Tencent's Hy team released Hy3, a 295B-parameter Mixture-of-Experts model that activates only 21B parameters per token. It is available under Apache 2.0 license with a 256K context window, targeting reasoning, agentic, and long-context tasks. Hy3 scores 78.0 on SWE-Bench Verified and lower hallucination rates. Free tier available on OpenRouter until July 21, 2026.

A deep dive into the core mechanics of LLMs: next-token prediction, training via backpropagation, logits and softmax, the role of temperature, and the probabilistic nature of outputs. Explains why models hallucinate, how chain-of-thought prompting improves accuracy, and what it means for a model to 'know' something.

sqlite-utils 4.0rc4 is the last release candidate before the 4.0 stable release, mainly implementing feedback from a detailed review by Claude Fable 5.

The article explores the challenges of building real-time voice AI for low-resource languages like Azerbaijani, comparing end-to-end speech-to-speech models (OpenAI Realtime, Gemini Live) and cascaded pipelines (LiveKit, Pipecat, Vapi). It details failure modes, component availability, and provides a checklist for evaluation. Key findings: Gemini Live spoke well but was too slow; OpenAI Realtime had accent issues. Available components include Azure TTS, ElevenLabs, and Scribe v2. The cascaded stack offers flexibility but requires engineering latency.

An interactive report from May 2026 analyzes AI search visibility for Australian general insurance brands, enabling head-to-head comparisons between two brands.

Covering Thariq's talk on Fable: unhobbling models, finding unknowns, emotional shifts, and being unreasonable. Plus AI news: Tencent Hy3 open-source, agent benchmarks, Anthropic J-Space, inference efficiency, world models, and speech.

OpenAI has introduced two new Realtime models for low-latency voice and multimodal applications. The gpt-realtime-2.1-mini is a reasoning model priced like its predecessor, while p95 latency is cut by at least 25% via improved caching. This article covers model details, pricing comparisons, and WebRTC integration.

This paper presents a robust nonlinear lateral control framework that accounts for varying longitudinal speed and acceleration, addressing limitations of existing constant-speed assumptions and parameter uncertainties. It derives a tracking error model, uses feedback linearization, and proposes two robust designs: Lyapunov redesign and incremental nonlinear dynamic inversion. Simulations and real-vehicle tests demonstrate enhanced tracking accuracy and robustness.

This paper addresses sampling-based motion planning for continuous-time stochastic systems under process and measurement uncertainty, with probabilistic guarantees on safety and performance. It models robot dynamics as a continuous-time linear stochastic differential equation and uses discrete-time sensor measurements. A hybrid belief propagation model is derived, where belief evolves via continuous-time ODEs between measurements and undergoes discrete Kalman filter updates at measurement times. A belief-barrier-function-based safety checker enables segment-level probabilistic verification, detecting inter-sample chance-constraint violations missed by conventional node-based checks. The method is integrated with RRT and SST planners and evaluated on benchmark environments, showing high su…

DREAMSTEER is a deployment-time steering framework for pretrained vision-language-action (VLA) policies that requires no finetuning. It leverages a latent world model and a value model to sample, imagine, and rank candidate actions, significantly improving task success rates and instruction-following accuracy.

This paper proposes AmpAttention, a novel attention mechanism inspired by differential amplifiers in analog circuits, to suppress attention drift in multi-view robotic manipulation. The RVAF model, built on AmpAttention, achieves state-of-the-art average success rates on 18 RLBench tasks (249 variations) while reducing training time by 33.3%. Its extension RVAF++, incorporating the SAM2 image encoder, reaches a 91% success rate on the 'insert peg' task. The work has been accepted at IROS2026.

CLEAR introduces a closed-loop training system using reinforcement learning for end-to-end autonomous driving. By learning a residual waypoint policy from pretrained VLA models and employing a heterogeneous simulation pipeline, CLEAR achieves state-of-the-art performance on CARLA longest6 v2 and Bench2Drive.

TACO is a tactile-aware world-model-driven framework for scalable VLA post-training in contact-rich manipulation. It converts real failures into imagined visuo-tactile corrections via a Recognize-Imagine-Label loop, achieving a 44% absolute success rate improvement over the base policy.

This paper investigates metallic ultrasound waveguides as distributed tactile sensors using a single proximal transducer. Experiments show a linear relationship between force and reflection/transmission coefficient ratio, and a load-independent parameter for material classification. The approach enables contact localization, force estimation, and material discrimination, reducing system complexity.

Researchers present MorphQuad, a morphable quadrotor with independently articulated rotor systems via two-axis gimbals, enabling omnidirectional thrust vectoring, global stability, and superhuman maneuverability, manipulation, and resiliency. The platform demonstrates multi-revolution rotation, valve turning, perching, and wind rejection.

