AI·The blog post argues against constantly telling users to ask an LLM for help.
Search
What are you looking for?
60 results for “rl”
AI·A developer reverse-engineered the Grok build CLI and detailed exactly what information and commands it transmits to xAI servers.
AI·Virginia Tech researchers found that resistance training (modeled as mouse weightlifting) outperforms running for reducing fat and improving insulin sensitivity and glucose control in obesity and Type 2 diabetes models.
AI·The iroh.computer blog introduces Mesh LLM, a distributed AI computing system running on the iroh network protocol.
AI·The Economist explains bizarre anti-drone camouflage tactics used in Ukraine, including vivid black-and-white striped patterns on Russian lorries designed to confuse machine-vision systems on Ukrainian drones.
AI·Article analyzes neoclouds CoreWeave and Nebius' massive GPU builds funded by circular financing: Nvidia's $4B equity stakes plus a $6.3B GPU purchase backstop, offset by $122B+ hyperscaler leases (Microsoft/Meta) that shift capex to opex and enable rapid Blackwell/Rubin deployment.
AI·A female US rower has completed a historic solo journey from California to Hawaii.
AI·A 2016 article argues that while doctors die differently from the general population due to lifestyle factors, this unique mortality pattern should still be acknowledged.
AI·SQLite strict tables enforce column types to prevent datatype errors like inserting text into integer columns, allowing only standard types or ANY for flexibility, but cannot be added to existing tables and require SQLite 3.37+.
AI·ClickHouse scaled PgBouncer throughput 4x by running a fleet of processes sized to cores, using so_reuseport to balance connections on one port and inter-process peering for query cancellation.
AI·The article dissects UPI's complex payment transaction architecture, detailing seven parties, NPCI's central role, and the shift from person-to-person to merchant payments.
AI·The article traces the early history of the singular value decomposition, documenting its development from initial ideas to practical applications in the 1960s and 1970s.
AI·Site offers 80+ structured courses for building real systems including Redis, Git, databases, compilers, and kernels from scratch in Python, Go, Rust, and others.
AI·Blackwell RTX Spark is demonstrated on a real machine at Bilibili World, with CPU and GPU directly soldered together for running 120B-parameter models on laptops.
AI·Alibaba's Lingbo released six embodied AI models from July 7-10, 2026, introducing 'Embodied Native' — a new paradigm that trains foundation models from scratch on physical-world data, interactions, and causal physics rather than migrating internet LLMs, with LingBot-VA 2.0 as the flagship VLA model achieving real-time 150Hz inference and strong generalization.
AI·SpaceX filed FCC for 100,000 Gen 3 Starlink satellites with 10x bandwidth upgrades, explicitly linking the constellation to AI workloads, data centers, autonomous vehicles, and humanoid robots for global compute connectivity.
AI·The embodied data industry has attracted nearly 100 players with 4.47 billion yuan in funding over the past year, yet the question remains whether revenue can be generated through data sales.
AI·Momenta listed on HKEX and soared over 6% on day one, securing global capital, while Mushroom Car Network's ToG Robobus model lags due to heavy hardware investment and cashflow pressure despite earlier physical AI claims.
AI·Nearly 100 players have entered the embodied data sector, securing 4.47 billion yuan in funding within one year and raising questions about sustainable revenue from data sales.
AI·DeepSearch-World is a self-distillation framework for web search agents that uses self-generated experience for training in verifiable environments.
AI·The paper identifies a failure mode in critic-free RL methods for LLMs where implausible tokens receive uniform credit, and proposes tail-aware credit calibration to improve reinforcement learning.
AI·The paper introduces PLURAL, a large-scale value-focused preference dataset grounded in the Integrated Values Survey across 92 countries to improve LLM representation of diverse non-Western value systems.
AI·The paper probes internal representations of Eternis-Forecaster 8B for forecasting, training a representation-pooling method to assess calibration and faithfulness of internal CoT reasoning.
AI·TypeProbe recovers type representations from hidden states of pretrained code models using parallel Java and Python datasets.
AI·Large-Language-Models-as-a-Judge enable theory-agnostic adaptive metric-alignment for prototypical networks in personality recognition.
AI·The paper enriches Roy Harris's Integrationism theory with Barenholtz's Autogenerative Theory to address gaps in computational language approaches.
AI·Global first embodied native pre-trained model released as open-source version LingBot-VA 2.0, designed from physical world for robotic brains.
AI·The book "RISC-V Microprocessor System-on-Chip Design" provides a comprehensive guide to designing RISC-V-based SoCs.
AI·UniClawBench is the first capability-driven benchmark for proactive agents, using live Docker containers and closed-loop multi-agent evaluation to assess five foundational skills across 400 real-world tasks.
AI·The paper introduces IdeaGene-Bench (IG-Bench) to benchmark AI systems on scientific lineage reasoning and lineage-grounded idea generation.
AI·Dual Latent Memory in Vision-Language-Action models for robotic manipulation interleaves historical experience fluidly in the native latent embedding space, overcoming Markovian limitations in long-horizon tasks.
AI·LingBot-World 2.0 is an advanced world modeling system with four upgrades: unbounded interaction horizon, real-time 60 fps video support, diverse interactive actions and events, and multi-agent control for multi-player virtual environments.
AI·Single-Rollout Asynchronous Optimization is an agentic RL method that replaces group-wise sampling with single-rollout updates and adds token-level clipping to improve stability and performance on coding and reasoning benchmarks.
AI·Talos-XII is a Rust-based CLI simulator for Arknights: Endfield gacha probabilities that trains small neural nets for uncertainty modeling and decision policies without using PyTorch or tch-rs.
