llm

A study reports that large language models are significantly better at detecting mistakes in academic papers than trained human reviewers, suggesting a potential shift in how peer review quality assurance could be assisted by AI.

Sente Labs open-sourced extensible-mcp, an MCP proxy that dynamically loads MCP servers without polluting the LLM context while enforcing zero-trust policy controls (Rego, CEL, and now Cedar) over tool calls. Tool definitions live in a queryable vector store, and policies and credentials are kept outside the LLM's reach, distinguishing user intent from granted rights.

Google DeepMind 论文提出 CALLIOPE，用 LLM 基于关键词（而非嵌入聚类）构建主题树，在严格隐私约束下生成更清晰的主题地图，人类评分在所有数据集上优于 URANIA。

A Wired reporter deployed a swarm of AI agents to manage a fantasy Premier League squad, finding they still underperform humans, while new research shows frontier models like GPT-5 increasingly excel at real-world prediction tasks and forecasting benchmarks such as Prophet Arena.

A user shared a jailbreak prompt called "iLang Soul v1.1," claiming it bypassed Grok Bot's restrictions in under 20 lines — easier than trying with local QWen3:8b.

The author built WhisperNow, a voice-to-text app, and fine-tuned three LLMs (LFM2.5 1.2B/2.6B, Qwen3.5 4B) on a 66k-row dataset to clean up disfluencies locally, outperforming several cloud models in factual accuracy for short dictations.

Simon Willison argues that October 2026 marks an inflection point for vibe-coded game development, as Claude Opus 5.5 now produces genuinely fun browser games with minimal guidance, outpacing GPT-6 Astra in game design sense.

A practitioner discusses the compounding error problem in multi-step AI agents (95% per-step reliability drops to ~36% at 20 steps) and advocates for keeping control flow in code while reserving autonomy for judgment calls, asking the community for fully autonomous production examples and reliability tips.

A Chinese-language explainer on reinforcement learning terminology for LLM readers, using the "Is OpenAI open?" example to walk through concepts such as policy, state, action, reward, return, value, advantage, and GAE, and explaining how scores are attributed to individual tokens when a judge only rates an entire response. Set against the backdrop of DeepSeek-R1-Zero's RLVR training, this is the fourth installment in the post-training series.

作者发布了开源项目 DecisionFormer，可将普通 LLM 转换为直接预测选择而非生成文本的决策模型，适用于决策模型实验与推理方式探索。

OpenLLM 是一款开发者工具，让你把订阅、API 密钥和本地模型接入 Codex、任意编码代理，甚至 Mac 上的 ChatGPT 应用，突破限额后可无缝继续，并支持从任何地方启动编码会话。

提出了ETAMem多视角记忆框架，通过解耦异构对话信息并构建Event、Affect、Time三个互补的层次化视图，提升长期对话智能体的记忆检索效果，在LoCoMo和LongMemEval-S上分别实现12.46%和7.96%的相对准确率提升。

该论文提出 PERPL 框架，通过 LLM 模拟读者的相对偏好来学习个性化事实核查解释改写，在保持事实忠实性的同时优于更大的通用 LLM。

Researchers present the first open Speech Language Model for Basque, based on the LLaMA-Omni architecture with a Latxa backbone, using translated and TTS-synthesized data to overcome low-resource scarcity for end-to-end spoken interaction.

This paper proposes a small candidate formalizer trained on offline teacher-generated lists to bridge the interface gap between open-ended LLM workloads and fast decision models, reporting up to 52.6% task success on CALVIN-D with 2.09s end-to-end latency, compared to 11.8% without the formalizer and slower baselines like Qwen3.5-27B.

本文提出了一种基于标签保持扰动的输入歧义检测框架，先检测再消歧，包含黑盒规则库与白盒重要性引导两种检测器，在四类问答数据集上优于输出侧不确定性基线与LLM改写扰动基线。

该论文提出 LM-DSB，一个结合 RAG 与本地部署 LLM 的轻量级多智能体系统，用于自动生成建筑工地每日安全简报，在消费级硬件上实现了接近人工质量的文档生成与专家认可。

FinLLM-KG 是一个模型无关的框架，将因果 LLM 与金融知识图谱结合，从 SEC 10-K 文件中注入公司级、时间对齐的知识以改进金融情绪分析与算法交易，同时提供可解释的预测依据。

MARQ is a modular, LLM-agnostic prototype that infuses expert-derived systems engineering knowledge into LLM-based requirement quality assessment, showing significant F1 improvements over off-the-shelf baselines on an expert-labeled NASA requirement set across Claude Opus 5, GPT-5.6 Sol, and Gemma 3 12B, though gains are model-dependent.

FedSEAM 提出一种联邦持续学习框架，通过子空间专家共识合并（SEAM）在客户端之间选择性共享稀疏专家知识，缓解路由漂移与合并擦除导致的灾难性遗忘，并在 TRACE 基准上使用 Qwen3-4B 和 Llama-3.2-3B 进行了评估。