Models

US policymakers are debating whether to create regulatory risk around Chinese open-weight models. The release of Moonshot AI's Kimi K3 reignited the argument. Enterprises face not just performance questions but whether these models will remain easily accessible in a year.

The release of Chinese open-source AI model Kimi has triggered a public feud within Trump's AI advisor circle. Kimi rivals the intelligence of paid models from OpenAI and Anthropic but is free, creating economic and political problems for the president. Factions disagree on whether to embrace open source or impose tighter controls.

Large language models often succeed on one formulation of a problem while failing on an equivalent formulation. Whether these failures arise from distinct internal circuits or different activation states of a shared circuit remains unknown. This study investigates whether arithmetic heuristic neurons are form-invariant across symbolic arithmetic, natural language word problems, and Python code in Llama-3 models. Using a two-stage pipeline of attribution patching and activation patching, we identify a compact set of neurons shared across all formats. Targeted interventions show this shared circuit is necessary and sufficient for late-layer arithmetic computation. Transferring activations from successful to failed executions recovers over 97% of errors for addition and subtraction, indicating cross-format failures arise from activation states rather than distinct circuits. Shared neurons consistently belong to the same heuristic families, demonstrating neuron-level form-invariance.

This paper presents SpecLA, a speculative decoding runtime for stateful linear-attention models. It verifies chains and trees with topology-aware kernels, stores compact factors to recover accepted states, and uses confidence pruning plus a target-aligned EAGLE-style drafter. On an NVIDIA H100 with a GDN-1.3B target, SpecLA achieves up to 1.70x end-to-end speedup over autoregressive decoding.

OpenLanguageModel (OLM) is an open-source PyTorch library for building and pretraining small language models with visible internals. Model code mirrors architecture using components like Block, Residual, Repeat, and Parallel. It integrates tokenizers, datasets, optimization, mixed precision, callbacks, checkpoints, and hardware-aware execution, enabling seamless transition from teaching notebooks to full pretraining. The library includes 27 presets across nine model families and documentation from fundamentals to research. Validation shows close agreement with reference implementations, 90.6% weak-scaling efficiency on four GPUs for a 348M-parameter model, and positive usability feedback. OLM is MIT-licensed and available on PyPI, GitHub, and its documentation site.

Existing memory systems for long-horizon LLM agents often retrieve prior traces as passive context rather than converting them into executable capabilities. This paper proposes MSCE, a training-free Memory-Skill Co-Evolution framework that organizes agent experience into grounded step traces, reusable procedural policies, and declarative environmental cognition. MSCE crystallizes evidence-backed L2 policies into callable skills and introduces reflection-weighted value backfilling. Experiments show significant improvements over state-of-the-art baselines.

This paper presents the NOWJ team's methodologies across all five tasks of COLIEE 2026, including legal case retrieval, entailment, statute law retrieval, textual entailment, and judgment prediction, using adaptive pipelines and deep learning.

This study compares human and language model (LM) next-word prediction using EEG event-related potentials (ERPs), finding that only the 'surprisal' measure correlates with language-processing ERPs, especially for open-class words, and that scaling LMs does not guarantee more human-like cognitive processing.

A new study uses a simple car-wash question to reveal that language models often pre-commit to an answer before reasoning, failing to derive logically correct conclusions. Experiments on Qwen3-8B show systematic wrong commitments (recommending 'walk' when 'drive' is the only valid option). Activation-level analysis suggests hidden states already lean toward the wrong answer before output, even for rollouts that eventually answer correctly. The findings highlight a pre-reasoning decision bias in LLMs.

Small language models in retrieval-augmented generation are highly sensitive to noisy evidence. This paper introduces RIMS, a three-stage preference optimization framework featuring synthetic chain-of-thought data generation, a differentiable soft aggregation mechanism, and preference optimization. Experiments show consistent gains on multi-hop QA benchmarks.

