Archive
GitHub 熱門專案、開源工具、框架、runtime 與重要 release。
代理工具
A Google team used multi-turn reinforcement learning to train Gemini 3.5 Flash with simulated patients, document tools, and multidimensional clinical rewards. The model missed fewer red flags in adversarial consultations, but the training system has not been open-sourced or prospectively validated with real patients.
最新模型
New experiments find that when Muon trains a Transformer’s hidden matrices while AdamW updates its embeddings and output head, the model learns modular arithmetic faster but cannot reliably retain the solution. Freezing either parameter group prevents collapse, suggesting that the problem stems from joint drift at the representation–readout interface.
代理工具
Shepherd records model calls, tool operations, and file changes as persistent execution traces, allowing users to inspect or discard modifications before they enter the workspace. The latest documentation also explicitly states that the meta-agent delegation, in-flight forking, and automated retry interfaces demonstrated in the paper and on the project homepage are not yet directly usable features in v0.3.0.
最新模型
Rather than directly reconstructing masked gene values, CellWorld predicts representations from a target encoder using neighboring cells and limited expression hints. The code and training configurations have been released under the MIT License, but the pretrained weights remain in an approval-gated private repository.
最新模型
AgentPatch first identifies weaknesses introduced by model merging, then restores only the expert models’ unique residuals and critical FFN neurons. It produces a single static checkpoint with no routing or inference-time agents, but the complete evaluation adapters and model weights have yet to be released.
GitHub Repo
AIPOCH’s Apache-2.0 desktop workbench integrates model agents, Python and R kernels, scientific data connectors, and immutable artifact versioning. The August 9 update makes context compaction visible and fixes Codex MCP approval and Windows notebook execution issues, but still falls short of fully portable replay.
代理工具
Rather than distributing the outcome of an entire agent run evenly across all tokens, DiDPO compares similar sub-diffs across multiple trajectories. The public verl-code repository includes training entry points and data recipes, but no author-released model checkpoint yet, and the experiments do not cover full repository-level software repair.
代理工具
P-Bench requires agents to select statistical methods, execute R code, and report p-values themselves, rather than merely checking whether the code runs. The 14B model leads the listed open models on the strict metric for difficult tasks, but the public repository currently contains only documentation.
推論系統
A community deployment recipe running on a single DGX Spark uses CUDA Graph and a one-layer MTP speculative decoder to increase Ling-3.0-flash INT4 generation speed from 20.8 to 38.7 tokens/s. More importantly, it finds that mainline vLLM does not yet support the V3 architecture: forcing the model through an older attention path may not raise an error but can produce plausible-looking yet incorrect content.
最新模型
MiniMax H3 packages text, images, video, and audio into a single sequence and jointly predicts visual and audio latents, producing up to 768p video at 24 FPS with 32 kHz stereo audio. The official workflow claims regeneration at up to 2K, but prompt orchestration and the high-resolution module remain available only through the hosted API.
代理工具
Rather than retaining every executable synthetic task, CalibForge repeatedly revises tasks until different solver agents exhibit predefined success and failure patterns. Two models fine-tuned on calibrated trajectories improved across three coding-agent benchmarks, but full training required 64 NVIDIA H20 GPUs, and some results were based on only a single run.
AI 安全
A new benchmark uses four paired contexts to distinguish whether models correctly evaluate evidence or simply ignore all external information. By changing only the configuration of DPO preference data, SCOPE reduces answer flips caused by misleading contexts while preserving accuracy in correct and irrelevant contexts.