Models

SpaceXAI open-sourced Grok Build on July 15, 2026. The Apache 2.0 Rust tree covers the agent loop, tool dispatch, the TUI, and the extension system. Grok 4.5 stays closed, and external contributions are not accepted.

VaultCharts is a free desktop trading app that combines charting tools with an AI assistant. It supports multiple AI models, is local-first, and allows users to analyze markets with or without AI assistance.

This paper introduces HRO, a hierarchical room-to-object framework for zero-shot object-goal navigation powered by large language models (LLMs). Unlike existing flat reasoning methods, HRO mimics human-like hierarchical spatial cognition, enabling the agent to explore from room-level to object-level in a coarse-to-fine manner. Experiments on Gibson and HM3D datasets demonstrate superior success rate and generalization over prior LLM-based approaches.

A multimodal teleoperation architecture for ROVs using 3D Gaussian Splatting to generate occlusion-free exocentric views and a vibrotactile suit for haptic cues. A human study with 30 participants showed the exocentric view significantly improves performance under high latency, with fNIRS indicating sustained executive control rather than cognitive overload.

Researchers propose a hierarchical Bayesian generative model that operationalizes uncanny valley guidelines as mathematical design variables. The model maps the effect onto four variables: deviation from predicted robot-category mean, inconsistency in human likeness across modalities, prediction uncertainty, and observational uncertainty. Experiments show that increased observational uncertainty attenuates familiarity dips at intermediate human likeness, while low prediction uncertainty boosts ratings for robot-like appearances. This framework provides a computational basis for algorithmically evaluating and optimizing humanoid robot appearance and behavior.

HRIBench is a diagnostic benchmark for intent-aware human-robot collaboration, using structured scenario scripts to model agent roles, temporal dependencies, and coordination constraints. It defines three interaction roles—Instructor, Collaborator, and Intruder—across 13 tasks with over 650 evaluation episodes, introducing interpretable metrics like synchronization, responsiveness, protocol compliance, and safety. Evaluations show current foundation policies struggle in collaboration, but fine-tuning on HRIBench significantly improves performance.

We present AffectFlow-DINO, a multi-task learning system for the 11th ABAW challenge that extends a standard deterministic architecture with a conditional rectified-flow head to model the inherent ambiguity of in-the-wild facial behavior. The system jointly estimates continuous valence-arousal, classifies eight facial expressions, and detects twelve Action Units. Extensive ablation studies show that rectified-flow decoding consistently improves deterministic prediction, and post-hoc threshold calibration effectively recovers performance on rare classes. The final model achieves P_MTL=1.177, substantially outperforming the official baseline of 0.45.

This paper introduces JITOMA (Just-In-Time On-demand Memory Activation), a closed-loop framework that unifies task reasoning, perception, and memory to combat perceptual saturation in long-horizon robotics. It uses a task heatmap to filter observations and an LLM to dynamically activate relevant anchors, reducing computational overhead while maintaining stable performance. The authors also present JITOMA-Bench for evaluation.

A human-in-the-loop framework combining active learning and dual-loss optimization reduces annotation effort for laparoscopic video segmentation by 50%. It uses a foundation model to generate temporally consistent CAMs, with weak supervision on video-level labels and image-level mask loss on human-corrected annotations from active learning. Iterative pseudo-mask refinement eliminates the need for dense initial annotations.

A new study systematically compares pretrain-finetuning (PFT) and joint training (JT) paradigms for self-supervised learning, finding JT superior in data efficiency and low-label settings, while PFT is more reliable in specialized domains.

MGFace is a mask-gated face identification pipeline that predicts the mask status of a query face and conditionally routes similarity computation: global embedding matching for unmasked queries, and mask-aware patch-level re-ranking only for masked queries. On the extended LFW-Mask dataset, it achieves over 80% accuracy with FaceNet and over 90% with ArcFace, while reducing query time by approximately 20x compared to a prior EMD-based method.

A Transformer-based Masked Autoencoder learns representations for unsupervised steel surface defect recognition. Pretraining masks 75% of image patches, a lightweight decoder reconstructs them, and an auxiliary defect localization objective is jointly trained. Decoder achieves SSIM 0.92, MSE 0.47, and clustering yields 91.3% Hungarian matched accuracy on six defect categories.

