Model releases drive changes across AI products and infrastructure. This hub tracks frontier models, multimodal capabilities, open weights, context windows, benchmark signals, API changes, and deployment paths so readers can judge whether a new model changes cost, quality, or availability.
Open models like NVIDIA Nemotron enable enterprises to build AI that uniquely addresses their business needs, offering full control, customization, and cost efficiency, driving the shift from AI adoption to AI ownership.
Open models provide enterprises with full control to customize, inspect, and improve AI for specific business needs.
Post-training and domain-specific tuning allow open models to achieve frontier-level accuracy at a fraction of the cost of closed models.
Researcher Dave Kuszmar discovered multiple systemic vulnerabilities that let him bypass LLM safety and obtain dangerous instructions. These exploits worked across nearly all major LLMs, revealing an industry-wide security problem. Kuszmar calls for slowing deployment, increasing transparency, and large-scale research into LLM safety before further integrating these systems into society.
Researcher found 'Time Bandit' and 'Inception' exploits to bypass LLM safety. Vulnerabilities affect major LLMs like GPT-4o, Claude, Gemini, and others.
Kuszmar obtained instructions for making weapons and drugs. Companies largely unresponsive to disclosures.
An exploration of the historical links between semiotics and large language models, drawing on thinkers like Saussure, Barthes, and Derrida to analyze the relationship between linguistic signs and AI-generated content.
Semiotics offers a theoretical framework for understanding LLM semantics
Saussure, Barthes, and Derrida's theories illuminate AI language models
Mistral AI introduces a vision model that enables robots to navigate unknown environments using only a single RGB camera and natural language instructions.
Mnemo AI is a local agentic AI assistant built with LangGraph and LangChain, supporting multiple LLM providers including Ollama, Bedrock, OpenAI, Anthropic, and more. It features MCP tool integration, RAG, user profile learning, episodic memory, and an ACE Playbook that learns from both successes and failures. The tool also offers web search, image analysis, file operations, bash execution, and many other capabilities.
Supports multiple LLM providers (local and cloud)
Integrates MCP tool system and RAG for document indexing
Demis Hassabis, CEO of Google DeepMind, proposes a global AI watchdog with the power to halt dangerous frontier models. He argues the US should lead the effort and hopes to establish the organization by year-end.
Hassabis proposes an AI regulator modeled after FINRA, composed of independent experts and open-source representatives.
The body would evaluate frontier models pre-release and coordinate industry-wide slowdowns if risks are too high.
Hayden Bleasel has released Blume, an open-source, MIT-licensed documentation framework. It reads a folder of Markdown or MDX and generates a hidden Astro project, shipping static, AI-ready docs with local search, 30+ MDX components, llms.txt, and a built-in MCP server.
Blume is a zero-config documentation framework that turns a Markdown folder into a full docs site.
It uses a hidden Astro and Vite project, supports hot reload, and can eject to a standalone Astro app.
Mistral AI introduced Robostral Navigate, an 8B embodied navigation model. It moves robots from a plain-language instruction using only a single RGB camera, with no LiDAR or depth sensors. The model reaches 76.6% success on R2R-CE validation unseen through a pointing method, prefix-caching training, and CISPO online reinforcement learning.
Robostral Navigate is Mistral AI's first 8B model for embodied navigation.
Achieves 76.6% success on R2R-CE validation unseen using only a single RGB camera.
This paper proposes a human-efficient post-training pipeline that enables a small number of human operators to supervise multiple robots through specialized division of labor and automatic trajectory segmentation using VLAC-CUT. Validated on four real-world manipulation tasks, the final policies achieve 80%-95% success rates and improve task throughput by 1.7x-4.2x over the base model.
Proposes a human-efficient post-training pipeline with role specialization to reduce task switching and training costs.
Introduces VLAC-CUT, an automatic trajectory segmentation tool for filtering useful rollout data.
This paper proposes a risk-field enhanced closed-loop digital twin framework for safety validation of autonomous driving systems. The framework integrates physical data acquisition, virtual reconstruction, risk-aware scenario generation, and algorithm evaluation, using a driving risk field as a unified intermediate representation to identify high-risk scenarios and provide safety guidance for reinforcement learning policies. Experiments show the method improves targeted validation and interpretability, but its effectiveness is bounded by model fidelity and sim-to-real transfer.
