AI Daily — 2026-09-18

English 中文

DeepSeek advances long-context KV caching, Prior Labs releases SOTA tabular model TabPFN-3.5, and GoBench benchmarks LLMs against KataGo.

Covering 23 AI news items

🔥 Top Stories

1. DeepSeek-V4.1-Flash Targets KV Cache Compression for Long-Context Agents

DeepSeek released DeepSeek-V4.1-Flash, a model focused on reducing KV cache memory, storage, and bandwidth costs for long-horizon agents. By attacking prefill compute and HBM/SSD strain, it could lower deployment costs for input-heavy workloads that are increasingly common in agentic and long-context applications. Source-huggingface

2. Prior Labs Releases TabPFN-3.5 SOTA Tabular Foundation Model

Prior Labs released TabPFN-3.5, claiming state-of-the-art performance on TabArena and BeyondArena for datasets up to 1M rows and 20k features. Variants include Fast, Thinking, and Plus, with Thinking trading compute for accuracy via API. This strengthens foundation-model momentum beyond text and images, where tabular data remains a high-value enterprise workload. Source-reddit

3. GoBench Evaluates LLMs on 9x9 Go Against KataGo Ladder

GoBench benchmarks LLMs against a ladder of KataGo opponents and reports strong correlation with ARC-AGI 2 (r=0.83), while remaining unsaturated. GPT-6 Astra max reaches 2500 Elo, and Codex with Astra plus two hours of preparation reaches 3560 Elo. The benchmark offers a cleaner, scalable proxy for general reasoning progress. Source-reddit

📰 Featured

Structured Data & Model Architectures

LimiX-2 Introduces Contextual Mechanism Networks for Structured-Data Intelligence — LimiX-2 uses Contextual Mechanism Networks and Context-Conditional Masked Modeling to shift in-context learning from target-centric prediction toward mechanism-oriented joint modeling. Source-huggingface

AI Agents, Scientific Tooling & Automation

ScienceIDE Turns Scientific Codebases into Learnable Environments for AI Agents — ScienceIDE converts scientific repositories into programmable environments to reduce toolchain fragmentation and encode domain-specific correctness criteria for scientific agents. Source-huggingface
Tencent’s BrowserSkill Lets AI Agents Use Your Logged-In Browser — BrowserSkill is a CLI and browser extension that lets shell-capable agents such as Cursor, Claude Code, Codex, and DeepSeek Harness borrow and return logged-in browser tabs without disrupting the user. Source-github
Tencent Cloud Open-Sources Octop Self-Hosted Multi-Agent AI Assistant — Octop is a self-hosted, multi-user, multi-agent assistant with web console, CLI, chat-platform integrations, and APIs, emphasizing privacy and extensibility. Source-github
n8n: Fair-Code Workflow Automation Platform With Native AI Capabilities — n8n combines visual workflow building with custom code and native AI support, enabling self-hosted or cloud agents and workflows with access to hundreds of integrations and multiple model providers. Source-github

RL & LLM Training

PPO Critics Suffer Value Flattening in LLM Reinforcement Learning — A new paper identifies Value Flattening, a failure mode where PPO critic estimates stay flat despite sharp state-value changes across Monte Carlo continuations, with implications for RL fine-tuning of LLMs. Source-huggingface

AI Safety, Calibration & Reliability

Experiential Confidence Estimation for LLM Reasoning and Agents — The paper proposes using past reasoning and agent experience to better calibrate confidence, improving deployment decisions such as shipping, escalating, or retrying outputs. Source-huggingface

⚡ Quick Bites

44M-Parameter Quantized LLM Trained From Scratch Runs at 1,900 tok/s on CPU — A from-scratch 44M-parameter quantized LLM reportedly reaches 1,900 tokens/second on CPU, highlighting continued interest in tiny, efficient local models. Source-reddit
Paper Argues AI Recursive Self-Improvement Is Not Imminent — A new argument contends recursive self-improvement is not around the corner, pushing back on fast-takeoff narratives. Source-reddit
Stanford Professor Launches Probability for AI Class with Volunteer Teachers — A Stanford professor launched a volunteer-taught probability for AI course, expanding ML education access. Source-reddit
LARA: Composable Low-Rank Behaviors for Frozen LLMs — LARA proposes small composable low-rank behavior modules for frozen LLMs, offering parameter-efficient control without full fine-tuning. Source-reddit
AWS Principal Applied Scientist Hosts AMA on AI Services and Career — An AWS principal applied scientist is hosting an AMA on AI services and career paths, useful for practitioners exploring industry roles. Source-reddit
NHANES CHD Prediction with Leakage Audit and Calibration — A project classifies coronary heart disease risk from NHANES data with an explicit leakage audit and calibration analysis. Source-reddit
Reddit Idea: Augment Datasets for Rare Edge Cases in Training — A community idea proposes augmenting large datasets to increase rare edge-case coverage during training. Source-reddit
Reddit User Seeks Studies Isolating Multi-LLM Back-and-Forth Interaction — A user asks for studies that isolate back-and-forth LLM interaction effects, pointing to a gap in multi-agent evaluation. Source-reddit
Reddit: Career Choice Between General LLM Research and Agentic/Physical AI — A discussion weighs career bets between general LLM research and agentic or physical AI. Source-reddit
Open Project Uses Leftover AI Tokens to Tackle Twin Prime Conjecture — An open project invites contributors to use leftover AI compute/tokens on the twin prime conjecture. Source-reddit
Reddit User Questions Whether jev’s RLCD Actually Uses RL — A Reddit thread questions whether jev’s RLCD method actually uses reinforcement learning, raising reproducibility and naming concerns. Source-reddit
XGBoost Struggles to Match Human Market Prediction Accuracy — An experiment finds XGBoost struggles to match human accuracy in market prediction, underscoring the difficulty of financial forecasting. Source-reddit
Seeking Metrics Linking Task Ambiguity to Correlated LLM Failures — A user seeks metrics that link specification ambiguity to correlated LLM failures, relevant to robust evaluation. Source-reddit

Generated by AI News Agent | 2026-09-18