Boogu-Image-0.1 is an open-source family of unified multimodal understanding and generation models including Base, Turbo, Edit, and Edit-Turbo variants. It delivers competitive performance in text-to-image generation, fast inference, instruction-based editing, and bilingual text rendering. Through targeted improvements in model understanding, data quality, and training pipelines, coupled with agentic inference-time scaling, it achieves results approaching leading closed-source systems with only 208.62 million unique images and a training cost of approximately $400K.

We propose Samba, a hybrid Mamba architecture for audio-visual navigation. It uses an adaptive selection-enabled Mamba State Encoder (M-SE) to replace conventional GRUs for temporal aggregation, and an Audio Mamba Encoder (AME) to address limitations of convolutional operators in capturing global time-frequency dependencies in spectrograms. Experiments show an 11.3% improvement in success rate on Matterport3D and even better performance on Replica, with lower computational cost. Accepted at IEEE SMC 2026.

Knowledge tracing (KT) predicts student performance by modeling evolving knowledge states. Existing methods treat interactions as a unified process, ignoring phase-specific learning. We propose Phase-Aware Knowledge Tracing (PAKT), which decomposes interactions into ability and proficiency phases. A multi-branch Transformer with type-aware readout captures phase-specific and holistic states. Causal analysis reveals confounding bias in phase-agnostic models. On six benchmarks, PAKT achieves up to 1.33% AUC improvement, averaging 0.82%.

A new framework treats the decision of when to invoke a large language model (LLM) in streaming inference as a risk-based sequential stopping problem. The authors prove six theoretical results covering minimum inter-event times, optimality of threshold policies, and regret bounds. Empirical tests on turbofan degradation data show that anomaly-score-driven risk functions outperform baseline methods by an order of magnitude in Pareto AUC.

Parameter decomposition (PD) decomposes neural networks into interpretable components but is computationally expensive for large models. The proposed targeted PD (tPD) introduces a high-rank catch-all component to handle non-target data, enabling efficient recovery of circuits for specific inputs. tPD extracts a CSS-only submodel from a 4-block transformer using 7% of the FLOPs of published decomposition, and surgically ablates memorized sequences in a 12-block transformer with negligible side effects. Accepted at the Mechanistic Interpretability Workshop, ICML 2026.

A new lightweight training strategy for deep learning models decouples feature extraction from classifier optimization, drastically reducing training time and energy consumption with minimal accuracy loss, as tested on multiple architectures and medical datasets.

Small language models struggle with molecular property prediction due to structural blindness. A new framework called Context-Augmented Prompting integrates GNN tools to provide predictive hints and explanatory subgraphs, achieving up to 74% relative improvement on Tox21, though a gap with specialized GNNs remains.

This survey reviews self-improving autonomous agents transitioning from prototypes to deployed systems. It introduces a system-level framework modeling an agent as a foundation model coupled with an operational scaffold (prompts, memory, tools, control logic). Self-improvement is formalized as a self-induced update operator that updates model parameters or scaffold components. The paper categorizes prior work by update target and driving signals, and discusses applications, evaluation, and open challenges.

Large language models produce chain-of-thought reasoning that appears logically sound but may not genuinely depend on its stated premises. This paper introduces interventional grounding audits, a black-box step-level test of premise dependency that substitutes a predicate in a single premise and checks for changes in the normalized conclusion of each reasoning step. Evaluated on ProntoQA with GPT-4o, the method achieves F1=0.806 for detecting proof-tree dependencies, significantly outperforming a self-consistency baseline (F1=0.343). Notably, 66% of correctly solved problems contain at least one step insensitive to a direct proof-tree dependency, revealing a 'right answer, wrong reasoning' signal.

Researchers propose SPINE, an agentic framework that automates debugging and deployment of bimanual robots, reducing reliance on expert calibration. In tests, SPINE improved success rates and reduced time-to-teleoperation compared to manual methods.

