All Articles
17198 articles total
IB-Flow: Information Bottleneck-Guided CFG Distillation for Few-Step Text-to-Image Generation
IB-Flow, a framework using the Information Bottleneck principle to achieve high-fidelity 2-step text-to-image generation by reducing inference latency.
The Patchwork Problem in LLM-Generated Code
Researchers identify the 'patchwork problem' in LLM-generated code, where locally correct code snippets cause global structural failures in the overall repository.
SCATE: Learning to Supervise Coding Agents for Cost-Effective Test Generation
SCATE is introduced as a framework to automate the supervision of coding agents for more cost-effective and comprehensive test generation, reducing reliance on humans.
AlphaZero in Sparsely Rewarded Games: Limits and Auxiliary Supervision
A study on AlphaZero's limitations in sparsely rewarded games suggests that auxiliary supervision (AZAL) is needed to achieve perfect play.
Model Agnostic Graph Prompt Learning for Crystal Property Prediction
A new multilevel graph prompt learning framework improves crystal property prediction by capturing latent chemical and structural features.
Correlation-Aware Contextual Bandits with Surrogate Rewards for LLM Routing
New algorithms for contextual bandits with surrogate rewards are proposed to improve the efficiency and cost-effectiveness of LLM routing.
Phone Segmentation and Recognition through Phonological Activation Mapping
The SPAM method leverages self-supervised speech models for simultaneous phone segmentation and recognition with minimal phonetic transcription.
Video Generation Models are General-Purpose Vision Learners
GenCeption demonstrates that pre-trained video generative models can serve as general-purpose vision learners for tasks like depth and pose estimation.
Evolutionary Intelligence for Scientific Discovery: From Evolutionary Computation to Cumulative Discovery Systems
This review proposes 'Evolutionary Intelligence' (EI) to transition scientific discovery from simple candidate refinement to cumulative discovery systems.
Quantum Logic as the Logic of Contexts
A mathematical exploration of quantum logic as a contextual calculus, arguing that classical logic is a simplified, information-losing image of it.
On Locality and Length Generalization in Visual Reasoning
Research suggests that local, sequential vision models (foveated glimpses) generalize better on complex visual reasoning tasks than global image models.
Backtrack-Free Cursive
A discussion on a technique for backtrack-free cursive writing.
The social physics of conversation: Communication patterns matter
An exploration of communication patterns and the 'social physics' of conversation.
Sam Neill has died
News regarding the passing of Sam Neill.
TheBioCollection: Unified Pre-Training Scale LLM Corpus for Biology
Introduction of TheBioCollection, a 52.6B-token pre-training corpus for biology-focused LLMs, along with a corresponding evaluation suite.
Prompt-Driven Exploration
Proposed Prompt-Driven Exploration (PDE) for Reinforcement Learning, using VLMs to refine prompts based on rollout videos to improve sample efficiency.
A Novel Parallel QCNN Architecture with Efficient Classical Simulability
A new parallel QCNN architecture that allows 128-qubit models to be simulated efficiently on classical machines through hierarchical partitioning.
Eluna: An Agentic LLM System for Automating Warehouse Operations with Reasoning and Task Execution
Eluna, a graph-guided multi-agent framework for automating warehouse operations via SOP execution and asymmetric episodic distillation.
NL-PAC: Specification Ambiguity and Certified Minimax Risk Floors in LLM-Mediated Supervision
Introduction of NL-PAC, a framework for certifying minimax risk floors in LLM-mediated supervision to address specification ambiguity.
MultiView-Bench: A Diagnostic Benchmark for World-Centric Multi-View Integration in VLMs
MultiView-Bench, a diagnostic benchmark for evaluating 3D world-centric integration in VLMs, and ViewNavigator, a multi-agent framework for improving multi-view perception.