Molr is an open-source, locally deployable Personal Agent developed by author fcten and programmer Tibo, using a tech stack of Go + React/TypeScript + SQLite; it supports permission control, browser automation, local and remote sandbox execution, WeChat connection, as well as basic capabilities like long-term memory and task planning; the project is still under development with no stable version released yet, and the GitHub repository is molr-ai/molr.
ddter released an independently developed web radio tool, roamdial.com, which supports random access to approximately 35,000 real-time radio stations worldwide. It features second-level station switching, high-bitrate filtering, language filtering, song title recognition with favorites, lyric sing-along, and star/earth view browsing; favorite data is stored only in the local browser, with no accounts and no ads; the tech stack consists of a React + three.js frontend and a Node + SQLite backend, with the radio directory sourced from the Radio Browser community project and automatically refreshed weekly.
IneqAgent is an agent framework for Lean4 that integrates large language model reasoning with symbolic computation backends (sum-of-squares, Bernstein, successive difference-substitution), coordinating lemma construction, goal rewriting, and backend selection via Monte Carlo tree search; across four benchmarks totaling 186 problems, it solved 175 problems (94.09%) using DeepSeek-V4-Flash, the highest among evaluated methods; it also obtained Lean-verified strengthened forms of Dittert's conjecture in dimensions 4 and 5, as well as certification of the Holens--Djokovic (4,4) inequality, thereby implying the 4-dimensional Merris conjecture.
AlphaDiana is a modular agent measurement system that supports flexible integration of datasets, agents, environments, and scorers. It enables macro-level agent analysis and micro-level component interventions (such as tools, skills, and memory) under conditions controlling file reuse and memory retention. The authors validated their findings using three open-source reasoning models and three open-source harnesses across five answer-based benchmarks and two interactive benchmarks. They found that harnesses do not consistently improve accuracy; skill guidance and history retention also did not reliably improve accuracy; and output token usage efficiency needs to be evaluated jointly with correctness.
The paper points out that the rapid improvement in AI capabilities has led to benchmark saturation, causing traditional metrics such as accuracy to lose discriminative power; the authors propose using the complete problem-solving transcripts (rather than just scores) of models on saturated benchmarks to evaluate capabilities, and tested methods including pairwise judging, single-transcript scoring, and prediction based on deterministic features or embeddings by simulating saturation scenarios with historical model and benchmark data; results show that pairwise judging performs best, reducing error by 5.3 times compared to a baseline using only saturated scores, and capability signals are distributed across specific information channels such as reasoning and tool calls, but some methods have limited generalization under distribution shifts.
The paper introduces RealtimeBench, containing 180 tasks across six game categories, designed to evaluate LLM agents while games are running continuously. Experiments show that GPT-6.1-sol outperforms Jev in paused environments but lags behind when running in real time. The authors propose the Two-Timescale Offline Recursive Self-Improvement (TORSI) method, where a slow teacher analyzes failures in paused games and generates instructions, while a fast student executes them during runtime. After two rounds, Jev's average test score increased from 13.51 to 24.19 (on a 0–100 progress scale), and instruction transfer improved the performance of three other models.
Paper arXiv:2610.20639 indicates that the phenomenon where Reinforcement Learning with Verifiable Rewards (RLVR) improves MMLU accuracy after mathematical training partially stems from models learning to place answers in \boxed{} format rather than truly enhancing reasoning capabilities; the authors designed a new answer extractor and found that 82% of the cross-domain multiple-choice question gains disappeared across four pairs of 7B SimpleRL baseline-RL models, indicating that these gains were actually format gains; independent LLM reading and prompts forcing output in \boxed{} format further verified this conclusion.
The paper presents DeepUnfolding_Agent, a pattern-driven framework based on FastAPI that provides deep unfolding-based adaptive control as an external service. It separates control intelligence from domain execution through components such as a strategy registry, diagnostic adapters, and a controller factory, supporting execution paths via classical methods, Qiskit Aer, and IBM quantum hardware. The authors validate its reusability and traceability through multi-dimensional testing and scenarios in finance, healthcare, genomics, transportation, and quantum simulation.
Lobby Domains is an independent domain search tool developed by Dale Hurley. It uses Chrome's built-in local AI models to generate domain candidates without relying on cloud AI; domain verification calls the registry's HTTPS RDAP interface directly, bypassing Lobby Domains' proxy; it supports sharing search links, saving shortlists locally, and exporting snapshots; there are no analytics scripts, cookies, accounts, or remote fonts; optional non-tracking registrar redirect links are provided; a WebMCP interface allows compatible agents such as ChatGPT or Claude to invoke its functions.
The Doubao Phone Assistant published the SAEP protocol rejection list after the end of the 30-day rule publicity period. A total of 155 apps refused to authorize AI automated operations, including WeChat, Taobao, Alipay, Meituan, Honor of Kings, and AI applications such as DeepSeek and Kimi; the second-generation phone NaviX Ultra switched to a 'knock-to-access' strategy, prioritizing MCP/A2A protocols with simulated clicks as a fallback, but mainstream platforms still generally refuse; the article analyzes that the core motivation for major companies refusing integration is to protect user relationship chains, recommendation rights, advertising duration, and data ownership, pointing out that the current contradiction is essentially a game of redistributing default distribution rights in the mobile internet.
Amazon has continuously banned shopping agents from Grok Bot, Meta Muse, Perplexity Comet, OpenAI, and Google, citing violations of terms of service and potential security risks. Although a preliminary injunction was obtained in the previous lawsuit against Perplexity, it was overturned by the Ninth Circuit Court of Appeals, which ruled that agent operations are an extension of user behavior. Meanwhile, Amazon is vigorously promoting its self-built shopping agent Rufus and the upgraded Alexa for Shopping, using the Buy for Me feature to allow AI to place orders on behalf of users at external merchants. Shopify, conversely, opens its product catalog to agent integration, highlighting the ecological divergence in the competition for agent entry points among platforms.
