Marktechpost AI

Meet Blume: An Open-Source, Zero-Config Documentation Framework That Ships AI-Ready Docs From a Markdown Folder Drop Markdown into a folder, run one command, and you get a full docs site. No app boilerplate to write or maintain. How it works: the CLI scans your content and generates a hidden Astro + Vite project that it drives for you. The core theme ships no client-framework JS, so pages stay fast by default. Run blume eject any time to get a standalone Astro app. What's built in: → 30+ MDX components, nothing to import → Local search (Orama), on in dev and prod → llms.txt, .md at every URL, and a Copy/Open-in-chat menu → A hosted MCP server with 4 read-only tools for Claude Code, Cursor, and VS Code → OpenAPI/AsyncAPI reference via Scalar, changelog from GitHub Releases, 36 locales → OG images, sitemap, RSS, JSON-LD Needs Node.js 22.12+. Works with Bun, pnpm, npm, or yarn. MIT-licensed. Full analysis: marktechpost.com/2026/07/14/mee… Repo: github.com/haydenbleasel/… Project: useblume.dev

@MistralAI x.com/Marktechpost/s…

Announcing Robostral Navigate, our first model for embodied navigation: an 8B robotics navigation model that guides robots to autonomously perform tasks specified with natural language. Single RGB camera. State-of-the-art on R2R-CE.

Mistral AI Releases Robostral Navigate: An 8B Model Enabling Robots to Navigate Complex Environments Hitting 76.6% on R2R-CE With One RGB Camera. No LiDAR. No depth sensor. No multi-camera rig. Here's how it works. 👇 1. Pointing, not metric commands The model predicts the pixel coordinates of the next target in the camera view, plus the arrival orientation. Working in pixel space keeps it robust to camera intrinsics and world scale. When the target leaves the frame, it falls back to local displacements ("2m forward, 1.5m left, turn 25°"). 2. Grounding-first No open-source VLM base. It starts from Mistral's grounding model (pointing, counting, localization). Navigation emerges once the model knows where things are. → ~400,000 trajectories across 6,000 simulated scenes 3. Prefix-caching for training A tree-based attention mask packs a full episode into one sequence — all time steps in a single forward pass. → 22× fewer training tokens; months of training done in days 4. Online RL on top After supervised training, CISPO adds trial-and-error learning to fight distribution shift from behavior cloning. → +3.2% success rate from RL alone 5. The numbers (R2R-CE, Matterport3D) → 76.6% success on validation unseen → +9.7 pts over best single-camera approach → +4.5 pts over best depth/multi-camera system The key takeaway: state-of-the-art continuous VLN without a sensor stack — grounding-init, pixel-space actions, prefix-cached SFT, and online RL, on one RGB camera. Full analysis: marktechpost.com/2026/07/14/mis… Technical details: mistral.ai/news/robostral… @MistralAI @MistralDevs

Stanford Researchers Introduce TRACE: A Capability-Targeted Agentic Training System That Turns Recurrent Agent Failures Into Synthetic RL Environment Instead of RL on the whole task or generic synthetic data, it: → contrasts passed vs failed rollouts to find missing capabilities (kept only if Δ ≥ 0.20 and failure coverage ≥ 0.10) → synthesizes 1 verifiable RL env per capability → trains 1 LoRA adapter each via GRPO (base frozen) → composes them with token-level MoE routing (top-1) Results on Qwen3-30B-A3B: • τ²-Bench: 32.9 → 48.2 (+15.3) • SWE-bench Verified: 26.0 → 41.0 Pass@1 (+15) • beats GRPO & GEPA with <1/4 the rollouts And a 27B open-weight model hits 73.2% Pass@1 on SWE-bench Verified — above GPT-5.2-Codex (72.8%). Paper: arxiv.org/abs/2604.05336 Full analysis: marktechpost.com/2026/07/13/sta… Code: github.com/ScalingIntelli… Project page: hgkang02.github.io/trace-blog/ @TarunSures41845 @hangoo_kang

