2026-07-17

Their verified account. English version just released:

Structured pruning is a hardware-friendly way to compress LLMs, but it is mostly validated on multiple-choice recognition tasks, while the same compressed checkpoints can collapse on the free-form generation that deployment actually requires. Two observations trace this gap. First, greedy pass@1 nearly vanishes after compression, yet pass@k recovers substantially under repeated sampling: useful generations are demoted, not erased. Second, the recoverable regime fails mainly through suffix repeti

As Large Language Models (LLMs) evolve into autonomous agents, the need for unified evaluation infrastructure becomes critical. However, current evaluation pipelines remain highly fragmented and tightly coupled, hindering reproducibility and causing redundant engineering. To address this, we introduce AgentCompass, an open-source, lightweight, and extensible infrastructure for evaluating LLM-based agents. AgentCompass organizes the evaluation process around three independent components, namely B

Cloudflare-native Email Infrastructure for Developers Discussion | Link

The gap is no longer measured in months. Open-weight models are catching up in real time, and Kimi K3 is already performing near Fable level. At this point, open models are not simply following the frontier anymore. They are on the verge of overtaking it. Kimi K3 may be one of the clearest signs yet that the closed-model lead is about to disappear.

We ran the same Qwen3.6-27B locally three ways on one RTX 6000: baseline, MTP, DFlash. The tasks were: quicksort, write a Steam library in JSON, solve a logic puzzle and write a sci-fi story. Outputs: Baseline: 44 tok/s · 1.00x MTP: 65 tok/s · 1.45x · 71% accepted DFlash: 98 tok/s · 2.20x · 30% accepted DFlash drafts 15 tokens in a row, so it flies through repetitive or structured stuff where long runs actually stick, like JSON (152 tok/s, 3.4x). On creative text most of the guesses are wrong, s

We introduce OvisOCR2, a 0.8B document parsing model. OvisOCR2 is designed as an end-to-end parser: given a document page image, it generates a Markdown representation in natural reading order, covering text, formulas, tables, and visual regions. We build a data engine that combines filtered real-document annotations with synthetic pages whose rendered images and Markdown targets are derived from the same HTML source. The training recipe includes supervised fine-tuning, reinforcement learning on

Google says these changes could endanger user privacy and security.

The company endorsed landmark AI transparency laws in California and New York last year, but its head of US state and local policy says they may already be outdated.

With this new update, Google is expanding AI Mode beyond answering questions and into completing tasks across the apps they use regularly.

I ported modern Linux to a ten-year-old HTC QWERTY phone as a handheld terminal. I wrote it up here: You can see the code here: Comments URL: Points: 35 # Comments: 6

为破解具身智能行业发展瓶颈构建了新一代“进化底座”

arXiv:2607.13049v1 Announce Type: new Abstract: Foundation models have given robots a sophisticated brain for complex decision-making, yet deploying that intelligence into a physical platform still demands tedious, expert-driven calibration. This deployment gap, the robot's spinal cord, remains a primary bottleneck to scalable Embodied AI. Hence, we propose SPINE (Scalable Physical Integration with ageNtic Expertise): an agentic framework for systematically debugging and deploying bimanual robot

Exploration is essential for reliable autonomy in multi-agent systems, yet it remains unclear whether large language model (LLM) agents can explore effectively when interacting with one another. We show that modern LLM agents fail to do so, often exhibiting myopic and polarized interaction patterns that lead to suboptimal coordination and increased regret. We formalize this challenge as the Multi-Agent Exploration problem, modeling it as a partially observable stochastic game (POSG) problem in w

Turn company operations into a living map for agents Discussion | Link

Build and hire autonomous AI employees from a single prompt Discussion | Link

Elon Musk's xAI files first lawsuit against Grok user accused of making child sex images.

Coca Cola said dairy production at its Fairlife unit will "remain suspended" in the United States following a hack.

One period tracker app tested by Mozilla was 'squeaky clean,' while another app was seen sharing users' health data with an analytics company, underscoring vast differences in user privacy among these apps.

Google already baked NotebookLM into Gemini, now it's changing the name to reflect the tighter integration

arXiv:2607.13069v1 Announce Type: new Abstract: Large language models produce chain-of-thought (CoT) reasoning that appears logically sound yet may not genuinely depend on its stated premises. We introduce interventional grounding audits, a black-box, step-level test of premise dependency: we intervene on a single premise by substituting its target predicate with a fresh symbol, re-run the model, and check whether each reasoning step's normalized conclusion (canonical predicate form) changes. We

Image guardrails are typically trained and evaluated under a fixed safety policy, implicitly treating safety as an intrinsic property of an image. Real deployments are different: the same image may be allowed in one product, restricted in another, and newly disallowed when a policy boundary changes. We study policy-adaptive image guardrailing, where a model must decide whether an image violates the currently supplied policy and generalize to held-out policy definitions. We introduce PolicyShiftB

Unbelievable to see kimi k3 beat frontier models that were 'too dangerous' for public use.

