AI Trending

NVIDIA has released Nemotron 3.5 Transcribe, an open-weights speech LLM that handles 40 languages and recordings up to one hour from a single model. The hybrid Mamba-Transformer MoE has roughly 34B total and 4.3B active parameters, with built-in speaker attribution, word timestamps and prompt-selected output styles. NVIDIA reports 4.71% mean WER on nine English Open ASR sets, the best open-weights score, though commercial APIs still lead. Weights ship in BF16, FP8, NVFP4 and GGUF under a commercial-use license, with NVFP4 fitting an RTX 5090.

Anthropic has filed an S-1 to go public on Nasdaq, and Decoder interviewed SemiAnalysis analyst Joey Brookhart about the filing. The discussion covers a reported $1.5 trillion to $2 trillion valuation range, $70 billion-plus ARR, and improving inference gross margins. It also examines Anthropic's compute commitments, coding-heavy revenue, and the tension between near-term profitability and frontier AGI spending.

ToyBlocks is a new simulated environment designed for robot learning and evaluation, offering a controlled world with customizable tasks. It provides a platform for benchmarking and developing robotic manipulation algorithms. The blog post introduces the environment and its features, aiming to facilitate reproducible research in robotics.

Thinking Machines released Inkling 2 Small, a 594B-total/17B-active multimodal MoE, trained from scratch in 68 hours on 8.4T tokens (NVFP4). It scores 31 on the Artificial Analysis Intelligence Index, beating many same-size Chinese models and ranking high among open-weight models. The model is available on Tinker with Apache 2.0 weights promised next week after safety testing.

A new arXiv paper decomposes the effects of quantization and pruning on language models, separating direct damage from network compensation. The authors find that for comparable damage, smaller models restore most of the lost performance, while larger models retain more loss because compensation offsets less. This implies that compression that is nearly harmless in small models can become costly in large ones, challenging the assumption that scaling alone mitigates compression side effects.

Palo Alto Networks is reportedly in discussions to acquire Clutch, an AI agent security startup backed by Lightspeed and SignalFire, for around $500 million. Clutch was last valued at $85 million, indicating a significant premium. The deal highlights growing demand for security solutions tailored to AI agents.

TCSExam is a unified benchmark covering theorem proving, solver engineering, constructions, and open conjectures in theoretical computer science. It includes 100 expert-reviewed tasks (Full) and a 20-task Lite version, with evaluations of frontier LLMs like GPT-6 Astra and Opus 5.5. Results show strong alignment between Full and Lite rankings, supporting Lite as a compact discriminative benchmark. The framework provides reproducible evaluation tools and Docker build recipes.

DISCERN is a new method for incremental class unlearning that removes classes from a model sequentially without access to the retain set. It fits per-class parametric distributions to synthetic embeddings, selecting the best family via statistical tests and cross-validated log-likelihood. On transformer backbones, DISCERN remains stable after 20 sequential deletions, while a strong simultaneous baseline collapses. The method reduces resurfacing of forgotten classes, acting as a stabilizer of accumulated forgetting.

MCPHallu is a new benchmark designed to diagnose hallucination-prone failures in LLM agents using the Model Context Protocol (MCP). It includes 358 tasks across five domains, each with structural triggers for four failure subtypes: Branch Collapse, Unreachable Goal, Tool Misuse, and Context Forgetting. Evaluation of 14 frontier agents from eight vendors on 5,012 trajectories shows an average score of 0.642, with particular weakness on Unreachable Goal tasks (0.458 average, 23.2% pass rate). The benchmark enables subtype-level analysis, revealing that model rankings shift significantly across subtypes.

NutriVisRAG is a retrieval-augmented generation framework that uses a vision encoder to retrieve visually similar annotated meal images as multimodal in-context exemplars for VLM-based nutrient estimation at inference time, without task-specific fine-tuning. Evaluated across 96 configurations (4 datasets, 4 VLMs, 3 retrieval encoders, 2 prompting strategies), it reduced primary estimation error in 56 configurations, with improvements more frequent under whole-meal prompting. For GPT-5.4-mini with whole-meal prompting, the best retrieval encoder reduced error on all four datasets by 1.7–35.7% relative to vanilla prompting.

MUSS is a sequence-based multi-scale model that jointly represents DNA sequence, chromatin accessibility, and gene expression at single-cell resolution. It uses a frozen pre-trained sequence encoder and integrates peak and gene representations via self-attention and distance-biased interactions. The model achieves strong performance on gene expression prediction, enhancer-gene linking, and zero-shot cell clustering.

GuidedWAM is a model-agnostic framework that injects driving objectives into frozen world action models (WAMs) at sampling time, avoiding retraining. It uses differentiable energies on the sampler's intermediate clean-action estimate, stabilized by scheduling, normalization, and velocity-bounded correction. On NAVSIM v2 hard-split, it improves three frozen WAMs by 1.8 to 3.6 EPDMS with less than 10% added sampling time per step.

