News Around the World — 2026-10-13

AI

• Hugging Face Papers — AgentSWE — A 25-task coding-agent benchmark finds agents rarely deliver what the requirement actually asks for: the strongest builder (DeepSeek-V4-Pro in the Codex harness) averages a held-out score of just 19.1 on Creation and 31.6 on Editing, and Creation runs scoring 90+ on visible dev cases average only 9.9 on held-out cases.
• Hugging Face Papers — Scaling Agentic Data for Long-Horizon Terminal Intelligence — TerminalHorizon, a synthesis engine for long-horizon terminal-agent trajectories, lifts Qwen3.5-27B’s Terminal-Bench 2.0 score from 41.6% to 55.8% (+14.2 points) after fine-tuning on ~2,850 generated trajectories, with gains across all 12 benchmarks tested.
• Hugging Face Papers — Thinking Inertia — Disabling an LLM’s thinking mode doesn’t reliably stop visible reasoning: DeepSeek-V4-Flash still showed explicit inference in 99.9% of open-ended responses with thinking off, and forcing strict answer-only prompting cut mean accuracy by about 15 percentage points.