World Action Models (WAMs) improve robot policy learning by jointly modeling actions and future visual observations, using future scene evolution as dense supervision for physically grounded action generation. However, a common design in existing WAMs is to explicitly generate future videos at inference time, incurring substantial computational overhead and hindering real-time closed-loop deployment. GigaWorld-Policy addresses this issue with an action-centered formulation, where future visual d

Vision Transformers (ViTs) are known to exhibit high-norm patch-token outliers that degrade feature map quality, a problem effectively mitigated by register tokens. As diffusion models increasingly adopt transformer architectures and move toward pixel-space training, they become closer in form to ViTs, raising the question of whether register tokens are also useful for Diffusion Transformers (DiTs). In this work, we show that DiTs differ from ViTs in a key respect: they do not exhibit patch-toke

Cars24 uses OpenAI-powered voice and chat agents to handle 1M+ monthly conversation minutes, recover 12% of lost leads, and bring agentic workflows to teams across the company.

Large language model (LLM) agents are beginning to automate machine learning engineering (MLE) by coupling planning, code execution, debugging, and empirical feedback. Translating this capability to medical imaging remains difficult because each task imposes modality-specific experimentation and strict requirements for validation protocols and prediction artifacts. Here we introduce AMID, an autonomous multi-agent framework for medical imaging model development. AMID first proposes Data-Conditio

AngelSlim/Hy3-GGUF at hugging face (I am not affiliated just testing) I dont really make posts but I wanted to make sure everyone is aware that there is now a 1bit quant of Hy3 and now that models are getting larger I wanted to test how it behaves. Its a normal iq1m quant by compression it takes it down to 89 gigs at the smallest end. And mostly I am testing since model are going to get larger and larger and wanted to see how coherent the smallest version still are. And Im surprised honestly. He

I’m sick of “opt-out” toggles for automatically enabled generative AI features. It’s past time to make “opt in” the default setting for sensitive features.

Netflix says roughly 300 titles on its platform used generative AI, most of which occurred in post-production. The streaming service revealed the news in its second-quarter earnings report released on Thursday, saying it's "increasingly leveraging these tools to deliver higher quality output more quickly and at a lower cost." It also provided some examples of […]

New York Governor Kathy Hochul might have just signed a moratorium on new AI data centers in the state, but she's not against using the technology herself. During an interview with Bloomberg's Odd Lots podcast, Hochul said that her team is using "AI to analyze every single rule, regulation, [and] policy" to check for outdated […]

Today, the European Union ordered Google to give its AI rivals greater access to Android, the open-source operating system that powers billions of devices worldwide. The demand is hardly surprising. It may look like a defeat on paper for Google, which has spent years resisting exactly this kind of access, but it is a regulatory […]

arXiv:2607.13292v1 Announce Type: new Abstract: Autoformalization translates informal natural language into formal, machine-verifiable languages. While most work focuses on individual statements, real formalization efforts are inherently theory-level: they require an entire web of axioms, definitions, and lemmas before target theorems can even be stated. In this position paper, we argue for theory-level autoformalization: formalizing complete theories, including all their inter-dependencies, as

Lila is betting that science, not the internet, is the last untapped source of training data. We went to find out what that actually looks like in a room full of robots.

arXiv:2607.13172v1 Announce Type: new Abstract: We address the problem of safely training an agent policy and deploying a good and safe policy, in settings where the environment dynamics are unknown and no suitable reward function is available. In the context of safety-critical environments, we consider traditional reinforcement learning impractical and resort to the resource of human input. We introduce DROPJ, a human-centred method for both safe training and deployment. We first learn a world

arXiv:2607.13220v1 Announce Type: new Abstract: Most AI-for-science systems focus on scaling a single reasoning process through better models, larger context windows, long-horizon agentic execution, or digital co-scientists working with one principal user. However, challenging scientific problems are rarely solved by one reasoner alone. They are solved by teams whose members bring different priors, experimental backgrounds, tacit knowledge, and domain-trained intuitions. The open problem is ther

Hierarchical Interest Representation is a research area for Meta Ads. We’re exploring an upstream representation layer over the universe of Ads entities – users, advertisers, products, services – learning unified embeddings that connect users’ inferred interests with the breadth of what advertisers offer in their deep funnel ads. The innovations in Hierarchical Interest Representation are [...] Read More... The post Exploring Hierarchical Interest Representation For Meta Ads Deep Funnel Optimiza

Empty your backlog. Pay only when it ships Discussion | Link

I’ve always had this idea but can’t prove it. I think Anthropic and OpenAI don’t really have any secret sauce, their moat is just scale. Rumor has it Opus has 5T parameters and Mythos/Fable are 10T parameter models, while open models stayed under 1T for a long time. Only recently was that ceiling broken by DeepSeek V4 and now Kimi K3, and we’ve seen a significant jump in performance as parameter size increased. What do you think?