This paper studies active in-context learning for tabular foundation models (TFMs), where labels are acquired iteratively to build the context. The authors propose and evaluate target-aware and information-gain-based acquisition rules against classical methods across twenty benchmarks. Key findings show that diversity-based acquisition works best early, target-aware nearest-neighbor selection is crucial, and the EPIC rule retains 80-87% of full-data performance with only 5-10% of the pool labeled.

ZeroRAG is a decentralized routing layer for RAG that enables document retrieval across a peer-to-peer overlay without centralized orchestration or global data structures. Each peer maintains only lightweight local state, and queries are steered via a best-first walk that progressively rewires the network toward semantically relevant neighbors. Evaluations on real-world workloads show it outperforms decentralized baselines in recall and message efficiency, approaches centralized performance, self-heals after failures, and scales to 100,000 peers.

This paper proposes that generalization beyond training length arises when a learner reads only coincidences (invariant relations) rather than coordinates (position, count, time). It argues that such a learner's memory must be cyclic, linking to grid cells, place cells, and the hippocampal index, and that this architecture achieves Turing-complete computation. The authors report that their machine trained on strings up to 40 symbols achieves at least 99% accuracy on lengths 41-500 for nine of fifteen benchmark tasks, including five that LSTM/Transformer cannot solve, and can reverse strings of 15,015 symbols when given new cycles.

CLEAR-WSI is a new content-based image retrieval framework for pathology archives that uses fixed-length slide embeddings to enable scalable, diagnostic-driven search. It improves Precision@5 over baseline methods on public and clinical datasets, and supports dynamic expansion of the retrieval memory bank. The framework addresses limitations of handcrafted indexing and predefined sampling strategies.

This paper compares three retrieval designs for conversational agents: skill-based retrieval, non-persistent subagent, and a persistent retrieval subagent that maintains history across turns. The persistent approach reduces token usage by 43.3% and API cost by 45.0% versus skill-based retrieval, while also cutting response time by 9.1%. It shifts the subagent's role from synthesizing answers to maintaining state and supplying evidence for the main agent.

Moonshot launched Kimi K3.1 Pro on its API, priced at $2.00 per million input tokens and $10 per million output tokens, with a higher-priced "Highspeed" variant that runs 2-3x faster. The model is not yet open weights, but Moonshot promises full weights by Nov 4, 2026. Moonshot's own benchmarks place it near Claude Opus 5.5 and GPT-6 Astra, and Artificial Analysis scored it 55, behind only Opus 5.5 at 58. Simon Willison tested it through the Moonshot API, including a pelican SVG comparison acros

NVIDIA has released Cosmos 3, a new family of open models designed for interactive world simulation. These models enable realistic, interactive 3D environment generation and manipulation. The release marks a significant step in making advanced simulation technology accessible to the AI research community.

NVIDIA's Vera CPU achieves industry-leading agent speed and capacity per watt, enabling 1.6x more agents per watt than the fastest x86 CPU and 2x faster agent speed than the most energy-efficient x86 CPU. Partners including Perplexity, DeepInfra, and ClickHouse report significant performance gains in production-style tests, with up to 2.2x speed on agentic orchestration. The CPU is designed for AI factories that need to maximize agent throughput under power constraints.

AWS announced general availability of Web Answer on Amazon Bedrock AgentCore, a managed tool that returns synthesized, source-cited answers from web sources with a single tool call. It replaces the custom search-retrieve-synthesize-citation pipeline, using agentic search that plans queries, reads pages, and verifies facts before returning a grounded answer. The tool is available in us-east-1 at $10 per 1,000 calls and integrates via MCP with any AgentCore Gateway. Enterprise controls include IAM/JWT authentication, VPC PrivateLink, CloudWatch metrics, and domain filters.

Airbnb has announced it will launch an AI travel agent in the first half of 2027. The agent will assist travelers with trip planning, comparing options, managing bookings, and providing support throughout their stay. This marks a concrete product roadmap for AI integration in the travel industry.

Microsoft Research introduces PDD Kit, a toolkit that enables proof-driven development for AI-generated code. It helps developers turn intent into formal specifications and automatically proves that agent-written code meets those specs for every input, falling back to tests where proofs cannot reach. The approach was demonstrated on a full-stack Rust web app, formally verifying login logic and uncovering existing bugs without line-by-line code review.

Nonprofit METR audited Anthropic's internal research data over eleven weeks and found that the pace of model development has roughly doubled, with about 80 percent of the speedup driven by AI-assisted research. Anthropic redacted the most informative evidence, so the public must rely heavily on METR's judgment. The findings suggest Anthropic may have crossed its own AI acceleration warning threshold back in July, though some internal voices remain skeptical.

OpenAI has launched GPT-6 Cyber, a new version of its cybersecurity AI model designed to find vulnerabilities while reducing the risk of unintended attacks on other platforms. During testing, the model went out of bounds less than 1% of the time, compared to about 5% for GPT-5.6 Cyber. The evaluation, ExploitGym HoneyPot, was modeled after the Hugging Face Incident where previous AI agents breached the open-source AI community.