Mixture-of-Experts (MoE) scales Transformers by routing tokens to a subset of experts, but existing routers use shallow or isolated token representations, leading to unstable and semantically inconsistent routing. This work proposes Multi-level Context Fusion MOE (MCF-MOE), which integrates cross-layer semantic aggregation and local token-level interactions for more context-aware representations. Experiments show improved routing consistency and downstream performance.

This paper proposes a token-level cross-modal transformer with contrastive multi-task learning to integrate genomic and clinical data for joint breast cancer subtype classification and survival prediction, overcoming limitations of coarse modality interaction, simple fusion, and independent optimization.

This paper introduces 'weights to words', a method that automatically discovers natural language-described preference dimensions from choice data, addressing under-determination and opacity in preference learning. Experiments show improved prediction accuracy and enable real-time user inspection and editing.

This paper proposes an operator-aware mixed-precision tolerance calibration method that mines accumulated cloud GPU run data to automatically determine optimal absolute tolerances for tensor kernel correctness tests, achieving much tighter tolerances than hand-picked ones and significantly improving bug-detection recall.

LLMs deployed in critical systems pose risks due to their inability to forget sensitive data. This survey examines gradient-based unlearning methods, questioning whether they genuinely remove knowledge or merely suppress expression, highlighting security and robustness challenges.

This paper presents ARGO, a smart eyewear platform leveraging the STM32N6 microcontroller and its integrated NPU for on-device machine learning, minimizing latency and preserving privacy. Through holistic co-design of hardware, firmware, and AI, an optimized YOLOv11 model is deployed for real-time urban obstacle recognition, introducing Head-wise Parallel Attention (HPA) for efficient NPU execution. The model achieves mAP50-95 of 24 with only 2.483 MB memory footprint. The platform integrates multimodal sensors, runs at 10 FPS, and provides ~113 minutes of autonomy on a 200 mAh battery, demonstrating the feasibility of high-performance, privacy-preserving assistive devices.

This paper proposes DocOCR-Eval, an annotation-free evaluation framework for automatic OCR assessment and selection. It employs a three-staged correction and ranking strategy to approximate annotation-based tool ordering without ground-truth labels. Experiments show that aggregating multiple MLLMs improves alignment with human rankings, enabling reliable tool selection in label-limited settings.

A new reinforcement learning method, W2SPO, uses weak auxiliary models to inject short segments into LLM reasoning paths, improving performance and training speed on math reasoning tasks.

Recent growth in reinforcement learning (RL) requires diverse training environments. World models can simulate environment states, but autoregressive models suffer from left-to-right bias. Masked diffusion language models (MDLMs) overcome this via bidirectional anchor-aware denoising, achieving better coherence and rollout diversity than LLMs 4x their size. A GRPO training framework is introduced, achieving up to 47% absolute gains on zero-shot transfer to out-of-distribution environments. The dataset and code are open-sourced.

This paper introduces Generative Ontology Induction (GOI), a domain-agnostic framework that induces a generative blueprint—entities, dimensions, properties, relationships, and constraints—from a corpus of examples and exports it as a typed graph (six node types, seven edge types) in YAML/JSON. A novel evaluation metric, Node Coverage Score, measures structural coverage. Validation on four diverse ontologies shows 95-100% coverage with GOI prompting, while a generic template drops significantly on unfamiliar domains.

arXiv:2607.16200 presents agrepl, a CLI framework for deterministic replay of AI agent executions. Using a MITM proxy, it records external interactions and replays them in isolation, achieving perfect fidelity (F=1.0) and 98.3% latency reduction.

A new research paper introduces PlanFlip, a framework of four planning-phase prompt injection attacks against multi-agent LLM systems. The study finds that stronger models like GPT-5 are more vulnerable, homogeneous backbones create a correlated-agent blind spot, and reasoning-augmented models like DeepSeek-R1 resist attacks. Two defenses are proposed with high detection rates.

A new study introduces a cross-domain framework to test LLMs' risk attitudes under uncertainty, finding that most models exhibit stable and consistent risk preferences both within and across tasks, with a distribution more restricted than humans.

