Announcing Autoresearch Bench A benchmark for coding agents autonomously tackling research problems While existing coding benchmarks are saturating quickly, the difference between models in autoresearch loops is stark

Gemini 3.8 Flash Cyber gives defenders a decisive advantage with expert vulnerability detection and autonomous patching. Here’s how it helps teams stay ahead of emerging threats 🧵

For the past 3 months, I've been heads down building and extending our multimodal data capture rig for Physical AI training. What you're seeing here is a single synchronized capture from our system — RGB, depth (stereo IR camera) feed; real-time 3D hand reconstruction, and a 23×20 tactile pressure grid from sensor gloves, all aligned and timestamped together. Getting here took a lot of unglamorous iteration. We reworked the pipeline to deliver clean, gap-free depth at scale after identifying early depth frame drops caused by data throughput bottlenecks in our capture box. We built per-channel calibration models that convert raw resistance signals from our flex sensor gloves into accurate, drift-corrected force values — solving the creep behavior that made a steady grip falsely appear to fade over time. Even though we’ve stayed quiet on socials, word of mouth spread — our rigs have already been deployed in data factories, manufacturing lines, and homes across the US, Europe, China, and Southeast Asia by our customers and ourselves. In the next few days, I'll be sharing more about how we're tackling rights-cleared data capture from the hardware level, and our closed-loop data-model flywheel — using model evaluation results to identify gaps and guide what data our customers need. More soon, stay tuned…

We’re introducing a new capability to our latest Gemini models: agentic video understanding. This allows developers to process long-form video content with more accuracy, while using up to 88% less tokens. See how it works 🧵

Complementary Field-Effect Transistor (CFET) technology achieves major static random access memory (SRAM) scaling beyond traditional nanosheet gate-all-around (GAA) devices by vertically stacking nFET and pFET transistors. This folding eliminates lateral n-to-p separation spaces, reducing 6T-SRAM bitcell area by up to 40% to 55% compared to standard nanosheet or forksheet benchmarks. From the below.

Instead of choosing between training a world model and training a language conditioned robot policy, why not do both? LDA-1B s a new foundation model that is trained on 30,000 hours of human and robot interaction data. Part of the secret is that LDA-1B jointly learns forward dynamics, action prediction, and visual forecasting, all in a structured DINO latent space which avoids the pitfalls of redundant pixel-level prediction which isn’t necessarily aligned robot action. This approach works on both dexterous hands and simple robot grippers; it also generalizes across objects, tasks, and scenes. @JiangranLyu joins us to explain. Learn more on Episode 98 of RoboPapers, with @micoolcho and @chris_j_paxton!

Microsoft just released an executable benchmark for tool-using agents on Hugging Face ThinkingBox-Bench evaluates agents on 507 stateful business tasks across retail, travel, insurance, and more.

Banger paper from Microsoft. It's on agent reliability in real business workflows. (bookmark it) Thinkingbox is a sandbox with isolated MCP-compatible tool sessions, plus a benchmark of 507 policy-conditioned workflows across retail, hospitality, auto insurance, neobank IT, and consulting support. Every attempt is graded on the backend state the agent leaves behind. Executable checks accept valid trajectories and reject wrong, missing, or extra effects, so collateral damage counts against you. The strongest model reaches 65.36% pass@1 and 25.25% pass^20. Many failed trials terminate cleanly with valid state-changing tool calls. Watching the response or the tool call tells you very little about whether the task actually completed. Paper: Track more trending AI papers in our academy:

JSEF v1.10.0 is out 🎉 A model-agnostic LLM eval toolchain + 20+ new hard vulnerability samples. ✅ Swap any OpenAI/Anthropic-compatible model (GLM-5.3, Mythos, GPT…) ✅ One-click cross-model leaderboard (Recall/Precision/F1/accuracy) ✅ DeepSwe-style trials stability (Pass@1, N-run) ✅ Multi-state gate, JWT 4-step chain, async taint chain Star → #LLMSecurity# #AppSec# #WebSecurity# #JavaSecurity# #VulnerabilityDetection# #AISecurity#

🧵 New repo: Awesome Egocentric Dataset — a curated list of first-person vision datasets, refreshed beyond the classic Ego4D-era list. ✅ Recent 2023–2026 releases (Ego-Exo4D, EgoSchema, AEA, EgoExoLearn, EgoLife…) ✅ Dead/broken links removed, migrated ones fixed ✅ A section on egocentric-driven VLA papers (training source + data-processing), with first-principles + fishbone analysis 🚀 Open-source, contributions welcome. 🔗 #EgocentricVision# #FirstPersonVision# #VLA# #ComputerVision# #Datasets#

灵巧手的数据格式： task_info/task_327.json： { "episode_id": 649078, "task_id": 327, "task_name": "Picking items in Supermarket", "init_scene_text": "The robot is in front of the fruit shelf...", "lable_info": { "action_config": [ {"start_frame": 0, "end_frame": 435, "action_text": "Pick up onion from the shelf.", "skill": "Pick"}, {"start_frame": 435, "end_frame": 619, "action_text": "Place onion into the plastic bag.", "skill": "Place"} ] } } 亮点：指令不是整段一个，而是按帧区间切片 + 原子技能(skill)标签分层标注。 这比我们 language_instruction.json 只标一个 segment 更细——它把"拿起→放下"拆成两个子动作各配指令。 这正是高质量示教数据的做法。

1,000,000 rigid bodies. All simulated on the GPU. No GameObject per body. No CPU readback — the CPU never sees a single pose, or even how many got drawn. Bodies, broadphase, solver, culling and the instance stream all live on device. Then I dropped four tornadoes on it.