Proposes a risk-field enhanced closed-loop digital twin framework
Driving risk field as unified representation for multiple risks
UAV swarms have potential in SAR and environmental monitoring but face limitations in situational awareness, connectivity, and cybersecurity. This paper proposes LAUS, an LLM-centric agentic AI framework integrating perception, memory, reasoning, and action for adaptive swarm behavior. It reviews enabling technologies, analyzes threats like Priority Manipulation Attacks, and identifies open challenges including hallucination-resistant reasoning, onboard LLM deployment under SWaP constraints, and standardized security benchmarks.
Proposes LAUS, an LLM-centric agentic AI architecture for autonomous UAV swarms.
Reviews enabling technologies: edge computing, 5G/6G, multimodal intelligence, and cybersecurity.
EgoSteer is a full-stack system that enables steerable dexterous manipulation by pre-training a VLA model on 9.6K hours of egocentric human videos and post-training on robots. It achieves robust execution of free-form instructions across 40+ tasks, with failure recovery and few-shot adaptation to long-horizon tasks like box folding at 75%+ success.
EgoSteer scales dexterous VLA pre-training from 9.6K hours of egocentric human videos with 9x throughput improvement.
The system integrates EgoSmith data pipeline, unified robot stack, and world-model-enhanced VLA.
Real-image diffusion inversion faces a quality-cost trade-off. This paper reveals two mechanisms: element-wise compression asymmetry and trajectory binding, leading to Noise-Anchored Reverse Correction (NARC), a training-free method that outperforms baselines with drastically reduced storage.
A paper accepted at ECCV 2026 presents a new approach to wearable motion capture that works with any combination of consumer devices like smartphones and smartwatches, introducing the WHIP model and a comprehensive dataset spanning 50 activities, along with a systematic study of sensor complementarity.
Proposes WHIP model for full-body motion reconstruction from arbitrary wearable sensor subsets
Introduces large-scale multi-modal dataset with consumer-grade sensors and ground-truth 3D motion across 50 activities
Multi-Modal Knowledge Graphs (MMKGs) enrich entities with modalities like text and images, but entities with highly similar multi-modal features remain hard to distinguish. Temporal information can serve as an additional modality for disambiguation, yet existing approaches rarely treat time as a separate modality due to sparse temporal semantics and noise from multiple timestamps. This paper proposes Time Imprint, a framework that treats time as an entity-level modality and aligns temporal, textual, and visual representations via a three-view contrastive objective. It also designs a compact timestamp subset selection with attention pooling to balance specificity and robustness. Experiments on three MMKG benchmarks show state-of-the-art link prediction, with Hits@1 improvements up to 6.07% overall and 58% on the top-1% ambiguous samples.
Treats time as a separate modality in multi-modal knowledge graphs with three-view contrastive alignment.
Addresses multi-timestamp ambiguity via compact timestamp subset selection and attention pooling.
This paper presents the first demonstration of a low-power MCU-based edge device for Automatic License Plate Recognition (ALPR). It uses a 9-core RISC-V processor (GAP8) with a QVGA ultra-low-power grayscale imager, employing SSDlite-MobilenetV2 for detection (38.9% mAP) and LPRNet for recognition (>99.13%). The system achieves 1.09 FPS at 117 mW, is 73x more energy-efficient than a Raspberry Pi 3 solution, and works on license plates as small as 30x5 pixels.
First MCU-based ALPR edge device using a 9-core RISC-V processor (GAP8).
Multi-model pipeline: SSDlite-MobilenetV2 for detection (38.9% mAP) and LPRNet for recognition (>99.13%).
A new AI system called ReflectWorld-MM enables assistants to continuously process and remember open-ended video streams by organizing memory around persistent entities rather than frames, achieving state-of-the-art results on six benchmarks.
ReflectWorld-MM organizes video memory around entities, not frames, improving long-term tracking.
The system has three components: perception front-end, hierarchical long-term memory, and a real-world realization.
RSLoRA is a training-free, gradient-free method for allocating LoRA ranks based on activation-space geometry. It introduces virtual representational probing to identify high-sensitivity layers, outperforming state-of-the-art allocators like AdaLoRA and GoRA.
RSLoRA eliminates the need for iterative training-time adjustments and backward gradients.
It uses Effective Rank and Fréchet Distance to measure manifold displacement from structured low-rank noise.
WiCAT, a multi-subject model using self-supervised pretraining, outperforms single-session models and enables zero-shot behavior decoding on unseen subjects in widefield calcium imaging.
WiCAT introduces an atlas-grounded tokenization scheme without session-specific components, learning globally shared spatiotemporal representations.
The pretrained model supports lightweight downstream decoding and transfers across subjects, tasks, and datasets.
Researchers propose DUNE, a training-free framework that refines diffusion models by detecting and suppressing early-stage fluctuations in deep latents, reducing artifacts and hallucinations while improving fidelity across both U-Net and Transformer backbones.