Strata released version v0.1.44, adding native support for Apple Silicon Macs (M1–M5, requiring macOS 15+), using the Metal backend; this is currently experimental support. It added the lightweight Qwen3.6-35B-A3B model, which runs only on NVIDIA/AMD GPUs, offering three quantized versions including UD-Q4_K_XL, with memory requirements starting from 32GB. The existing Qwen3.8-Flash-Next 125B model remains supported, and both models support MTP acceleration.
The author uses the XIAO ESP32-S3 development board to implement an iPhone on-device AI control agent: the ESP32 creates a Wi-Fi hotspot as an AirPlay receiver, forwarding H.264 video streams via Bluetooth to a mobile app; the app calls any vision model compatible with the OpenAI interface to analyze the screen and generate operation commands, which are then sent back to the ESP32 via Bluetooth, where it simulates USB mouse/keyboard inputs combined with iOS AssistiveTouch to execute clicks; supports automatic calibration, keyframe transmission, and pointer zeroing mechanisms; demo videos and code are open-sourced.
The article notes that Kimi launched its personal agent Kimi Claw as early as February 2026, seven months ahead of Meta Muse, yet it remained in Beta stage with slowed updates after seven months. The core reason is that the C-end monthly package at 199 yuan struggles to cover the exponential computing consumption caused by complex Agent tasks, leading to a financial model inversion, while over 70% of Kimi's revenue comes from B-end API calls. Big tech companies like Alibaba, Tencent, and ByteDance are betting on Agents as the third-generation entry point, using GMV commissions, private domain LTV, and content traffic protection respectively as their backward discounting logic. The article further proposes that payment authorization is the ultimate destination for Agents; ByteDance's launch of utility bill payment features aims to cultivate users' mindset for binding cards and automatic deductions within AI, but faces payment barriers constructed by WeChat Pay and Alipay.
Beeline has launched the Multiplayer Agent Harness, enabling humans and multiple coding agents (such as Claude, GPT, Gemini, and DeepSeek) to collaborate within the same Room. Users can log in via GitHub on mobile devices, run `npx usebeeline connect` to integrate local proxies, and then trigger agents through mentions to complete planning, coding, checking, and submitting changes for approval in the corner.
Booko is a cross-platform book excerpt management tool independently developed by author jiayifun. It supports note imports from WeChat Reading (via one-click import through Skill API), Kindle, Apple Books, Douban Read, Dedao, and Zhangyue. It has launched on iPhone, iPad, Mac, and Apple Watch. Features include quick capture via an 'Inbox', many-to-many organization via 'Collections', 'Random Review', and persistent display via desktop widgets/Apple Watch faces. The Mac version adopts a three-column design and supports keyboard-centric workflows. All features were implemented through AI Coding collaboration, with local storage + iCloud sync. Basic functions are free, while the Pro version is a one-time purchase for 88 yuan.
bimbaloo resumed operation of the "Qiaosi Toolbox" WeChat mini program, which was abandoned 5 years ago. Based on the Taro + Go + MySQL tech stack, it aggregates 234 tools covering 6 major categories. The AI intelligent module includes 16 image capabilities (OCR/cutout/license plate recognition, etc.) and 11 text capabilities (summarization/error correction/translation, etc.).
NVIDIA, in collaboration with the AI4Bharat lab at IIT Madras and Bodhan AI, released Nemotron-pretraining-Indic-v1, a synthetic pretraining corpus covering 22 Indian constitutional languages, totaling 2.8 trillion tokens (2601 billion native script tokens + 224.6 billion romanized tokens), counted using the Gemma-4 tokenizer; simultaneously open-sourced a GPU-accelerated data generation pipeline based on NeMo Data Designer, including document-level translation, MQM-inspired quality assessment, feedback-driven repair, and dual-track script processing (native script + romanization); this corpus was used for Hindi continual pretraining of Nemotron 3.5 Super VL (with 10% of the 10 billion tokens being Indic mixed data), improving performance by 4.98–9.82 percentage points across 5 Hindi benchmarks such as MILU and Indic MMLU-Pro; the corpus has been uploaded to Hugging Face, supporting streaming loading configured by language-script-region (e.g., hi-Deva-IN), with accompanying open-source plugins providing adaptable prompt, rewriting, and script mapping functions.
Xiaomi has open-sourced the spatial audio understanding large model MiDashengLM-Spatial. Based on MiDashengLM-7B, it adds spatial perception and reasoning capabilities, supporting sound content understanding as well as sound source localization and spatial distribution modeling. The model weights, inference code, technical report, evaluation code, and demo page are all publicly available under the Apache 2.0 license.
On June 26, 2026, at its 10th-anniversary launch event, Guanyuan Data introduced its new product, DecideX. Positioned as a decision intelligence solution, its architecture includes an enterprise-grade data foundation, a decision context layer, the DecideX Runtime intelligent hub, and a human-machine collaboration layer. Its core innovation lies in structuring business judgment rules (such as discount strategies and order suspension conditions) scattered across systems like meeting minutes and ERPs into callable decision contexts and Skills modules. Meanwhile, the official website continues to list the Question Answering Agent and Insight Agent released in 2025 for sale alongside it. The three implementation cases disclosed by the company (Unilever Hefei Logistics Center, Laiyifen, and Qiutian Manman) all originate from this launch event, with no independent third-party verification observed. Guanyuan states that delivery adopts the FDE on-site model, emphasizing the need for manual accumulation of customers' implicit business logic, which differs significantly from traditional BI software delivery methods.