- Claude = coding. ($20/mo) - Supabase = backend. (Free) - Vercel = deploying. (Free) - Namecheap = domain. ($12/yr) - Stripe = payments. (2.9%/transaction) - GitHub = version control. (Free) - Resend = emails. (Free) - = AI job Agent （Free） - Clerk = auth. (Free) - Cloudflare = DNS. (Free) - PostHog = analytics. (Free) - Sentry = error tracking. (Free) - Upstash = Redis. (Free) - Pinecone = vector DB. (Free) Total monthly cost to run a startup: ~$20 There has never been a cheaper time to build.

“AREX: Towards a Recursively Self-Improving Agent for Deep Research” Deep research agents often fail not because they need more search, but because they don’t know which parts of an answer are already verified and which constraints are still unresolved. This paper turns research into a recursive loop: answer, verify constraint-by-constraint, preserve evidence, then refine only the weak parts. AREX adds learned context updates and step-aware RL, achieving extremely strong results with 4B and 122B-A10B models across various agentic benchmarks.

Inkling-small is out today! With SGLang, you can get 648 tok/s decode with DSpark (simulated acc len=4) and 288 tok/s w/o DSpark, under the same setup (8x @NVIDIAAI B200, TP 8, NVFP4, bs=1). What makes this model different is the size. 276B total with 12B active is a sweet spot for RL, and both LoRA and full-parameter training become well within reach. Miles is ready and verified for multimodal RL on Inkling-small, so you can turn your multimodal data into real capability gains. At ~1/4 the size, Inkling-small matches the bigger version in capability and even wins on some benchmarks. Run Inkling-small with SGLang, and customize it with Miles.

"Metis: Memory Foundation Model" Most AI agents still use memory as an external RAG-style module, so the model retrieves old text instead of actually remembering. This paper makes memory native to the Transformer. So past interactions are compressed into dynamic layer states and read through memory attention during normal forward passes. The model weights stay frozen at inference, but its memory state updates without gradients, giving the model persistent memory inside the backbone. Still early and lossy, but this is yet another paper with a big step toward agents that remember natively instead of outsourcing memory to a database.

🔥 𝐍𝐞𝐰 𝐛𝐥𝐨𝐠: 𝐓𝐨𝐰𝐚𝐫𝐝𝐬 𝐋𝐨𝐨𝐩𝐞𝐝 𝐌𝐨𝐝𝐞𝐥𝐬 𝐃𝐨𝐧𝐞 𝐑𝐢𝐠𝐡𝐭 — 𝐏𝐚𝐫𝐭 𝐈 Looped models reuse the same weights across depth, promising a better compute–parameter trade-off, especially for reasoning. 𝐁𝐮𝐭 𝟏) 𝐝𝐨 𝐭𝐡𝐞 𝐠𝐚𝐢𝐧𝐬 𝐬𝐮𝐫𝐯𝐢𝐯𝐞 𝐰𝐡𝐞𝐧 𝐛𝐨𝐭𝐡 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐚𝐧𝐝 𝐢𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 𝐅𝐋𝐎𝐏𝐬 𝐚𝐫𝐞 𝐦𝐚𝐭𝐜𝐡𝐞𝐝? 𝟐) 𝐀𝐧𝐝 𝐰𝐡𝐢𝐜𝐡 𝐚𝐫𝐜𝐡𝐢𝐭𝐞𝐜𝐭𝐮𝐫𝐚𝐥 𝐜𝐡𝐨𝐢𝐜𝐞𝐬 𝐚𝐜𝐭𝐮𝐚𝐥𝐥𝐲 𝐦𝐚𝐭𝐭𝐞𝐫? We run 𝐚𝐩𝐩𝐥𝐞𝐬-𝐭𝐨-𝐚𝐩𝐩𝐥𝐞𝐬 ablations spanning Ouro to Huginn. Huginn performs better overall, with the largest gains coming from the loop-in-the-middle (sandwich) design and input injection, though they provide different benefits. Trained on 𝟓𝟎𝟎𝐁 tokens, an 𝟖𝐁-𝐀𝟎.𝟖𝐁 Huginn MoE approaches or surpasses a 𝟑𝟐𝐁-𝐀𝟑.𝟐𝐁 feedforward MoE on several reasoning benchmarks, including GSM8K (83.6% vs. 80.8%), while using 𝟕𝟓% 𝐟𝐞𝐰𝐞𝐫 resident parameters under 𝐦𝐚𝐭𝐜𝐡𝐞𝐝 𝐭𝐫𝐚𝐢𝐧𝐢𝐧𝐠 𝐚𝐧𝐝 𝐢𝐧𝐟𝐞𝐫𝐞𝐧𝐜𝐞 FLOPs. More details and the blog link in the thread ↓