DUNE identifies and mitigates artifacts by analyzing abrupt early-stage fluctuations in deep latent variables.
The framework operates without retraining, using an EMA-based criterion for detection and backbone-specific suppression.
This paper investigates the feasibility of training a reasoning language model in Japanese. By applying GRPO to a Japanese continually pretrained model based on Qwen-3-Swallow-8B, the authors find that reasoning-language control is achievable, yet performance at best matches English-reasoning baselines. On Japanese cultural benchmarks, the model performs worse, indicating that reasoning in Japanese does not automatically improve culturally relevant tasks.
Explores training a reasoning model to reason in Japanese.
Developed a Japanese-reasoning variant of Qwen-3-Swallow-8B using GRPO.
This work adapts an open-source spoken language model (SLM) to the Singaporean Home Team domain using LoRA fine-tuning, a surrogate text-QA dataset, and a multi-task objective with CoBa reweighting. The resulting model, HT-Moonstone (5B), matches or outperforms SLMs 7x its size on most tasks and achieves top accent and gender recognition with less than 2% loss in original speech QA ability.
Combines LoRA, surrogate dataset, and CoBa reweighting to adapt SLM to sensitive domains
Builds HTD-multilingual-QA, a 504,853-sample multilingual QA dataset
A new study shows that post-training quantization can silently alter how large language models reason even when task accuracy is preserved. Using a six-category failure taxonomy, the researchers classified 30,000 chain-of-thought outputs and found that hollow convergence exhibits a size-dependent shift under NF4 quantization, while shortcut collapse and confidence snowballing undergo qualitative changes. Hollow convergence cannot be reliably detected from surface-level text features, posing a deployment risk.
Post-training quantization can silently alter LLM reasoning while preserving accuracy
Hollow convergence decreases sharply for smaller models under NF4 but remains stable for larger ones
This report studies on-device English-to-Traditional-Chinese subtitle translation for Taiwan under short inputs, short outputs, batch-size-one inference, low latency, and privacy constraints. The authors replace the original 151k-token vocabulary with a 64k-token subtitle-domain tokenizer, perform embedding calibration and fine-tuning, achieving a 59.2% tie-excluded win rate against Google Translate on a subset of OpenSubtitles2024, and a 1.63x speedup on Apple M2.
On-device English-to-Traditional-Chinese subtitle translation optimized for short inputs, low latency, and privacy.
Replaced 151k-token vocabulary with a 64k subtitle-domain tokenizer; embedding calibration and fine-tuning applied.
A new benchmark framework evaluates the faithfulness of LLM-generated clinical trial summaries across three stakeholder audiences. Using 200 stratified trials from ClinicalTrials.gov and a six-dimension annotation schema, the study identifies 'Unsupported Claims' as the dominant failure mode. A knowledge-graph-augmented retrieval system shows statistically significant improvements in faithfulness scores.
New benchmark evaluates LLM faithfulness in clinical trial summaries for three audiences.
Unsupported Claims is the dominant hallucination across all tested models.
A language-model forecasting system for merger arbitrage, utilizing long-context reasoning over technical documents, outperforms market-implied probabilities and frontier LLMs on a dataset of over 400 large deals across 42 countries.
The system predicts three outcomes: closing at announced terms, higher bid, or deal termination, using expert-guided context engineering and finetuning on hindsight reasoning traces.
It achieves a class-balanced Brier score of 0.151, 24% lower than calibrated market-implied probabilities, 19% lower than XGBoost, and 25-42% lower than frontier language models.
CLIR-Bench is a benchmark for evaluating models on question answering over irregular clinical time series. It is constructed from de-identified ICU records using a principled four-stage pipeline, comprising 6,600 QA instances covering 11 clinical variables, organized into four capability dimensions and 11 tasks. Experiments reveal that current generalist models struggle to retrieve and reason over sparse clinical evidence, highlighting the need for stronger irregular time-series reasoning methods.
CLIR-Bench contains 6,600 QA instances across 11 clinical variables and 11 tasks.
It focuses on irregularly sampled clinical time series QA, filling a gap in existing benchmarks.
Researchers introduce a reference-based membership inference method to detect whether large language models are distilled from other models. By comparing a student model's preference for outputs from different candidate teachers against an earlier checkpoint, the method identifies the most likely teacher with near-perfect accuracy, handling unknown distillation pipelines and open-world settings.
