All Articles
16580 articles total
Kernelized Linear Attention: Breaking the Capacity Wall with Symmetric Cones
Kernelized Linear Attention (KATA) breaks the capacity wall of linear attention using symmetric cones, achieving significantly higher throughput than FlashAttention-2 on long sequences.
HyCoRec: Hypergraph-Enhanced Multi-Preference Learning for Alleviating Matthew Effect in Conversational Recommendation
HyCoRec is a hypergraph-enhanced learning paradigm designed to alleviate the 'Matthew Effect' in conversational recommendation systems by learning multi-aspect preferences.
Multilingual Sentence Embeddings for Linguistic-Integrated Reliability Audit
A study evaluating whether multilingual sentence embeddings can replace translation-based processes for reliability audits in linguistic assessments.
DecoyFace: Beyond Obfuscation via Controllable and Imperceptible Identity Misdirection for Privacy-Preserving Face Recognition
DecoyFace is a privacy-preserving face recognition framework that misdirects unauthorized reconstruction attempts toward plausible but incorrect identities.
Retrieval-Augmented Interpretable Learning: Towards Task-Specific Zero-Shot Models in Healthcare
RAIL is a probabilistic meta-learning framework for healthcare that enables zero-shot generation of interpretable models for clinical procedure prediction.
LG to Ban Residential Proxies from Smart TV Apps
LG is implementing a ban on residential proxies within its Smart TV applications to prevent unauthorized access and abuse.
The Anthropic-Physical Intelligence rumor roiling AI Twitter
Rumors are circulating on AI Twitter regarding potential acquisition sprees by Anthropic and OpenAI, specifically involving Physical Intelligence.
Debate-on-Graph: Reliable and Adaptive Reasoning of Large Language Model on Uncertain Knowledge Graph
Researchers introduce Debate-on-Graph (DoG), a framework that combines LLMs and Uncertain Knowledge Graphs to reduce hallucinations through a multi-agent debate mechanism.
Between Safe Boundaries: Exploiting Temporal Consistency for Jailbreaking Text-To-Video Generation Models
The BSB framework explores temporal consistency in text-to-video generation models to create more efficient and structured jailbreak attacks.
AIGB-R1: Self-Evolving Generative Auto-Bidding via Hierarchical Planner-Executor Optimization
AIGB-R1 is a hierarchical self-evolving auto-bidding framework that uses LLMs to optimize online advertising bids through a planner-executor architecture.
SAGA: Synthetic Agentic Graph Architecture for Temporal Benchmark Generation
SAGA is a system for generating large-scale, semantically rich temporal graph benchmarks for training graph neural networks using a four-phase pipeline.
Lookahead Branching for Neural Network Verification
This research explores lookahead branching strategies to accelerate neural network verification in branch-and-bound verifiers like Marabou and alpha-beta-CROWN.
WAR: Workload-Aware Rollouts for Synchronous Agentic Reinforcement Learning
WAR is a workload-aware rollout system designed to accelerate synchronous agentic RL by optimizing decoding and scheduling to reduce KV-cache recomputation.
The Optimization Trilemma: Efficiency, Comfort and Fairness in Decentralized Multi-agent Coordination
This paper proposes a model for decentralized multi-agent coordination that balances system-wide efficiency, individual agent comfort, and fairness.
TAPAS: Throughput-adaptive Perception for Autonomous Systems
TAPAS is a throughput-adaptive perception strategy for mobile/edge platforms that uses RL to dynamically allocate resources and save energy in autonomous systems.
The Miso That Went to Space
A Hacker News discussion regarding a piece of miso that went to space.
VLA-ReID: Video-Level Association for Re-Identification in Multi-Object Tracking with Highly Similar Objects
VLA-ReID is proposed to improve multi-object tracking for visually similar objects by reformulating re-identification as video-level association.
DADIR: Density-Aware Data-level Imbalanced Regression Framework
DADIR is a new framework for imbalanced regression that uses density-aware partitioning and a conditional VAE to handle underrepresented data regions.
Talaria: Session-Aware Serverless Serving of Hundred-Billion-Parameter LLMs
Talaria is a session-aware serverless serving system designed to optimize the serving of 100B+ parameter LLMs by improving session continuity and KV locality.
Auditing Question-Order Effects in Large Language Models with the QQ Equality: Mechanism Characterization and a Saturation Caveat
Researchers audit question-order effects in LLMs using the QQ equality, highlighting a saturation caveat when using next-token log-probabilities.