Google Research has open-sourced Agile Deliberation, a framework that lets users define fuzzy 'gray area' concepts before an AI classifier learns them. The system splits a concept into in-scope and out-of-scope subconcepts, surfaces borderline samples for user labeling, and has Gemini rewrite the definition as a VLM prompt. Across 18 live sessions it beat zero-shot Gemini 2.5 Flash by 10.5 F1 points on average, with code released under Apache 2.0.
Zero-shot Learning
零样本学习
Ability to perform new tasks without task-specific fine-tuning
Curated Intelligence
Unitree founder Wang Xingxing revealed that the G1 humanoid robot has produced 10,000-20,000 units, making it the world's best-selling model. Tests show frontier LLMs like GPT-6 can directly control robots for tasks like painting and Go without training. The company is developing a self-evolving system called 'Unitree Preview', aiming for voice-controlled task completion by next year.
A developer introduced KillSwitch, an esoteric programming language designed to probe whether LLMs truly understand code logic or rely on pattern matching from training data. By embedding hidden lies in the spec and exploiting weaknesses like character counting and safety guardrails, it serves as an adversarial benchmark for testing coding agents' robustness.
Magic Atom's VLM Agent framework EchoPolicy-0 ranked first on the global MolmoSpaces benchmark with a 75.4% average success rate, while its sibling framework MagicBrain secured fourth place. This achievement marks a shift from 'demo-ready' to 'production-ready' embodied intelligence, validating the effectiveness of hierarchical architectures and instruction-following capabilities in complex zero-shot environments.
Huawei Cloud and East China Normal University have launched the 'Qiyao' Ascend-native time-series foundation model, which has topped the prestigious TIME-leaderboard and GIFT-Eval (Pretrained category). Deeply optimized for the Huawei Ascend ecosystem, the model offers high accuracy and robustness, supporting zero-shot prediction for scenarios like power grids and finance, while enabling low-cost fine-tuning for vertical industry customization.
Artificial Analysis announced the upcoming release of AA-Robotics v0.1, a benchmark designed to evaluate whether frontier language models can control real robot arms zero-shot without coaching. The announcement is currently a teaser with no specific methodology, dataset details, or results disclosed yet.
OpenAI released a GitHub repository containing 722 mathematical manuscripts generated by an unreleased internal model, tackling major problems like the Quasi-Riemann Hypothesis and 4D Kakeya Conjecture. Sam Altman described this as the dawn of a new era of discovery, sparking significant interest in both AI and mathematics communities.
Musubi Labs released and open-sourced PolicyLM-1.7B, a lightweight decision model designed for real-time content moderation. The model applies natural language policies to messages in under 50 milliseconds without requiring retraining when policies change, significantly reducing compliance iteration costs and latency.
To address high inference costs in complex AI search caused by query decomposition, the 'Retrieve-for-Train' method is proposed. It replaces zero-shot LLM dynamic planning with retrieval-augmented training, significantly reducing reliance on expensive thinking budgets for efficient set retrieval.
Google introduces TimesFM-3, a time series foundation model enabling zero-shot multivariate forecasting in a single forward pass. It breaks the univariate limitation of previous versions, significantly outperforming other models on major benchmarks and offering direct access via BigQuery's AI.FORECAST command.