Proposes reference-based distillation detection using earlier checkpoints to identify teacher models
Achieves near-perfect accuracy in single-teacher distillation scenarios
A new study reveals that coding agents need minimal context when editing code: the signal is only in the code being edited, natural-language summaries fail to answer behavioral questions, surrounding context (UML skeletons) performs no better than deleting it, and compressed context matches full files at one-third the tokens. Temperature-0 inference introduces a ~9% noise floor. The authors release their instrument including gold-validated environments, deterministic patches, and pre-registered hypotheses.
The signal for editing lives solely in the code being edited; natural-language summaries answer almost none of the behavioral questions that source code does, regardless of summarizer size.
Surrounding context rendered as UML skeletons resolves no more issues than outright deletion (N=70, p=0.75).
A new paper introduces MawForge, a system that enables practical local inference of Sparse Mixture-of-Experts (MoE) language models on memory-constrained unified-memory machines by storing the model on disk and materializing expert tensors on demand into a bounded cache. The system is effective as a measurement substrate but not as a cache-maximization policy.
MawForge stores the full MoE model on disk and materializes routed experts into a bounded execution cache.
It is designed for local inference on constrained unified-memory machines.
This paper proposes a novel two-level taxonomy for GNN-based knowledge graph technologies, covering construction, embedding, reasoning, and applications, and reviews various GNN models, discussing their strengths, limitations, and future directions.
Proposes a two-level taxonomy combining KG pipeline and GNN perspective.
Comprehensively reviews GNN models like GCN, GAT, and HGNN across KG tasks.
This paper presents a closed-loop control framework using a small language model (SLM) aligned via Group Relative Policy Optimization (GRPO). The system integrates an action agent, a digital-twin validator, and a reprompting agent to iteratively correct outputs. In thermal control simulations, it achieves 91.5% action-alignment accuracy with 3.84s inference latency, demonstrating viability for edge autonomous control.
Compact 1.5B parameter SLM (Qwen2.5-1.5B) aligned via GRPO for control reasoning
Multi-agent architecture: action generator, symbolic/digital-twin validator, and reprompting agent for iterative correction
YUKTI is a novel framework for robust decision-making from natural language, using uncertainty-typed proposition graphs and Assumption-Robust Pareto Frontiers (ARPF). It reduces mean and tail regret by over 90% under misspecification, outperforms a status-quo baseline by 34% on a real dataset, and incurs 47x less regret than an LLM-based approach.
YUKTI replaces fragile point-value optimization with uncertainty-typed proposition graphs and assumption resampling.
It introduces Assumption-Robust Pareto Frontiers (ARPF) to score action robustness and prove a regret bound.
A new study investigates how message format affects information fidelity in multi-hop LLM agent relays, finding that effects are tier-dependent. Under strong relays with faithful instructions, loss is minimal, while weak relays show large inter-format variability. Structured formats provide a faithful, error-localizing channel, not an error-correcting code.
The study tests five message formats over six hops using a controlled relay testbed.
Strong relays are nearly lossless under faithful instructions; weak relays show an 8.7x spread in recall across formats.
This study introduces the Format Sensitivity Index (FSI) and Parseability Sensitivity Index (PSI) to measure how prompt wrappers affect LLM accuracy and answer parseability. Experiments on 140,000 generations show mean FSI varies by over 30x across models, largely explained by compliance failures. Parseability remains a strong predictor of accuracy even after controlling for task, model, and wrapper. Recommendations for robust benchmarking and structured-output deployments are provided.
Introduces FSI and PSI to quantify accuracy and parseability ranges due to wrapper choice.
Across 140k generations, mean FSI varies over 30x across models, mainly due to compliance failures.
This paper proposes a structured diagnostic assistance framework based on the Toulmin model of argumentation, decomposing image-based ML diagnoses into claim, grounds, warrant, qualifier, rebuttal, and backing. Using a specialized biomarker extractor, a MedGemma agent for medical knowledge, and MedSigLip for image similarity, the system presents human experts with interpretable components for critical assessment of ML outputs.
Decomposes ML image diagnoses using the Toulmin argumentation model for interpretability.
MedGemma agent analyzes the warrant linking grounds to the claim.
A new book claims AI has been built on a flawed assumption dating back to Alan Turing's famous 1950 paper. Peter J. Denning argues that the most important parts of human intelligence, including common sense, intuition, culture, and practical know-how, cannot be encoded into computers. He believes this makes true human-level AI impossible, regardless of how large language models become.