This study presents an open-source MATLAB framework for compact deep learning models for affective touch recognition in soft interactive companions. After exploring 468 CNN models, a 13.2k-parameter dilated 1D CNN achieved 75% test accuracy and 85% leave-one-subject-out cross-validation accuracy. A hybrid pipeline combining heuristic filtering and CNN classification enables real-time 20 Hz operation on a microcontroller, enabling privacy-preserving emotional touch interpretation.

Concentrate is a managed LLM gateway that provides a single API to access over 130 models from major providers. It offers features such as model routing, spend tracking, security controls, and fallback redundancy, designed for teams scaling AI in production.

Dean Ball, OpenAI's chief of strategic futures, argues that open-weight models are inherently decelerationist because they deter capital expenditure. This article examines the economic logic behind the claim, considering China's capacity to build AI infrastructure, shifting training paradigms, and where value accrues in the AI stack. It concludes that open-weight models may actually be accelerationist by lowering costs and broadening participation, even if they disrupt current business models.

Italian engineer Vincenzo (JustVugg) created Colibrì, a proof-of-concept that runs the 744-billion-parameter GLM-5.2 model (1.5TB) on a modest CPU with only 25GB RAM and 1GB/s NVMe. Despite extremely slow speeds (0.05-0.1 tokens per second), it leverages the Mixture-of-Experts architecture to load experts per token, achieving frontier-level answer quality. The project is open-source and aims to explore running large models on consumer hardware.

A live Twitch stream pits GPT-5.6 Sol against Kimi K3 in a speedrun of Kerbal Space Program, showcasing AI real-time decision-making in complex games.

Alibaba’s Tongyi Lab has released Qwen-Audio-3.0-TTS, a production-oriented TTS system with two variants: Flash for real-time interaction and Plus for high-quality generation. The hosted model covers 16 languages and 20 Chinese dialects, features natural-language style control and 86 fine-grained inline tags, and ranks first on the Artificial Analysis leaderboard.

Anecdotes show coding agents have slashed the cost of reverse-engineering home devices. The ROI equation has changed: low effort for automation, low cost of failure, and less psychological baggage from maintenance.

Ben Thompson proposes US legislation to clarify that training data collection is fair use, and to bar terms of service that forbid distillation, in order to help US open models compete with Chinese counterparts. Additionally, Alibaba's release of Qwen 3.8 Max as open weights may have been influenced by Xi Jinping's recent speech encouraging open source.

Moonshot AI released Kimi K3, a 2.8T parameter MoE model with open weights, ranking high on benchmarks. The article discusses the narrowing gap between Chinese and US AI models, China's commitment to open source, economic impacts of open models, and China's efficiency advantages.

Adobe's Indigo camera app, initially focused on natural iPhone photography, now adds generative AI editing tools via 'AI Playground', using Google's Nano Banana model. Features include AI styles, object removal, photo guidance, and custom editing. The experiment offers free access to a small user group and may become paid.

Inertia-1 is a unified motion foundation model that tackles fragmentation in motion datasets. It uses large-scale self-supervised learning on wrist accelerometer data to create representations that generalize across body placements and sensor types, and reveals key design choices for real-world performance.

Opaque pricing and backward-looking bills force enterprises to rethink their AI model strategy.

A study by Meta's Oversight Board finds that leading AI systems, including those from U.S. companies, are more likely to refuse requests to criticize restrictive leaders or governments, raising concerns about extending state influence on free speech globally.

Chinese AI leaders Moonshot and Alibaba released models that claim to match top US systems at lower cost. Their open-source approach challenges US dominance and raises questions about the effectiveness of export controls and massive spending.

OpenAI shares lessons from deploying long-running AI models, highlighting new safety risks, observed failures, and improved safeguards through iterative deployment.

Public health departments across the US will test generative AI tools under a new program, PULSE, involving the Coalition for Health AI, OpenAI, Anthropic, and Accenture. The program will run trials in 10 jurisdictions, providing enterprise licenses for up to 2,000 practitioners. It covers five use cases including biosurveillance, social determinants of health, public communications, and automated clinical data retrieval. Pilots are scheduled for autumn 2026, with playbooks expected in 2027.

