constraint-based approaches.
ActivePusher is a novel framework that enhances data efficiency and planning reliability in nonprehensile manipulation by combining residual physics modeling with uncertainty-based active learning to prioritize informative data collection and guide control sampling toward more reliable actions.