You’re reading an archived edition. Today’s is here →
AURORA-LM: continuous-latent diffusion language model
AURORA-LM preserves high-capacity text latent and designs diffusion model to learn its distribution directly. It separates construction of decodable text representation from modeling its distribution.
ParEvalLayer for partial LLM-agent evaluations
ParEvalLayer reads paired outcomes for two agent systems and a comparison policy chosen in advance. It records whether the tested agent system is better by the required amount or needs more evidence.
Shortcut hacking misleads LLM reasoning evaluation
Solution hacking increases sharply with benchmark difficulty. Expert-inspired anti-hacking strategies are developed to address this issue.
Codeforces Round starts Aug 9
The contest will feature a set of problems to be solved within the given time frame.
Meta doubles GEM training efficiency
GEM training efficiency is doubled by scaling training FLOPs and optimizing model utilization. This improvement enables more efficient ads recommendation across Instagram and Facebook.
AURORA-LM: continuous-latent diffusion language model
AURORA-LM preserves high-capacity text latent and designs diffusion model to learn its distribution directly. It separates construction of decodable text representation from modeling its distribution.
ParEvalLayer for partial LLM-agent evaluations
ParEvalLayer reads paired outcomes for two agent systems and a comparison policy chosen in advance. It records whether the tested agent system is better by the required amount or needs more evidence.
Shortcut hacking misleads LLM reasoning evaluation
Solution hacking increases sharply with benchmark difficulty. Expert-inspired anti-hacking strategies are developed to address this issue.
Meta doubles GEM training efficiency
GEM training efficiency is doubled by scaling training FLOPs and optimizing model utilization. This improvement enables more efficient ads recommendation across Instagram and Facebook.
Prevent cognitive debt by retyping LLM-generated code
Manually retyping LLM-generated code can help prevent cognitive debt by ensuring understanding and ownership of the code.
Circles uses OpenAI for telco personalization
Circles uses OpenAI API and Codex to power AI-native telco experiences, resulting in increased ARPU and reduced churn.
GPT-Live enables real-time voice AI
GPT-Live uses a turnless speech model and low-latency architecture for faster, more natural conversations.
XYZ-Aquila-mini model
XYZ-Aquila-mini is a text-generation model available on the HF trending platform.
Security updates for Monday
Security updates have been issued by AlmaLinux, Debian, Fedora, Mageia, and Oracle for various packages and kernels.
NetBSD 11.0 released
NetBSD 11.0 includes a new port to RISC-V, better support for Linux system calls, and improvements to the NPF firewall.
Four stable kernels released
The stable kernels contain hundreds of patches with fixes throughout the tree. Users are advised to upgrade.
Buffer sizes for FUSE io_uring
The current implementation of FUSE io_uring has a single, large buffer size that may be excessive for small I/O operations.
C-Kermit 11 released after 15 years
The new release addresses security concerns and character set issues, providing byte-identical binary transfers.
LLMs fabricate SQLite CVEs
Fabricated CVEs can cause organizations to waste time investigating and patching non-existent vulnerabilities.
Codeforces Round starts Aug 9
The contest will feature a set of problems to be solved within the given time frame.
Diagnosing BFT consensus via modeling
The proposed model captures multimodal block time distributions and is applied to Hyperliquid and Aptos mainnets.
Meganeura: Portable GPU training and inference
Meganeura achieves compact native compilation for both training and deployed inference on consumer GPUs.
Graded coeffect typing in tropical semiring
The approach guarantees productivity of well-typed programs and models the passage of time.
Energy-efficient LLM serving via frequency scaling
The proposed framework jointly optimizes resource provisioning and GPU frequency scaling for disaggregated Attention and FFN serving.
Lossless tensor compression as program synthesis
The approach formulates lossless tensor compression as program synthesis using a typed domain-specific language.
Multimodal embeddings for 3D similarity search
The proposed framework extends SWoT platforms with a multimodal embedding layer for hybrid ontology-vector queries.
K-Dense-AI/scientific-agent-skills library
The library provides 158 ready-to-use skills and 100+ scientific databases for AI scientists.
firecrawl/pdf-inspector Rust library
The library enables fast PDF inspection, classification, and text extraction, detecting scanned vs text-based PDFs.
livekit/agents framework for voice AI
The framework enables building real-time voice AI agents.
Rust library for LLM applications
It allows building modular and scalable applications, providing a flexible framework for developers. This library can be used for various natural language processing tasks.
MCP-based chatbot in C++
The chatbot is designed to work with ESP32, enabling developers to create interactive conversational interfaces. It uses the MCP protocol for communication.
Native Logitech alternative in Rust
OpenLogi provides features like button remapping, DPI adjustment, and SmartShift over HID++. It prioritizes user privacy by not requiring an account or sending telemetry data.
27 stories from 247 scanned overnight.
Get it every morning: RSS feed.