News Around the World — 2026-10-21

Development

• InfoQ — TornadoVM 7 Enables GPU-Native Workloads From Java Applications — TornadoVM 7.3 drops JNI for the Foreign Function & Memory API, adds an optimized CUDA backend with cuBLAS/CUTLASS calls and multi-GPU execution, and ships jitLLM 2.0, a single-JAR Java AI inference engine positioned against llama.cpp and Ollama (no benchmark numbers given).

AI

• Hugging Face Papers — More Memory, Worse Agents: Error Reproduction and Anti-Persistence in LLM Agents — across nine backbone/benchmark settings on ALFWorld, WebShop, and ScienceWorld, discarding task-specific abstractions after each use (“Anti-Persistence”) beats trajectory, joint, and abstraction memory on held-out success (59.9% vs. 53.5%/49.3%/49.0%) while also costing fewer LLM calls — more persistent agent memory made results worse, not better.
• Hugging Face Papers — Harness Self-Evolution for Long-Horizon Terminal Agents (RiseHarness) — iteratively evolving the agent harness around a fixed model lifts DeepSeek-V4-Flash’s Terminal-Bench Pass@1 by 7.25% and Pass^4 by 14%, and GPT-6-sol’s Pass@1 by 5.25%, showing harness design (not just the model) drives long-horizon terminal-agent performance.