AI news

New paper from Google DeepMind on training models to argue well about moral dilemmas. The trained model became more decisive and stopped changing its answer when options were reordered. Sycophancy and reframing still moved its judgments. The setup uses real dilemmas from

View full image

DecisionPivot A diagnostic benchmark of 48 tasks and 40,033 items that shows how item accuracy alone hides large differences between decision models in LLM systems

"BeNice! Closing Looping’s Inference-Memory Gap via KV-Sharing" Looped Transformers reuse the same layers multiple times to improve performance without adding parameters, but their KV-cache memory still grows with every loop. This paper proposes a training method that lets different loops reuse the same KV-cache without losing accuracy, by combining stochastic-depth pretraining with a fine-tuning objective that makes earlier loops compatible with the final loop's cache. On a 1.3B model, KV-sharing reduces total GPU memory usage by 66.2% in a tested configuration. In simulated long-context workloads, it also improves throughput by up to 23.6% and reduces 90pt time-to-first-token by 83.6%, while preserving model quality. alphaxiv.org/abs/2610.31633

View full image

"Agentic World Action Models" What if robots could write their own code to plan actions, while also predicting how the physical world will respond? NVIDIA's new AW-0 combines three capabilities in one model: generating reusable skill programs, predicting future video frames, and producing robot actions. All three are generated together, which enables code-based planning to guide physical execution. On RoboTwin 2.0, AW-0 achieves 50.7% average success, compared to 41.3% for π0.5. Across eight real-world manipulation tasks, it scores 71.1% vs 46.0% for the strongest baseline, including tasks like folding towels and solving a Rubik's Cube. alphaxiv.org/abs/2610.30136

View full image

Highly recommended read on looped and depth-grown models. There is a lot of interest in looped and depth-grown language models for reasoning. The question here is what reusing layers does to in-context learning. At 3B parameters, both grown and looped models gain more from

Translated from Chinese · Show original

AI images never get the pose right? An open-source tool lets you pose them yourself AI images keep getting poses wrong and things in the wrong place? ArtCraft lets you first set up characters, props and the camera in 3D, then has AI draw the image and generate the video. Github: github.com/storytold/artcraft

Translated from Chinese · Show original

😱 Unbeatable! A 30B-class open-source agent pushed to the ceiling! Nanbeige4.3-33B is officially live! It outperforms Qwen-3.8 27B! It achieves SOTA agent performance among open-source models of about 30B parameters, crushing complex coding and professional collaboration tasks across the board. It is both a ready-to-use strong teacher model and can be plugged into real agent frameworks at low cost to get work done. What makes it great: 🔹Terminal-Bench 2.1 80.9 / 4.0 23.8 / DeepSWE 65.5, leading Qwen3.8-27B and others by a wide margin on several coding agent benchmarks 🔹Leads across collaboration tasks such as Automation Bench, JobBench and DeepPlanning, covering real workplace scenarios in many industries 🔹512K ultra-long context + cost-efficient inference, usable directly as a teacher to distill later small models such as Nanbeige4.3-3B How to use it: 1️⃣Go straight to HF to download the model: huggingface.co/Nanbeige/Nanbeig… 2️⃣Load with Transformers (trust_remote_code=True), enable enable_thinking in the chat template (for agent scenarios, preserve_thinking=True is recommended) 3️⃣Recommended deployment: vLLM / SGLang + --reasoning-parser qwen3 --tool-call-parser qwen3_coder; for agent tasks temp=1.0, top_p=0.95 4️⃣Plug it into your agent framework (OpenClaw or any mainstream harness works) and start on real tasks #LLM #Agent #开源模型 #Nanbeige #AgenticAI #30B模型 #Qwen3827B

GameScaling: miHoYo's study on scaling embodied agents in games Continued pre-training from 0.8B to 27B on 1,600 hours of gameplay across 80 diverse virtual worlds. Models are open-source on Hugging Face.

Tencent just released WeVisOne-VL on Hugging Face One vision-language model for detection, counting, pointing, OCR, segmentation and alpha matting — from objects to pixels.

The complete reasoning-from-scratch video series is now up (and linked) here: magazine.sebastianraschka.com/p… (Taking a short break but will be back with LLM model and research updates in November!)