Computer scientist Peter J. Denning challenges Turing's assumptions about AI in a new book
Denning argues tacit knowledge like common sense, intuition, and culture cannot be encoded in machines
Anthropic has launched Claude Sonnet 5, its most agentic mid-tier model, outperforming Sonnet 4.6 across all benchmarks and narrowing the gap to Opus 4.8. It introduces effort levels to control reasoning costs, offering great value at low/medium effort but potentially exceeding Opus 4.8 cost at extra-high effort. It is now the default model for Free and Pro plans and accessible via API.
Sonnet 5 beats Sonnet 4.6 on SWE-bench Pro, OSWorld-Verified, and HLE, approaching Opus 4.8. scores.
Pricing is lower than Opus 4.8: $2/$10 per million tokens intro (until Aug 31, 2026), then $3/$15.
Peter Gostev created DOOMQL, a Doom-like game that uses SQLite as the game engine, featuring a recursive CTE ray tracer. Simon Willison demonstrates how to play it and build a Datasette app to view the game state in real time.
DOOMQL is a Doom-like game where SQLite serves as the game engine
It uses a recursive CTE to implement ray tracing in SQL
Simon Willison shares a GitHub code frequency chart for his Datasette open source project, illustrating the impact of coding agents and Opus 4.5-class models, with a huge spike in activity in 2026 aligned with releases like Opus 4.8, GPT-5.5, Fable 5, and GPT-5.6 Sol.
GitHub code frequency chart shows weekly additions and deletions for Datasette project.
Large spike in 2026 aligns with Opus 4.8, GPT-5.5, Fable 5, and GPT-5.6 Sol.
OpenAI's latest family of models, GPT-5.6 Sol, Terra, and Luna, is now generally available on Amazon Bedrock. Sol is a flagship reasoning model with state-of-the-art performance, Terra offers balanced capabilities for production, and Luna provides fast, low-cost inference. Amazon Bedrock's next-gen inference engine provides burst handling, prompt caching with 90% discount, and hardware-enforced security. Additionally, OpenAI launched ChatGPT Work and Codex agents.
GPT-5.6 Sol, Terra, and Luna are now GA on Amazon Bedrock.
Sol sets new records in coding, security, and agent tasks; Terra is for everyday production; Luna for high-volume low-latency tasks.
Outlines is an open-source library that introduces deterministic certainty into LLMs' output generation process for better, more reliable generation of structured outputs.
Outlines masks illegal tokens during inference to enforce output structure.
It supports multiple-choice classification, JSON object generation, and pure JSON generation.
A new wave of social engineering attacks, dubbed 'SociaLLM Engineering,' targets AI agents powered by large language models. These attacks manipulate LLMs into revealing sensitive information or performing unauthorized actions by exploiting their implicit social understanding and lack of trust boundaries. Real-world cases include Instagram account takeovers, GitHub workflow data leaks, and 'Bioshocking' of AI browsers. The article examines why LLMs are particularly vulnerable—due to their design to please users, single-channel processing, and lack of memory—and suggests mitigations such as human oversight and robust guardrails.
SociaLLM Engineering uses social engineering techniques like impersonation and pretexting to manipulate LLM agents.
Notable incidents include mass Instagram account takeovers in 2026, GitHub Gitlost prompt injection, and Bioshocking attacks on AI browsers.
Indian companies are increasingly relying on Chinese large language models from DeepSeek, Alibaba, and Moonshot AI to curb AI spending, extending India's dependence on Chinese cutting-edge technology despite historical tensions.
Indian firms turn to Chinese LLMs to reduce AI costs
DeepSeek, Alibaba, and Moonshot AI are key providers
Stanford researchers present TRACE, a system that diagnoses missing capabilities from agent failures, synthesizes verifiable training environments for each, trains LoRA adapters via GRPO, and composes them with token-level MoE routing. It achieves +15.3 points on τ²-Bench and 73.2% Pass@1 on SWE-bench Verified.
TRACE identifies capability gaps via contrastive analysis of successful and failed trajectories.
Each capability gets a dedicated synthetic environment with algorithmic rewards.
This study identifies vascular metrics associated with navigation difficulty and develops an automated pipeline for quantitative feature extraction to enable future complexity grading. Vascular trees from 61 patients were analyzed using a Soft Actor-Critic RL algorithm for 120 s autonomous navigation. Results show that left-side bovine arch and type II/III aortic arch increase navigation time by 30.19 s and 37.92 s, respectively, while greater tortuosity prolongs procedure and reduces success. On the right side, type II/III arches extend time by 45.94 s, and each additional reverse curve adds 3.96 s. The pipeline provides a foundation for standardized complexity grading and RL model evaluation.
First demonstration that MT agent navigation difficulty is strongly influenced by vascular geometry.
Automated pipeline for quantitative characterization of vascular features developed.