⚓ Prime Intellect released verifiers 0.2.0 — a rewritten "v1" core for agentic RL and evaluations, shipped under the new verifiers.v1 namespace. It breaks down an environment into 3 parts: → Taskset (what): data, tools, scoring → Harness (how): Codex, Terminus 2, ReAct, or your own → Runtime (where): subprocess, Docker, or sandbox The main piece is a verifiers-managed interception server. It sits between the harness and the inference server, proxies every request, and records the trace on the fly. It can also rewrite tool responses to mitigate reward hacks. Why it matters: → Any taskset runs under any compatible harness → Message-graph traces grow linearly, not quadratically like v0 → Branches make compaction + subagents native → train past the context window → Ships with full prime-rl training support Dialects: OpenAI Chat Completions, OpenAI Responses, Anthropic Messages. Example: run Nemotron 3 Ultra on Terminal-Bench 2 under Codex with a short TOML + the CLI. Harbor datasets work out of the box; NeMo Gym + OpenEnv are alpha. 🔗 Full breakdown + interactive explainer: marktechpost.com/2026/07/13/pri… Technical details: primeintellect.ai/blog/verifiers… @PrimeIntellect @mikasenghaas

@Azaliamirh x.com/Marktechpost/s…

Check out TRACE, a new self-improvement approach where the agent identifies the missing capabilities behind its own failures and trains itself to address them. TRACE-trained Qwen3.6-27B reaches 73.2% on SWE-bench Verified, outperforming much larger models like Codex 5.2 and GLM 5, while beating GRPO and GEPA with <1/4 the training rollouts. By contrasting successful and failed trajectories, TRACE identifies its own weaknesses (such as bug localization or retrieval of the correct doc) and creates new synthetic environments to fix them. The result is a transferable and sample-efficient synthetic env / data generation + fine-tuning pipeline for agentic tasks. Great work led by @TarunSures41845 and @hangoo_kang!

@hangoo_kang @TarunSures41845 x.com/Marktechpost/s…

“TRACE: Capability-Targeted Agentic Training” got Spotlight @ ICML AIWILD 🎉 Beats direct RL, GEPA, & synthetic-agent data on SWE-Bench Verified and τ²-Bench. TRACE-Qwen3.6-27B tops GPT-5.2-Codex, GLM 5, & Claude 4.5 Sonnet on SWE-Bench. Co-led with @TarunSures41845. Thanks to @JonSaadFalcon and our advisor @Azaliamirh. Details below 👇

@PrimeIntellect x.com/Marktechpost/s…

Today, we are releasing verifiers v1 — an overhaul of our environment stack for the modern era of agentic RL and evals. We decompose environments into a taskset, a harness, and a runtime. Run complex agentic tasks like coding and computer use at scale, in any harness.

🧠 University of Michigan researchers introduce NeuroVFM, a generalist neuroimaging foundation model trained on 5.24M clinical MRI + CT volumes (566,915 studies, 20+ years of routine care). The base model, Vol-JEPA, extends I-JEPA and V-JEPA to volumetric medical imaging. How it works: → Self-supervised, vision-only — no labels, no radiology reports, no voxel decoder → Predicts masked 3D-patch representations in latent space → Student encoder + predictor + EMA teacher, trained with a smooth L1 loss → The research team call this paradigm "health system learning" Results across 156 diagnostic tasks (74 MRI / 82 CT): → 92.68 AUROC (CT), 92.49 AUROC (MRI) → Outperforms HLIP, PRIMA, NeuroMAE, DINOv3, BiomedCLIP → Full pretraining in <1,000 GPU hours Paired with Qwen3-14B (NeuroVFM-LLaVA) for report generation + triage: → Beats GPT-5 and Claude Sonnet 4.5 on 3-tier acuity accuracy → 92.6% vs 71.2% (GPT-5) balanced triage accuracy in a 1-week prospective study → Critical-finding miss rate 13.5% vs 50.3% (GPT-5) → >24× cheaper report inference Full analysis: marktechpost.com/2026/07/12/mee… Paper: nature.com/articles/s4159… Codes: github.com/MLNeurosurg/ne…