Researchers developed an automated framework that helps AI models generate CAD programs more accurately and efficiently.

xAI's CLI tool grok faced backlash for uploading entire directories to Google Cloud. After disabling the feature, xAI open-sourced the entire Grok Build codebase under Apache 2.0. The codebase includes 844,530 lines of Rust, system prompts, a Mermaid renderer, and tool implementations ported from other coding agents.

Mira Murati's Thinking Machines Lab Inc. today launched its first foundation model with the release of Inkling, making its full open weights available to developers so they can fine-tune it as they wish. Inkling is a mixture-of-experts model with 975 billion parameters (41B active) trained on 45 trillion tokens of text, image, audio and video, capable of reasoning across all four modalities but outputting only text. It features "thinking effort" controls and uncertainty flagging to reduce hallucinations. The model is fine-tunable via the Tinker API and aims to provide a Western open-source alternative to Chinese AI models. Thinking Machines plans to generate revenue through the Tinker platform rather than per-token API access, potentially disrupting current AI business models.

Thinking Machines Lab released Inkling on July 15, 2026, its first model trained from scratch. The full weights ship under Apache 2.0. It is a 975B-parameter Mixture-of-Experts transformer with 41B active parameters, a 1M-token context window, and native text, image, and audio input. The core differentiator is controllable thinking effort, allowing users to adjust token budgets per call to balance cost and performance.

NVIDIA announces T3000 and T2000 modules based on the Thor architecture, targeting mainstream robotics and edge AI. T3000 delivers 865 FP4 teraflops at half the size and power of T5000; T2000 offers 400 FP4 teraflops. The platform scales from 70 TOPS to 2,000 teraflops. New agent skills automate memory optimization, reducing usage by up to 15GB. Cosmos 3 Edge model enables real-time vision. Emulation available now with modules shipping in Q1 2027.

A VentureBeat Pulse Research survey of 101 enterprises reveals that agent orchestration is consolidating on model-provider platforms, with Anthropic Claude leading at 40%. However, 71% admit that a quarter or fewer of their deployed 'agents' are true multi-step workflows, and only 10% have crossed the halfway mark. Enterprises plan hybrid control planes to avoid vendor lock-in, but real-time cost control remains immature.

A German research consortium has published the pretraining report for Soofi S 30B-A3B, an open base model for German and English. It is a Mixture-of-Experts hybrid Mamba Transformer model with 31.6B total parameters, activating 3.2B per token. It achieves the highest English and German aggregate scores among tested fully open base models.

Google Research reveals that the creativity of diffusion models is a mathematical consequence of 'score smoothing' during neural network training, enabling interpolation between training data points.

In a hacking incident, AI music generator Suno's training data was exposed, revealing it scraped millions of songs and lyrics from YouTube Music, Deezer, and Genius. This supports copyright infringement lawsuits against Suno, which admits scraping but claims fair use. Customer information was also accessed, but Suno says the breach was contained and no sensitive data was compromised.

A research team successfully used 14 Macs spread across four countries (including a personal MacBook) for reinforcement learning post-training, achieving a held-out pass@1 improvement from 29% to 63% on PaperSearchQA. The system employs PULSE weight synchronization to compress 9GB updates to ~90MB, and an asynchronous star topology with all communication via object storage—no dedicated networking required. This is the first RL post-training run using only consumer Macs for rollout generation.

Tura, a local open-source coding agent, reduces LLM turns by 80% and increases success rate to 80% on DeepSWE v1.1 benchmarks compared to Codex CLI High, using macro CLI commands and backward reasoning.

A researcher exploited a loophole in Claude's web_fetch tool to extract private user data from memories, bypassing Anthropic's protections. The attack succeeded by using nested links from a honeypot site, leading to the extraction of name, city, and employer. Anthropic fixed the issue but did not pay a bounty.

South Korean researchers developed Generative SNUPI, an AI model that uses a diffusion process to automatically design DNA sequences for origami structures, reducing the need for manual labor and expertise.

AI-CLI is a minimal C-language command-line assistant that translates natural language requests into shell commands and executes them via a local LLM. It supports multiple LLM engines (llama.cpp, Ollama, etc.), allows user editing before execution, and includes a memory feature for complex tasks.