EVA-Client is an open-source framework for deploying, collecting data, and evaluating trained manipulation policies on real robots. Sitting between a policy server and the physical hardware, it unifies the real-robot stages of the policy iteration loop within a single codebase. The framework features a component-decoupled architecture, inspectable execution workflows (Debug, Collect, Eval), and evaluation-as-data-collection, where each evaluation run produces training-ready trajectory data. It consolidates major real-time inference strategies behind a single configuration surface.

This systematic study introduces WMBench, a benchmark for evaluating world models as surrogate robot policy evaluators. Analyzing 7 video world models, 4 action schemes, and over 324,000 simulated rollouts paired with real robot executions, the authors identify three key insights: evaluator quality depends on long-horizon action-faithful consistency, pretraining requires balancing general knowledge with robot-specific controllability, and architectural choices critically determine alignment with real-world behavior. Based on these, they present GigaWorld-1, an optimized world model, and release all code, models, datasets, and toolkits.

A visual analytics framework for exploring attention dynamics in diffusion models, enabling structured analysis of token-level cross-attention maps across generation steps. Case studies on a 60-prompt benchmark reveal interpretable patterns, supporting human-AI collaboration.

This paper presents an entropy-coded MS-VQ-VAE framework that leverages discrete latent representations and learned autoregressive priors to achieve ultra-low bitrate video compression. Operating at 0.043-0.064 bpp on UCF101, the method outperforms H.265 in perceptual quality while using 5-7.6× fewer bits.

DH-Active is a lightweight, training-free geometry back-end that uses LiDAR as a metric ruler to improve multi-view depth estimation. It anchors near-field returns via PnP, triangulates visual features, and selectively abstains in ill-conditioned regions. The method runs at 1.11 ms on CPU, ~38x faster than a DINOv2-L GPU branch, and achieves 1.4-6.7% median relative error on benchmarks.

This study proposes a two-stage U-Net framework for efficient prediction of time-averaged pedestrian-level wind speed over realistic urban morphologies. The model, trained on UrbanTALES dataset, uses a baseline U-Net (M1) for patch-wise prediction and a refinement U-Net (M2) based on inpainting to eliminate boundary artifacts. Results show reasonable reproduction of mean velocity and spatial variability, while maximum velocities remain underestimated. The framework offers an efficient surrogate for high-resolution pedestrian-level wind prediction.

This study presents a decision framework that quantifies the trade-off between labeling effort and recognition accuracy when transferring vision systems across marine habitats. The benchmark spans five datasets, three oceans, and three taxonomic groups. It finds that frozen self-supervised foundation features (DINOv2 + linear classifier) generalize well, requiring as few as 10-20 labeled images per species for reliable recognition at new sites, cutting annotation effort by roughly an order of magnitude.

Deep learning models for diabetic foot ulcer segmentation often report high accuracy on in-domain data but fail to generalize across clinical sources. This study benchmarks U-Net, DeepLabV3+, and SegFormer-B2 under a strict protocol, finding that Transformer-based SegFormer-B2 generalizes best across external datasets, while model complexity does not guarantee better generalization.

In outdoor driving scenes, sparse-view neural reconstruction is challenging. This paper proposes using Depth Anything V2 as a monocular depth prior, with selective supervision via photometric masks. On Splatfacto, PSNR improves from 14.903 to 15.932 and RMSE drops from 0.542 to 0.100, showing gains from selecting reliable low-error regions.

A study using interpretable machine learning predicts Parkinson's disease motor severity from QSM and fMRI features, achieving 75% accuracy within 5 points of clinical scores.

Video multimodal large language models struggle with precise local spatiotemporal perception when videos differ only in a short time span or small region. DELTAVID converts cross-video spot-the-difference into a trainable perception signal, and introduces DELTAVID-10K and DELTAVID-Bench for scalable training and evaluation. Experiments show improvements on multiple video understanding benchmarks, moving models toward fine-grained evidence reasoning.

VulcanVoxel is a method that learns free-space affordances via a masked autoencoder over 3D occupancy fields, enabling robots to discover feasible volumes for blade insertion in cluttered fabric bins. Trained on 10,000 real warehouse episodes without human annotation, it achieves a top-5 coverage of 0.89 compared to 0.71 for the best pose-based baseline, with a distilled student model reducing inference time from 1.4 s to 30 ms.

Cross-lingual retrieval-augmented generation (RAG) often suffers from language drift and unreliable evidence use when evidence is in English. This paper proposes TR-RAG, a teacher-regularized RL method that combines reward optimization with on-policy distillation on student-visited prefixes, and introduces a reward decomposition, significantly improving language adherence and evidence-grounded correctness across multiple benchmarks.

LLMs can achieve high theoretical CBT knowledge (up to 96% on exams) but fail to apply it effectively in dialogue. A new metric (Protocol Leverage Force) shows that even with multiple chain-of-thought prompting, the effect on behavior is minimal (under 1.5%), and models remain biased toward validation and reflection.

This paper introduces PraMem, a method that constructs experiential memory by practicing on long historical sequences to address challenges in long-horizon behavior prediction with LLMs, achieving superior performance.