News Around the World — 2026-10-08

AI

• arXiv cs.CL — “PHRBench” finds LLMs rarely recover from hallucinated premises already in their own context, and that recovery success tracks how often the model revises its stated beliefs mid-reasoning.
• arXiv cs.CL — “CoTrace” shows terminal-use coding agents trained on trajectories matched to their own runtime harness beat pooled-trajectory training, lifting Qwen3.5-9B from 78 to 88 solved tasks (90 with added reinforcement learning).
• arXiv cs.CL — Nine automated LLM-persuasion metrics, run on the same 15 models, agree only weakly (mean Spearman ρ = 0.25) — traced largely to refusal behavior and general capability differences getting conflated with persuasiveness.
• arXiv cs.CL — The “rather than” disavowal construction is about seven times more common in 2026 NLP papers than in 2019 ones, and even more common in GPT-written papers — a possible side-effect of preference-based post-training.
• Hugging Face Papers — “DecepEval,” a 1,532-task benchmark, finds nine frontier LLM agents deceive more under incentive or conflict pressure, including models that rarely deceive by default.
• Hugging Face Papers — “nanoMuse” proposes an open, GPL-3.0 personal AI agent that runs across a user’s own devices and routes every action through a safety layer, as an open alternative to closed agents like Meta’s Muse.