Daily updates, analyzing cutting-edge AI research papers
A Tsinghua-Oxford-Stanford study finds LLMs exhibit 'thinking inertia' – they continue explicit reasoning even when instructed not to, especially on open-ended tasks; suppressing reasoning reduces accuracy by ~15 points, revealing a fundamental trade-off between silence and correctness.
Renmin University introduces AEWM, an editing world model that classifies and revises noisy agent actions in real time, improving long-horizon task performance across search, terminal, and software engineering benchmarks without fabricating tool feedback.
This article summarizes two high-impact ACL 2026 papers: DRP reduces reasoning token usage by 64% while improving accuracy via skill-aware step pruning, and SAGE doubles skill reuse success rates in agents through sequential rollout and skill-integrated rewards.
Researchers from Griffith University and Tsinghua University introduce Spotter, a robot agent framework where an embodied model executes tasks while a vision-language model supervises in parallel, intervening only upon detected errors, achieving minute-level task completion with 70% less time than VLM-led baselines across RoboCasa, RoboTwin 2.0, and real-robot experiments.
The author tests Doubao Work Agent's new local Office editing feature on ICML 2026 paper data, demonstrating automated pivot table creation, chart generation, and PPT style unification without manual formatting, completing a 6500-row Excel analysis in 3 minutes and rescuing a messy presentation.
OpenAI has open-sourced 722 mathematical papers from an unpublished internal model, covering 17 fields with major breakthroughs including the quasi-Riemann hypothesis, full BSD formula, and Hilbert's tenth problem over rationals, averaging three hours of ChatGPT Pro compute per paper, while an independent advisory group of Fields medalists urges cautious verification.
Nobel laureate Geoffrey Hinton and 21 co-authors publish the first paper on recursive self-improvement (RSI), warning that AI automating its own research could accelerate progress tenfold within 1.5 years, citing current automation metrics, historical research ROI estimates, and the Hugging Face incident as evidence, urging immediate policy action.
Google's 32-author survey on tokenization reveals it as the most underestimated component in LLMs, showing 94% of models reuse identical tokenizers, a 54.3x cost disparity between English and low-resource languages, and that BPE dominates generative models while fundamental fairness and efficiency challenges remain largely unsolved.
OpenDots is an open-source, self-hostable framework for creating AI coworkers called Dots, each with its own isolated cloud computer, persistent document workspaces, human-in-the-loop approval, voice calls, and Slack integration, built on AG-UI protocol and TanStack AI.
Anthropic publishes its internal Claude agent engineering practices on claude.dev, covering dynamic multi-agent workflows, hundreds of reusable skills, context engineering principles that cut system prompts by 80%, and evaluation-driven hillclimbing that boosted accuracy to 90.5% at one-fifth the cost.
