Archive
1541 technical news articles
代理工具
A Microsoft Research team jointly updates synthetic applications, tasks, and graders based on agent failure trajectories instead of merely adding more static webpages. The 9B model’s average success rate across 14 evaluations rose from 36.5% to 67.1%, although only four environments have been publicly released as benchmarks.
代理工具
Alibaba’s Qwen team has released a technical report on a GUI agent spanning mobile, desktop, web, and DeepSearch environments. Its unified action space combines on-screen interactions, CLI commands, and batched actions. The model leads existing systems on several self-reported benchmarks, but its weights, training data, and complete execution framework have not yet been released.
代理工具
Using 1,019 human-annotated trajectories spanning the web, Windows, Ubuntu, and mobile platforms, OSReward tests whether vision-language models can correctly determine whether computer agents have completed their tasks. The team also released 100,000 reasoning-based judgment examples and trained self-hostable 9B and 35B OS-Shepherd reward models.
最新模型
NVIDIA has released the Cosmos 3 Edge weights and inference pipeline, combining autoregressive reasoning and diffusion-based multimodal generation in a 4B Mixture-of-Transformers model. The model accepts text, images, and action trajectories, and a robot policy variant fine-tuned on DROID data is also available, though physical consistency and edge-device latency still require independent verification.
AISPA 把系統提示拆成八項使用者保障維度，並分析 88 款商用 AI 產品中的 3,249 條指令。雖然 98.9% 的產品至少包含一項保護性規則，完整覆蓋八個面向的比例只有 24%。
推論系統
WIDE brings dynamic pruning down to token-level attention-head groups and FFN channel groups, supporting both prefill and decode. At 50% sparsity, the study reports up to a 4.95× decoding speedup for core layers, but the end-to-end gain falls to 1.55×.
AI 安全
The public review of the OWASP Agentic Skills Top 10 v1 draft has concluded. The framework organizes malicious skills, excessive permissions, dependency confusion, and cross-platform portability risks into ten control categories. It treats skills as the behavioral layer between models and MCP tools and proposes a universal format for declaring permissions, provenance, and integrity.
最新模型
VIDRAFT has released the weights for its POCKET series, quantizing a 35B model down to as little as 8.2 GB while retaining roughly 3B active parameters per token. The models run directly with upstream llama.cpp, although the compression method remains undisclosed and current performance and quality evidence comes primarily from the development team.
AI 安全
After Google connected Nano Banana 2 to real-world geospatial imagery, users quickly generated and shared screenshots of fabricated explosions, airstrikes, and other scenes. The company withdrew the feature the next day. Although generated images carried watermarks and were not added to shared maps, the incident shows that provenance labels cannot replace input restrictions and interface isolation.
最新模型
Huawei’s openPangu team has released a 505B mixture-of-experts model trained on Ascend NPUs, activating approximately 18B parameters per token. The model introduces hierarchical sparse attention and three-head multi-token prediction, though its performance figures currently come primarily from the team’s own evaluations.
科學機器學習
UNICON converts data from rivers, traffic networks, power grids, weather systems, and human mobility into a shared graph-sequence interface, using a small number of numerical examples to specify new tasks in context. In unseen domains such as web traffic, the model approaches the performance of specialized models, but it has so far been validated only on forecasting tasks, and neither its weights nor code have been released.
AI 安全
A new study simulates auditing agent populations under limited human-review budgets and identifies the tipping point at which distorted self-reported confidence causes ranking-based strategies to underperform random sampling. Five small and midsize open-weight models achieved confidence AUROCs of only about 0.5 on two benchmarks, although the findings are based on a single prompt and direct-answer setup.