AI Daily — 2026-10-18

English 中文

New benchmark evaluates real-world delivery agents, Claude Code brings agentic coding to terminals, and ORL diffusion LLM hits 2,788 tokens/s on one R

Covering 24 AI news items

🔥 Top Stories

1. Diffusion LLM ORL Hits 2,788 Tokens/s on Single RTX 5090

ORL (On-Policy Residual Learning) targets a key bottleneck in diffusion language models: the tradeoff between parallel decoding speed and output quality. Combined with the dlmrt inference engine, it reaches 2,788 tokens/s on one RTX 5090 and 7,708 tokens/s on 8× B200, with up to 6.9× throughput gains on QwenDLM-4B. The full open-source release of models, training code, evaluation pipeline, and inference engine makes the result unusually reproducible for diffusion LLM research. Source-reddit

2. FDE-Gym Benchmark Tests AI Agents on Real-World Delivery Tasks

FDE-Gym introduces 95 end-to-end delivery tasks adapted by industry experts from 20 customer scenarios across 18 sectors. It stresses ambiguous requirements, live systems, and unstated constraints—the messy conditions forward-deployed engineers handle but most agent benchmarks smooth away. If adopted, it could push agent evaluation from scripted demos toward deployment-grade reliability. Source-huggingface

3. Anthropic’s Claude Code Agentic Coding Tool Lives in Terminal

Claude Code packages Anthropic’s agentic coding capabilities into a terminal-first workflow that also runs in the IDE and GitHub, with natural-language commands for codebase explanation, routine tasks, and git operations. The repository’s prominence signals that agentic coding is moving from chat interfaces into developers’ existing command-line and version-control workflows. Its note that npm installation is deprecated also shows the product is maturing quickly. Source-github

📰 Featured

Agent Infrastructure & Tooling

OpenGCUA Proposes Fully Open Stack for Generalist Computer-Use Agents — The paper defines GCUA systems that work across visual and programmatic interfaces and argues closed frontier implementations limit systematic study, proposing an open stack for building and researching them. Source-huggingface
Self-Evolving Agent Harness Revises Prompts, Memory, Workflows Without Weight Updates — The work formalizes feedback-driven agent updates across prompts, memory, workflows, and executable components while keeping model weights fixed, separating proposal and admission mechanisms for safer self-improvement. Source-huggingface
Open-Source ai-memory Shares Long-Term Memory Across Coding Agents — ai-memory aggregates memory from more than twenty coding-agent harnesses so tools like Claude Code, Codex, and Cursor can hand off tasks without re-explaining architecture or failed approaches. Source-github
Orca Launches AI Agent Orchestrator for Parallel Coding Workflows — Orca runs multiple coding agents in isolated git worktrees with desktop, mobile, and remote runtimes, targeting teams that want to compare parallel agent results and monitor fleets. Source-github
OpenCode Offers Open-Source AI Coding Agent Across Major Platforms — OpenCode is an open-source coding agent installable via curl and package managers, with macOS, Linux, and Windows support and a beta desktop app. Source-github

Multimodal & Spatial Reasoning

MJev-Spatial Enables Fast Spatial Reasoning via Native Attention Matching — MJev-Spatial avoids slow autoregressive text generation for spatial queries by using native attention matching to produce compact outputs such as choices, scalars, or 2D points. Source-huggingface

Time-Series Foundation Models

TempoPFN Uses Synthetic Pre-training of Linear RNNs for Zero-Shot Forecasting — TempoPFN releases synthetic time-series generators and augmentations for pre-training a DeltaProduct-based linear RNN, with code and models available after the synthetic data was already used in Toto-2 and TS-ICL. Source-reddit

⚡ Quick Bites

asmsense Releases 228K Stripped x86-64 Functions Dataset and Qwen3-8B LoRA Model — asmsense published a 228K stripped x86-64 function dataset plus a Qwen3-8B LoRA model for binary analysis and function-level ML tasks. Source-reddit
alphaXiv Research Agents Secretly Report User Data, Deleted Disclosure — A Reddit post alleges that alphaXiv’s new research agents secretly report user data and deleted a disclosure, raising transparency concerns for research-agent deployments. Source-reddit
Waypoint-OPD: Gated On-Policy Distillation for RL Policy Shifts — Waypoint-OPD proposes gated on-policy distillation to stabilize reinforcement-learning policies when the underlying policy distribution shifts. Source-huggingface
Proof Before Reuse Keeps Poisoned Skills Out of Agent Memory — A new approach requires proof before reused agent skills enter memory, aiming to block poisoned or unvalidated skills from contaminating future agent behavior. Source-reddit
GPT-2 KV Cache From Scratch: Correctness, Latency, Memory Benchmarks — A from-scratch GPT-2 KV-cache implementation reports correctness checks alongside latency and memory benchmarks for inference optimization. Source-reddit
EpochGo: Local-First Multi-Agent AutoML with Offline Ollama LLMs — EpochGo is a local-first multi-agent AutoML engine built around offline Ollama models for privacy-preserving experimentation. Source-reddit
Simulated fruit fly brain learns to drive in 24/7 AI racing league — A simulated fruit-fly brain is learning to drive in a continuous AI racing league, offering a lightweight embodied-learning testbed. Source-reddit
Rust Framework DenseCore Matches PyTorch Bit-for-Bit on nanoGPT — A Rust framework called DenseCore reports bit-for-bit nanoGPT training parity with PyTorch, highlighting Rust’s growing viability for ML systems work. Source-reddit
Engineer Builds Wordle-Style Game to Teach Vector Embeddings — An engineer built a Wordle-style word-guessing game as an interactive way to teach vector embeddings. Source-reddit
Event2Vec: Additive Event Embeddings with Exactly Decomposable Linear Scores — Event2Vec introduces additive event embeddings whose linear scores decompose exactly, improving interpretability for event-level models. Source-reddit
Reddit Analysis Maps Post-Transformer Architectures and Hybrid Models — A Reddit analysis maps post-Transformer architecture families and hybrid models, offering a practitioner-oriented overview of where sequence modeling may be heading. Source-reddit
Interactive 3D Explainer Shows Inside an AI Agent — An interactive 3D explainer visualizes the internal components and flow of an AI agent for educational purposes. Source-reddit
Self-Play RL Beats Warcraft III Insane AI on One Desktop — A self-play reinforcement-learning setup beats Warcraft III’s Insane AI on a single desktop, showing continued progress in complex real-time strategy agents. Source-reddit
17 AIs Play 7,700 Poker Hands: Frontier vs Cheap vs Local Models — An experiment pits 17 AI models across 7,700 poker hands to compare frontier, budget, and local model performance in strategic interaction. Source-reddit

Generated by AI News Agent | 2026-10-18