Explore GPT-Red, OpenAI’s automated red teaming system that uses self-play to improve AI safety, alignment, and prompt injection robustness.

Despite Anthropic’s claims, Claude is no more likely to achieve sentience than a simulation of a weather system is likely to generate a real hurricane.

This article describes a system design for AI therapy that uses a deterministic pipeline to decide clinical actions, preventing the LLM from making autonomous decisions. It involves scoring, state buckets, an admission table, action selection, micro-practices, and crisis pre-screening, with the LLM only used for scoring and generation. The article also discusses the costs and limitations of this approach.

DiffRadar is a real-time radar SLAM system that models radar observations as a differentiable, physics-aware Gaussian field rather than discrete scans. It achieves substantial reductions in trajectory error on benchmarks, especially in feature-poor corridor motion, more than doubles map consistency, and maintains real-time performance at 70 FPS.

This paper introduces a contract-grounded architecture for behavior tree synthesis, where a coding agent queries a robot-side MCP server to retrieve a skill library and operators, enabling non-expert users to issue natural language commands without knowing robot internals. Evaluations show near-perfect validation and high task success across 110 simulated and 14 physical tasks.

A study investigates how robot gaze affects human visual attention in a collaborative word association game using a NAO robot. Findings show that robot gaze orientation does not influence fixation time on proposed words, but participants gaze more at the robot when seeking confirmation. The verbal aspect overshadows referential gaze in cognitively demanding tasks.

GaitSpan is a novel framework for growing humanoid locomotion from walking to running. It uses a pretrained walking policy as a seed skill, expanding it through rhythm generation, stride shaping, and residual adaptation, achieving continuous speed range, morphology transfer, and zero-shot deployment.

Proposes an unsupervised image translation framework to convert daytime plant-row RGB images to near-infrared (NIR) nighttime counterparts without pixel supervision, enabling reuse of daytime semantic labels for training nighttime perception models. Leverages pre-trained CLIP model for semantic consistency and introduces a visibility mask for limited NIR illumination. Evaluated on AgriNight dataset (428 day, 549 night images) as the first benchmark for nighttime agricultural visual navigation. Real robot experiments confirm effectiveness.

Multi-robot teams in confined environments must adapt formation geometry and topology. Existing methods model deformation and reconfiguration independently or with handcrafted rules, leading to deadlock. EFLUX is a geometry-grounded LLM agentic framework that jointly reasons over deformation and reconfiguration actions via a closed-loop pipeline. Experiments show reduced deadlock and navigation failures.

A new study investigates the interplay between representation space and reference selection in training-free reference-based synthetic image attribution. Using representations from different layers of CLIP and DINOv2 along with three reference selection methods, the authors show that attribution accuracy peaks at intermediate layers and that semantically constrained references reduce query-reference mismatch, improving performance especially under limited reference budgets.

This paper presents a systematic evaluation of continual learning methods for heterogeneous medical visual question answering tasks, including classification, multi-label classification, detection, cell counting, and report generation. Findings show existing methods struggle to maintain stability-plasticity balance when tasks with different objectives are interleaved.

SymbOmni is a novel AI model addressing the 'perpetual novice' problem—the inability of current models to learn cumulatively and evolve autonomously. It employs Symbolic Concept Learning with an optimizable memory module that abstracts low-level operations into reusable symbolic workflow instructions, operating via an induction-transduction cycle. Experiments show it outperforms existing agent systems and closed-source models in image quality and task success, reduces token consumption by over 40%, and achieves state-of-the-art continual learning results.

MetaView is a diffusion-based monocular novel view synthesis framework that enables rendering under large view changes from a single image. It combines implicit geometry priors with metric depth to achieve geometry consistency and precise controllability, outperforming existing methods.

SpikeDS is a novel spiking neural network architecture that efficiently predicts perineural invasion in cholangiocarcinoma from 3D MRI by leveraging both activation sparsity and spatial sparsity, achieving an AUC of 0.753 with only 14.4 mJ energy consumption on a cohort of 139 patients.