#Reasoning

Stanford has open-sourced Memorilla, a potential memory module for freezing LLM development

Stanford researchers have open-sourced Memorilla, a potential memory system designed for frozen LLMs. It converts the entire document collection into 16 vectors for use by frozen LLMs. On FactKG, the frozen Qwen3-8B paired with Memorilla achieves an accuracy of 84.33%, which is 3.9% higher than using all document contexts, but only requires 16 memory tokens, reducing the number of document tokens by 108 times.

Vercel releases UI SDK: Provides 50+ UI behaviors for AI coding agents

Vercel Labs has released the UI SDK, an open-source UI behavior library designed for AI-coded agents. The SDK provides over 50 behaviors and supports React, Vue, Svelte, Solid, and Web Components, reducing tool calls for agent task setup by 69% in Vercel's testing, decreasing task time from 41 seconds to 16 seconds, and lowering costs by 56%.

Tsinghua University releases the single-cell proteomics foundational model ProtCell, successfully completing the 'virtual cell' pre-training reproduction using MolThread S5000 intelligent computing cards

Tsinghua University releases the single-cell proteomics foundational model ProtCell, collecting about 380 million cells and 98 public datasets for pre-training and evaluation, MolThread S5000 intelligent computing cards complete the full process of pre-training reproduction, evaluation metrics differ from mainstream international GPUs by only about 0.2%

Google expands the Antigravity multi-agent system, launches the Teamwork system, and verifies multiple mathematical proofs.

Google announces expansion of its Antigravity multi-agent system, forming the Teamwork system with multiple professional agents driven by Gemini, capable of verifying complex mathematical proofs and optimizing software system performance.

Test-Time Baking (TTB) released: A strategy for efficiently rendering decoder-only large-view synthesis models

Test-Time Baking (TTB) strategy enhances the fidelity of new-view decoding for large-view synthesis models (LVSMs) by baking global scene information from cross-view attention into lightweight MLPs, while maintaining Test-Time Training efficiency. It achieves up to 23x rendering speed improvement with quality comparable to current state-of-the-art techniques on multiple benchmarks.

Claude Code v2.1.159 release

Claude Code v2.1.159 released, adopting a gradient alignment reward mechanism, using the Qwen3-8B model on arithmetic tasks, rewarding by calculating the gradient similarity between student output and expert demonstrations, experiments show the model may overfit the precise wording of expert examples.

llama.cpp 1.0.0 released: supports over 150 model architectures, significantly improving the ability to run large models locally.

llama.cpp version 1.0.0 has been released, supporting over 150 model architectures, marking API/ABI incompatible changes with the major version number; the ggml underlying engine has been enhanced, supporting dynamic loading of multiple hardware backends, improving the performance of locally deployed large models.

Amazon Bedrock AgentCore runtime V2 architecture mode: Accelerating session startup

Amazon Bedrock AgentCore runtime V2 reduces session startup time to about 2 seconds by preparing the environment once and using a snapshot for recovery, while also reclaiming unused memory. V2 supports building hooks to load large files or CLI, suitable for scenarios requiring resource preparation in advance. Session startup time can be measured via CloudWatch metrics, prioritizing optimization of platform overhead or code execution time.

GitHub Copilot 0.171.0-alpha.5 released

GitHub Copilot released version 0.171.0-alpha.5, providing code review, vulnerability detection, and secret scanning features, supporting enterprise-level code security and AI features.

Google Opens /teamwork Multi-Agent Workflow in Antigravity

Google has opened the /teamwork multi-agent workflow in Antigravity to external researchers and developers, which can be used in conjunction with Gemini 4 Argon, Carbon, and Hydrogen; /boost and Deep Review are also available.

langchain-modal 0.0.8 release: fixed file I/O issues

langchain-modal 0.0.8 has been released, fixing the `ModalSandbox` file I/O issue and updated to use the `Sandbox.filesystem` API

Running native DLA inference on NVIDIA Jetson AGX Orin using cuDLA through Python.

NVIDIA Jetson AGX Orin can run native DLA inference in Python via cuDLA, supporting the execution of specific neural network operations on DLA hardware to offload GPU inference tasks, thereby improving the energy efficiency of edge applications.

How AssemblyAI optimizes speech-to-text inference costs using EC2 Spot instances on AWS

AssemblyAI builds an elastic architecture on AWS by hybridly using EC2 On-Demand and Spot instances, and leveraging Amazon ECS, SQS, and EventBridge, to increase GPU instance utilization by 3.5 times, achieve cost reduction within the European AWS region, and meet service level agreements (SLA). Its inference platform is deployed on three GPU instance families: g4dn, g5, and g6.

Personal Agent Protocol 1.0 released

The Personal Agent Protocol (Poppy) v1.0 protocol is now released, and Shopify has enabled it for all merchants on the platform. Multiple design partners, including Mercury, Synchrony, and Turkish Airlines, have started going live. Meta's AI agent Muse is automatically prioritizing Poppy on every website visited. The protocol uses OAuth authentication to let customers control personal agent permissions and supports phased adoption, introducing an e-commerce extension co-developed by Shopify.

Perplexity releases Decider v1.3: Open-source decision model adds Locate feature

Perplexity releases open-source decision-making model Decider v1.3 and its lightweight version pplx-decider-flash-v1.3, adding a Locate feature to extract precise text segments.

Open Vision Banana: A completely open-source generative visual perception solution.

Open Vision Banana (OVB) has open-sourced its data, weights, code, and implementation details. Its 4B and 9B parameter-based open-weighted large model and 2.1M-3.8M training pairs surpass Vision Banana on Cityscapes with zero-shot performance of 73.9 vs 69.9 mIoU and can rival state-of-the-art affine-invariant depth estimators.

LangChain Deep Agents 0.7.38 released

LangChain Deep Agents 0.7.38 was released to fix the issue where `ls`, `grep`, and `glob` would incorrectly return results when soft link resolution exceeds the approved search scope.

How does Dyania Health use Synapsis AI to assist healthcare teams in improving patient care efficiency?

SparseRead: A reusable control layer for AI agents to efficiently sparse read external content

The SparseRead control layer reduces excessive reading of external content by AI agents through Read Intent Protocol, Adaptive Reader Backends, and Budgeted Read Gate, without modifying the agent model.

Google releases TPU Sync: an open-source KV cache transfer library for TPUs

Google has released the open-source C++ library TPU Sync, used to move KV cache and model weights on Cloud TPUs, treating TPU HBM, host DRAM, shared storage, and data center network as a unified memory hierarchy.

inclusionAI/Duo-1.0-distill-quant released: An FP8 quantized real-time audio-visual full-duplex multimodal world model with four-step distillation

inclusionAI releases the Duo 1.0 Realization weights for FP8-quantized four-step distillation, used for its causal base model, supporting full-duplex real-time audio-video interaction, and needs to be overlaid with the causal base model weights.

Understanding the decision model in long-term intelligent agents: Actor, Judge, or Teacher?

Decision models, even as actors, can effectively teach reinforcement learning strategies by evaluating complete action trajectories, improving success rates on ALFWorld and WebShop.

Players use retired Intel Optane memory to run the 2.8 trillion parameter Kimi K3 model, outputting 1.5 tokens per second

Users have successfully run the Kimi K3 model with Intel Optane Persistent Memory, which has been discontinued, without a graphics card, achieving 2.8 trillion parameters. The initial generation speed was approximately 0.85 tokens per second, and it increased to 1.50 tokens per second after disabling the DAX function. The 3-bit quantization test only reached 0.75 tokens per second, and generating 50,000 tokens took 18.5 hours.