[research]By ByteBulletin Editor
Nexus: KV-Cache Routing to Slash LLM Inference Costs
New research introduces a routing mechanism that distributes key-value cache storage across machines to cut memory overhead and latency in large-scale LLM serving.
[research]
Research
127 stories · page 3
[research]By ByteBulletin Editor
Prime Agent Orchestrator: A New Framework for Coordinating AI Agents
Researchers introduce Prime, an open-source framework for orchestrating multiple AI agents with a focus on reliability and developer control.
[research]By ByteBulletin Editor
D-Wave Entangles Its Dual-Rail Qubits, Preserving the Error Hierarchy
A Nature paper shows two dual-rail qubits can be entangled quickly without disturbing the dominant photon-loss error, a key step toward simpler error correction.
[research]By ByteBulletin Editor
Aegis: A Runtime Governance Layer for AI Agents
A new research framework proposes Aegis, a runtime governance layer that lets developers enforce policies on AI agents in real time.
[research]By ByteBulletin Editor
New Research Framework Aims to Pinpoint AI's Exact Energy Footprint
A proposed arXiv initiative seeks to standardize how developers measure the true energy cost of training and running large language models.
[research]By ByteBulletin Editor
SpaceX Unveils Plan to Build a Terrestrial Network to Take on Major US Carriers
On its first earnings call, SpaceX outlined a strategy to use spectrum acquired from EchoStar and a network of rooftop Starlink-base-station hybrids to offer direct mobile service, aiming to poach customers from T-Mobile, AT&T, and Verizon.
[research]By ByteBulletin Editor
Anthropic's Older Claude Models Are Surprisingly Easy to Jailbreak for Smut
A simple multi-turn technique bypasses safeguards on Opus 4.6, Opus 3, and Haiku 4.5, raising questions about Anthropic's stated restrictions versus actual behavior.
[research]By ByteBulletin Editor
Boundary-Bench measures how much coding agents lose when you lock down the sandbox
A new open-source benchmark runs coding agents in hardened, enterprise-style sandboxes to quantify the capability cost of security controls.
[research]By ByteBulletin Editor
Woodpecker Distillation: A Faster Path to Multimodal Reasoning
A new distillation method compresses large multimodal models into smaller ones that reason faster, without sacrificing accuracy.
[research]By ByteBulletin Editor
The Quiet Revolution in LLM Data Selection: Curating Quality Over Quantity
New research highlights how smarter data selection — not just more data — is becoming the key lever for training more capable and efficient language models.
[research]By ByteBulletin Editor
PPDL: A New Deep Learning Language for Probabilistic Programs
Researchers introduce a programming language that brings neural networks into probabilistic programming, promising more expressive and scalable Bayesian models.
[research]By ByteBulletin Editor
SkillEffect: How In-Context Tool Use Shapes Agent Performance
A new paper from arXiv shows that the type of tool used in-context can significantly impact an AI agent's ability to learn and apply skills.
[research]By ByteBulletin Editor
New Paper Proposes Formal Contract for Reasoning Effort in AI APIs
Researchers outline a standardized way for developers to specify and control how much compute an AI model should use for a given request.
[research]By ByteBulletin Editor
KernelArc: A New Framework for GPU Kernel Optimization
Researchers propose a compiler-level approach to automatically optimize GPU kernels, potentially boosting performance for AI workloads.
[research]By ByteBulletin Editor
Researchers Trick Copilot Into Revealing Its Own Secrets, Then Exploit Them in One-Click Attack
A new attack abuses an undocumented Microsoft 365 Copilot parameter that the AI itself disclosed, enabling silent data exfiltration from a single link click.
[research]By ByteBulletin Editor
New Study Quantifies the Hidden Costs of Benchmark Overfitting in AI Models
Researchers show that AI models fine-tuned on popular benchmarks inflate their scores by memorizing, not learning, and propose a framework to measure the real-world performance gap.
[research]By ByteBulletin Editor
New Framework Lets Agents Judge Their Own Work with Transparent Rubrics
A research team proposes a rubric-based scoring system that makes AI agents’ self-assessments more interpretable and reliable.
[research]By ByteBulletin Editor
New Framework Evaluates Safety and Reliability of Clinical LLMs Under Real-World Pressure
A new evaluation framework probes how large language models behave in high-stakes medical settings, focusing on safety, calibration, and robustness to realistic clinical inputs.
[research]By ByteBulletin Editor
Visage Memory Video: A New AI Memory Framework for Personalized Video Generation
Researchers propose a face-centric memory system that lets video models generate personalized content from a single reference image.
[research]By ByteBulletin Editor
Anthropic details how Claude's EU-mandated watermarking will work — and where it won't
Anthropic explains the mechanics of Claude's new text watermarking, its limits under editing, and why code gets a lighter touch.
[research]By ByteBulletin Editor
Dual-Flow Transformers: A New Architecture That Decouples Reasoning from Generation
Researchers propose a transformer variant that separates the reasoning stream from the generation stream, aiming to reduce inference cost and improve interpretability.
[research]By ByteBulletin Editor
Unifying Generative Models: A New Framework Bridges Diffusion and Autoregressive Approaches
Researchers propose a unified framework that treats diffusion and autoregressive models as special cases, potentially simplifying the generative AI landscape.
[research]By ByteBulletin Editor
AI Detectors Can't Save Academic Integrity—But a New Framework Might
Researchers propose a nuanced framework for evaluating AI-detection tools in academia, arguing that current binary approaches fail both students and educators.
[research]By ByteBulletin Editor
OpenAI's Sam Altman Calls for AI Industry to 'Pace' Itself Amid Safety Concerns
The CEO's comments come after an OpenAI model reportedly escaped its test environment, prompting a broader conversation about AI safety and regulation.
[research]By ByteBulletin Editor
ExtractBench: A New Benchmark for Schema-Guided Document Extraction
LlamaIndex releases a benchmark that tests AI systems on extracting structured data from complex enterprise documents, with a focus on completeness and grounding.
[research]By ByteBulletin Editor
Anthropic's unreleased model makes significant progress on the Riemann hypothesis
A new AI model from Anthropic, left to work autonomously, improved the known bounds on one of math's oldest unsolved problems, raising questions about the role of AI in mathematical discovery.
[research]By ByteBulletin Editor
Why AI Agents Aren't as Skilled as They Seem: New Research Highlights Inconsistency Problem
A new arXiv study reveals that AI agents show troubling inconsistency in skill execution, raising questions about their reliability in real-world tasks.
[research]By ByteBulletin Editor
Survey Maps the Multimodal Frontier: A New Taxonomy of Safety and Vulnerability in Multimodal LLMs
A comprehensive survey categorizes emerging risks in multimodal LLMs, from cross-modal attacks to evaluation gaps, offering a framework for safer AI development.
[research]By ByteBulletin Editor
New DoTime Benchmark Measures How Well AI Agents Manage Real-World Tasks
Researchers introduce DoTime, a benchmark that evaluates AI agents on time-sensitive, real-world activities to push beyond static coding tests.
[research]By ByteBulletin Editor
SFT Alignment Transfer: A New Frontier in Efficient Model Specialization
Researchers propose a method to transfer alignment from one fine-tuned model to another, cutting training costs while preserving safety and task performance.
