AI BriefingPyTorchKR news
The latest AI and developer news from PyTorchKR, with short summaries.
Visit website →Convai Unveils 'Laya', a Decision Model Without Text Generation
A Jev-compatible open-weight model that returns probability distributions in a single forward pass to address LLM hallucination and latency issues.
CLM-v0.1-8B Released, 13x Faster Decision-Making Compared to Jev
CLM-v0.1-8B, based on Qwen3-8B, released with candidate action selection speed up to 13x faster than Jev
Liquid AI Releases LFM2.5-VL-3B-DSpark, a Speculative Decoding Model for VLMs
Liquid AI has released LFM2.5-VL-3B-DSpark, a draft model for speculative decoding of its VLM LFM2.5-VL-3B, which improves decoding speed by up to 3.13x on the Apple M5 Max.
This Week's AI Papers: Loop Decoding, Self-Improvement, and Agent Benchmarks
Introducing recent research on performance enhancement and agent learning, including LoopCD, UniEvo-VL, SkillGym, and AIM.
UTAW Releases Report on AI-Driven Work Intensification
UK tech union UTAW released a report based on worker testimonies stating that AI adoption does not reduce workload but instead increases review burdens and stress.
llmwiki-serve Released: Connect Obsidian Docs as Context for Coding Agents
A local server that delivers existing Markdown/Obsidian documents to coding agents via CLI, HTTP, and MCP without modification.
ACP Client 'agent-shell' for Emacs Released
The Emacs package agent-shell, based on the ACP protocol, has been released, enabling control of 19 coding agents through a single UI.
Figure Announces Humanoid Helix 2.5
Figure achieves an average household task success rate of 56% across 30 unseen homes by pre-training on human behavior data
vLLM Introduces Hardware-Agnostic Layer Architecture to Support Diverse Accelerators
The new architecture maintains within 3.4% of native token throughput on NVIDIA H100 GPUs while enabling support for out-of-tree accelerators and older hardware.
vLLM rope_scaling Override Causes 36% Model Configuration Mismatch
Released the open-source tool entail to verify and prevent the issue where vLLM's rope_scaling override bug causes GSM8K scores to drop by up to 28%.
vLLM Releases 'Decision 1.0', AI Models Specialized for Decision-Making
Released six open-weight models and a Studio for testing that return the probability of the optimal choice among candidates instead of generating text.
PickAnthropic Releases Migration Guide for Claude Opus 5.5
Reflects API structural changes including a 30% output speed improvement over Opus 5, forced thinking activation, and default effort level changes.
PickTrillion Labs Releases GravityOCR, an OCR Model Applying Self-Speculative Decoding
GravityOCR, based on GLM-OCR, increased decoding speed by 1.32x without adding parameters while maintaining an accuracy score of 95.16.
NVIDIA Redefines AI Agent Evaluation Metrics
NVIDIA introduced a new framework for AI agent evaluation that emphasizes task completion and cost efficiency over tool-calling accuracy.
StepFun Releases Step 5 Preview
StepFun has released Step 5 Preview, offering intelligence on par with Kimi K3 at one-third the cost, with weights scheduled to open on October 15.
PickPostgreSQL Extension pgjev Released, Supporting Natural Language Filtering Without Embeddings
The PostgreSQL extension pgjev has been released, enabling natural language row filtering in SQL WHERE clauses without embeddings or vector columns.
NVIDIA Releases AIPerf, an LLM Inference Benchmarking Tool
NVIDIA has released AIPerf, an LLM inference benchmarking tool that eliminates client-side bottlenecks.
Google Research Unveils R4T, an RL-Based Search Distillation Technique
Google Research and UIUC unveiled R4T, which distills RL-trained query fan-out behaviors into a 53.9M parameter diffusion model, reducing search latency by 12–20x.
Bend 2 Released: A Proof-Oriented Language That Fails Compilation on AI Code Rule Violations
Bend 2, a proof-oriented language designed to fail compilation when AI-generated code violates human-specified rules (LAWS.bend), has been released.
Context Compression Plugin 'Fast Jev Compaction' for Claude Code Released
A Claude Code plugin has been released that selectively deletes unnecessary tool call results without summarizing the conversation, utilizing TypeSafe AI's Jev model.