CoEvo-RSI: Aligning evolving harnesses with model capabilities A recursive self-improving framework that uses bidirectional historical inheritance to solve brain-body mismatch, letting a 9B model outperform much larger frontier models across diverse agent tasks.

When should the cloud execute for the edge agent, and when should it just advise? We present a systematic study of edge-cloud collaboration in long-horizon agent execution. Using EC²-Bench, we evaluate four collaboration strategies across 128 executable tasks, jointly measuring task completion, resource cost, and privacy exposure. Here are three key findings: 1️⃣ Small cloud budget, big gains.(cost-efficiency). With Qwen3.5-9B on-device and GLM-5.3 in the cloud, edge-only scores 39.12; the best collaboration reaches 60.46 — about 73% of cloud-only (82.70) — at only $8.73, roughly 15% of cloud-only's $56.42. 2️⃣ Performance converges, privacy doesn't. On Qwen3.7-Plus, best collaboration hits 76.84 vs 82.70 cloud-only; strategy gaps shrink to just 3.67 completion points, yet privacy exposure spans 66.23 percentage points — and the best-performing strategy leaks the most. 3️⃣ The mode flips with edge capability. Weak edges do best with an external router that lets the cloud execute (2B…

JEPA learns what to keep from what it sees, and uses that to predict what comes next. But what about the details it can’t predict? Wouldn’t it be easier for the model to drop them? Look at this example Imagine a traffic light that changes color randomly. Its next color is unpredictable, but its current color helps predict whether a cyclist will stop or move. If we drop the light, the cyclist becomes harder to predict. ▪️ A recent research call this set-level view "predictive closure:" A feature’s value depends on its relationships with the other features the model represents. Keeping the light makes the cyclist easier to predict, but adds an unpredictable color to the model’s targets. Dropping it removes that target, but loses a useful clue. The researchers separate these two costs: 1. Aleatoric uncertainty: what remains uncertain even with the full current observation, like the light’s next random color. 2. Predictive leakage: useful information available in the observation bu…

Apple and UCLA present OpenVLHarness A training-free harness that gives multimodal LLMs a capability layer and persistent multimodal memory, boosting average scores by 9.7-12.6 points across 22 benchmarks.

1:50

Can LLMs improve their own harness? Self-improving harnesses are a hot topic. And one interesting question is whether LLMs can generate and improve them on their own. They find that two full rewrites of the harness raise held-out scores by only 2.80 points, and the 95% interval crosses zero. Each frontier model writes a complete harness from a fixed coding bootstrap, plays five development games through it, and then rewrites its whole codebase twice. Every version is scored on 20 games that never influence the rewrites. Across 27 lineages, first harnesses range from zero to near-ceiling. The revisions add very little. Gains on the development games overstate the held-out gains, and revisions that cut inference cost often lower the score. Human-steered systems still score higher on the same games, so better designs exist. The models just do not find them reliably on their own yet. arxiv.org/abs/2610.30499 Chat with Paper: academy.dair.ai/papers/can-llms…

Nice work on using skills more effectively for coding agents. (bookmark it) In this work, researchers treat skill granularity as a runtime decision. For each task, the agent can use primitive tool calls, a macro skill, or a macro skill that falls back to primitive steps when it fails. Fixed macro execution is efficient when the skill fits and brittle when it does not. The controller picks the level from task structure, execution state, and earlier macro failures, and a finite-state executor handles the fallback as an explicit transition instead of an informal retry. In a 12-task routing study, a learned router reaches 100% success where a hand-written heuristic fails. With macro-to-micro fallback, the agent recovers controlled macro failures that fixed granularity leaves unresolved. arxiv.org/abs/2610.31461 Chat with Paper: academy.dair.ai/papers/adaptive…

LEV: think deeper without a single extra token A decision model that loops its layers, scoring options after each pass. One checkpoint works at any depth. On JevBench, Lev-2.6B goes from 63.2% to 76.2% accuracy as depth increases from 1 to 4.

1:34

DeepSeek raised the bar, and SGLang pushed it even further 🧡 lmsys.org/blog/2026-10-2… Read our blog to learn more about the performance results, engineering optimizations, and how SGLang brings it all together in a production-ready serving stack! x.com/sgl_project/st…