Thinking Machines Lab (Mira Murati's lab) published a new report "The Future Worth Building Is Human." The core argument: most AI today is trained in a few places, then frozen. It never learns from the people who actually use it. The lab wants the opposite — AI that is distributed, customizable, and shaped by its users. Four technical directions they're pursuing: → Train strong models with multimodal interaction + customizability → Build tools (Tinker) that let you fine-tune and own the model weights → Develop interaction models that widen the human↔machine channel → Publish research on how models are actually made Two ideas worth understanding: (1) Distributed knowledge needs distributed AI. Much real know-how is tacit and local (think a chef's recipe) — you can't dump it into a database. So AI should help orgs cultivate that knowledge, not extract a snapshot and replace it. Chess and math are the exceptions: static goals, no hidden knowledge, so autonomy works. (2) Values belong in weights, not prompts. A prompt only changes surface behavior; deeper habits stay fixed. Their fix is fine-tuning your own values into portable LoRA weights you keep — so alignment becomes an ecosystem of diverse, owned models instead of one central spec. They also reframe evaluation: benchmarks like METR's measure how long a model works alone, not what humans + machines achieve together. Full breakdown + interactive explainer: marktechpost.com/2026/07/11/mir… Report: thinkingmachines.ai/blog/the-futur… @miramurati @thinkymachines

New tutorial: NVIDIA tile-based GPU programming, explained by building real kernels in Colab. Instead of writing code one thread at a time, tile programming operates on whole data tiles — load a tile, compute on it, store it back. The compiler maps it onto threads and tensor cores for you. What the tutorial does: → Probes your CUDA environment and tries the real NVIDIA cuTile backend → Falls back to Triton when standard Colab GPUs lack the cuTile stack (needs CUDA 13.1+, Ampere/Ada/Blackwell) → Builds 5 kernels: vector add, fused GELU, row-wise softmax, tiled matmul, flash attention → Checks every kernel against PyTorch for correctness and benchmarks it Flash attention is the capstone: online softmax, no full attention matrix materialized. Same tile idea shown in both cuTile and Triton spellings. Full Tutorial: marktechpost.com/2026/07/11/a-c… ⭐ Check and Star the Tutorial repo to follow along: github.com/NVIDIA/TileGym ⚡️ Check the Official NVIDIA repo: github.com/NVIDIA/TileGym @NVIDIAAI @NVIDIAAIDev @nvidiadeveloper

[Most robots react. This one thinks a step ahead.] Ant Group's Robbyant just published LingBot-VA 2.0 — a video-action foundation model built from scratch for robot control, not fine-tuned from a video generator. The usual approach takes a video generator made for content creation and bolts a robot policy onto it. LingBot-VA 2.0 argues that's the wrong starting point, and pretrains the whole causal stack natively instead. What stands out: → Foresight Reasoning — the robot predicts the next action chunk while executing the current one, then overwrites the imagined frame with the real observation. Prediction and execution stop waiting on each other. → 927 ms → 142 ms per chunk, across four cumulative optimizations. That lifts asynchronous control from 35 Hz to 225 Hz — a 6.5× speedup. → One shared latent space. A semantic visual-action tokenizer puts world states and actions in the same coordinates, so unlabeled web video carries action-relevant signal. → Sparse MoE video stream — 128 experts, top-8 routing. Roughly 2.5B of ~15.3B parameters fire per token. → Few-shot by design — adapts from 10–15 demonstrations, and a human demo video can replace the text instruction entirely. Full breakdown: marktechpost.com/2026/07/11/ant… Paper: github.com/Robbyant/lingb… Project Page: technology.robbyant.com/lingbot-va-v2 @robbyant_brain @AntGroup

@miramurati x.com/Marktechpost/s…