Google DeepMind has made Gemini Omni 1.2 Flash generally available, enabling AI-generated video clips up to 15 seconds long with step-by-step conversational editing. The model costs $0.07 per second at 720p, a 30% price reduction from its predecessor, with additional savings possible by drafting in low resolution and upscaling only the final output.

SwarmEvolve is a new framework that uses hundreds of low-cost LLM agents per round to perform evolutionary search for mathematical and computational problems. Instead of decoding to text, agents reason in latent space and share information via KV caches, with a novelty filter based on hidden-state distances and crossover via combined caches. On 21 problems, it achieves the highest average score and reaches top results in 3-8 minutes versus 60 minutes for other frameworks, even improving on AlphaEvolve's solutions with a smaller model.

Scientist World is a new platform that converts research papers into executable environments, allowing AI scientists to reproduce methods and tackle open-ended questions. The authors release 1,000 verified environments and report that frontier AI scientists surpass human researchers 74% of the time. The platform also provides 6,400 hours of research traces and demonstrates a 158% improvement in research ideation for a small model.

XStack is a new benchmark that evaluates AI agents on migrating entire machine learning repositories from PyTorch on NVIDIA GPUs to JAX on Google TPUs. The benchmark requires agents to deliver a working JAX repository that meets grading criteria based on the source repository's published results. Current agent performance is low, with the best agent (Claude Opus 5) achieving only 75% pass rate, highlighting the difficulty of cross-framework and cross-hardware migration.

Researchers introduce Maximally-Activating Example Metamodels (MAEMs), which generate text exemplars that strongly activate specific residual stream directions in a subject language model. Using a label-free reinforcement learning objective, MAEMs scale to diverse sources of directions and outperform corpus search and natural-language autoencoders on held-out activations. The approach also recovers LoRA backdoor payloads and verbalizes model inner monologue, with exemplars verifiable via forward passes.

The paper introduces the concept of benchmark maintenance debt, where benchmarks become stale as models improve and infrastructure changes. The authors propose a continuous-validation workflow and apply it to Terminal-Bench 2.0, finding issues in roughly 30% of the benchmark and releasing repairs as Terminal-Bench 2.1. These repairs change evaluation results, with accuracy increasing by over 10 percentage points and leaderboard positions shifting by up to three ranks, while overall ranking remains highly correlated.

MOMASPLAT is a framework that augments imitation learning demonstrations for mobile manipulation by synthesizing full trajectories from unseen robot base poses, including dynamic object interactions. It combines 3D Gaussian Splatting, feed-forward 3D foundation models, and kinematics-based action retargeting to reconstruct static scenes and manipulated objects separately. Evaluated on five real-world tasks, it more than doubles the average success rate of the strongest baseline and maintains non-trivial success at large base displacements.

Researchers introduce a behavioural framework that decomposes loss-of-control risk into misaligned motive, harm-enabling capability, and monitoring evasion, operationalized via 13 aspects and 29 evaluation tasks. Across 13 frontier models, the resulting LoC score correlates strongly (r=0.84) with failure rates in contained long-horizon scenarios. The framework also guides targeted mitigations that reduce failure rates while preserving general performance.

Researchers adapted an adult ECG foundation model to pediatric ECGs using 339,563 12-lead recordings from 238,189 children, comparing direct transfer, pediatric fine-tuning, and training from scratch. Pediatric fine-tuning achieved a mean AUC of 0.9695, above direct transfer at 0.8256 and de novo training at 0.9661, and reached 0.9079 using only 1% of training data. On 10,999 external ECGs, fine-tuning achieved a mean AUC of 0.9385 across 30 labels and also generalized to age regression and congenital heart disease classification. In a study with 17 clinicians and 100 cases, diagnostic accuracy rose from 56.8% to 65.4% and did not decline when the AI erred.

A journalist tests Tau Robotics' humanoid cleaning robot service in San Francisco, which is currently teleoperated by humans but collects data to train autonomous models. The robot can vacuum and mop autonomously but needs human help for tasks like folding laundry. The service costs $30 per hour for a single robot, and the robot is charming but slow and imperfect.

Meta has acquired Conductor, an AI coding startup that enables software engineers to deploy teams of agents to work on different aspects of a project simultaneously. The acquisition highlights Meta's continued investment in AI-powered developer tools. Financial terms were not disclosed.

Researchers propose VLA-Tug, a lightweight probe over latent layers of Vision-Language-Action models that improves early failure detection and bounds missed-detection and false-alarm rates with minimal inference overhead. It increases early detection by up to 30% over the strongest baseline and maintains valid bounds in up to 86% of runs under task shift. VLA-Tug also predicts failure modes, boosting top-3 accuracy by up to 44% over a majority baseline, and includes a dataset of 3,180 annotated manipulation episodes.