AI·UP is a universal RL objective that breaks the exploration-stability dilemma by allowing unclipped positive gradients while clipping negative ones, enhancing exploration without sacrificing training stability across various RL frameworks.
AI·Details not yet available
AI·LingBot-Video is a 13B-parameter sparse-MoE video diffusion transformer (1.4B active) post-trained as an action-conditioned world model with physical-plausibility RL, releasing weights, code, and Diffusers/SGLang integration for robot rollouts.
AI·RoboDojo is introduced as a new unified benchmark that combines simulation and real-world testing for comprehensive evaluation of generalist robot manipulation policies.
AI·WildCity introduces a real-world multimodal dataset of 18 autonomous-vehicle trajectories spanning 1500 km total, covering 18 trajectories averaging 83.7 km each, to enable city-scale spatial intelligence research via rendering, simulation, and urban digital twin baselines.
AI·SWE-Review closes the loop on AI-generated pull requests by using an agentic reviewer that explores repositories, accepts or rejects PRs, and supplies structured revision feedback, outperforming one-shot generation on SWE-Review-Bench and enabling test-time scaling.
AI·PhyMRI-SR reframes MRI super-resolution as a physics-aware problem that dynamically optimizes resolution-SNR trade-offs using 2D Gaussian Splatting, prior-aware representations, physics-constrained signal modeling, and meta-learning, achieving state-of-the-art on dynamic-resolution benchmarks.
AI·GeneBench-Pro is a new research-level benchmark with 129 problems across 10 domains in computational biology, testing AI agents on judgment-heavy tasks involving ambiguity, iterative experimentation, and decision-making with synthetic yet realistic datasets.
AI·OpenAI launched Codex Security plugin updates and the full GPT-5.5-Cyber model to accelerate vulnerability discovery, validation, and automated patching for defenders, alongside the Daybreak Cyber Partner Program and Patch the Planet initiative for open-source projects.
Long-horizon failure in world models is conventionally attributed to compounding error, a generic framing that does not distinguish what kind of error compounds. We propose a kinematic-vs-dynamic reframing: world models tend to imagine kinematically rather than dynamically. We operationalize this as the imagined Kinematic-Consistency Error, a per-step diagnostic that measures how far a rollout departs from a closed-form kinematic null, paired with a perturbation protocol that tests whether iKCE…
Samsung Electronics deploys ChatGPT Enterprise and Codex to employees worldwide, marking one of OpenAI’s largest enterprise AI rollouts.
Introducing LifeSciBench, an expert-authored, expert-reviewed benchmark for evaluating how AI systems handle real-world life science research tasks and decisions.
Learn how BBVA scaled ChatGPT Enterprise to 100,000 employees and partnered with OpenAI to accelerate AI-powered banking transformation worldwide.
We’re expanding access to Google AI Ultra subscribers globally and introducing a new capability powered by Street View.
Google DeepMind partners with global consultancies to bring the power of frontier AI to organizations around the world.
Gemini Robotics ER 1.6: Enhancing spatial reasoning and multi-view understanding for autonomous robotics.
Reinforcement learning (RL) for non-verifiable instruction following increasingly relies on LLM judges with prompt-specific rubrics as reward signals. While recent methods adapt these rubrics to the evolving policy during training, the training prompts themselves remain static, drawn from fixed corpora. This static approach often results in a critical misalignment between prompt difficulty and policy capability, leaving the judge unable to recover a discriminative reward signal when prompts fai…
Algorithms & Theory
Health & Bioscience
Visual policies learned from human videos, teleoperation, and robot demonstrations offer scalable motion priors, but often fail in contact-rich manipulation, where success significantly depends on local force and contact geometry. Tactile sensing provides these complementary signals, yet tactile data remain costly to collect and hard to generalize across sensors, robots, and tasks. We introduce OmniTacTune, a policy-agnostic real-world RL pipeline that adapts tactile feedback to pretrained visu…
Pixel-wise Earth-observation (EO) foundation models are now achieving state-of-the-art performance via generated spatial embeddings. However, how these models scale and how best to spend a pretraining budget remain poorly understood. We present the largest controlled scaling study for EO to date: 395 training runs on 1,024 GH200 superchips within a fixed pixel-wise Barlow Twins family, each evaluated on 15 downstream tasks. We find that pretraining loss barely predicts downstream performance (|…
The inherent complexity of video understanding makes it difficult to determine whether Video-LLM benchmark performance stems from visual perception, linguistic reasoning, or knowledge priors. While many benchmarks have emerged to assess high-level reasoning, shared criteria for evaluating video understanding remain largely overlooked. Instead of introducing yet another benchmark, we take a step back to re-examine the criteria for evaluating video understanding. In this work, we introduce Video-…
Large Audio-Language Models (LALMs) are increasingly integrated into daily applications, yet their generative biases remain underexplored. Existing speech fairness benchmarks rely on synthetic speech and Multiple-Choice Questions (MCQs), both offering a fragmented view of fairness. We propose VIBE, a framework that evaluates generative bias through open-ended tasks such as personalized recommendations, using human-recorded speech. Unlike MCQs, our method allows stereotypical associations to man…
Reinforcement learning (RL) has become a central component of post-training large language models (LLMs), yet little is understood about how RL adaptation is distributed across transformer layers. Existing approaches typically update all model parameters uniformly, implicitly assuming that every layer contributes similarly to the gains obtained during RL post-training. In this work, we challenge this assumption through a systematic layer-wise study of RL training. Surprisingly, we find that tra…