Moonshot AI’s Kimi K3, a 2.8-trillion-parameter open-weight model, uses mixture-of-experts, quantization, and attention caching to trade compute for memory, circumventing US chip restrictions. While it tops benchmarks in coding, deployment requires data-center infrastructure, pricing is high, and software support is incomplete.

Thinking Machines Lab has released Inkling, its first general-purpose open-weights foundation model. It is a multimodal MoE model with 975B parameters, 41B active parameters, and a 1M-token context window. Designed for customization, Inkling excels in reasoning, coding, agentic workflows, and multimodal tasks. This guide covers its architecture, training, benchmarks, deployment, and fine-tuning workflow.

Zlvox AI Humanizer is a free, unlimited online tool that uses Groq Llama 3.3 to convert AI-generated text into natural human language, bypassing detectors like GPTZero and Turnitin. It offers multiple humanization levels, tone adjustments, grammar fixing, paraphrasing, and summarization, all without requiring signup.

In mission-critical scenarios like disaster inspection and search-and-rescue, communication-limited robots must make reliable onboard decisions. Episodic memory reuse, though low-cost, can be unsafe due to changed topology or insufficient resources, leading to 'memory traps'. This paper presents MemoGuard, a lightweight adaptive runtime that validates memories against topology, resource, and outcome contracts before reuse, invoking fallback only when validation fails. In a corridor-inspection simulator, MemoGuard reduces battery safety violations by 76.6% over similarity-only top-1 reuse and reduces fallback calls by 21.4% over always reasoning. On an NVIDIA Jetson AGX Xavier with local llama3.2:3b fallback, it avoids 3.67 s and 36.97 J overhead per trial.

PACE proposes a framework for dynamically generating personalized, psychologically grounded robot personas via interactive Q&A. Integrated on the Ameca robot, it significantly boosts user trust, anthropomorphism, and interaction quality over static baselines.

This paper presents robust, repeatable, and scalable fabrication procedures for soft pneumatic actuators using two-part silicone pour casting, including methods to prevent internal cavity clogging and ensure airtight sealing, as well as a robust sensor embedding procedure for thin-film flex sensors. Finite Element Modeling, PID-controlled pneumatic experiments, and automated image processing calibration validate the actuator performance. Staircase and sinusoidal actuation tests demonstrate high repeatability and low hysteresis, validated across two operators and 24 successful fabrications.

Xiaomi Robotics Team presents Xiaomi-Robotics-1, a foundational VLA model capable of following diverse language instructions in unseen environments and efficient fine-tuning for novel tasks. The two-stage training uses over 100k hours of real-world trajectories with an auto-labeling pipeline. It achieves state-of-the-art results on RoboCasa365 (57.6%) and RoboDojo (20.07). Code and models will be released.

Cross-view geo-localization matches ground-level observations to satellite imagery. Recent methods use sequential queries like video clips for richer spatiotemporal cues, but overlook route descriptions. This paper introduces SeqGeo-VL dataset (~39K video-text-satellite triplets) and TrajLoc framework that processes both video and text, leveraging dense visual and linguistic semantics. TrajMod module conditions embeddings on trajectory geometry. Experiments show significant gains over state-of-the-art on video and text geo-localization.

A Partial Information Decomposition (PID) framework is used to select the most informative MRI contrast pair prior to training, reducing computational cost for multi-contrast 3D brain tumor segmentation. Applied to T1n, T1c, T2w, and T2-FLAIR, it selected T1c+T2-FLAIR, which achieved mean Dice 0.676 vs 0.687 for all four inputs on lightweight 3D U-Nets.

Multimodal large language models struggle with part-level grounding. The proposed Object-Part Hierarchical Reflective Grounding (OP-HRG) uses a coarse-to-fine reasoning approach, first localizing the parent object then the part, with self-check and re-encoding. A part-aware GRPO framework with stage-wise rewards trains a 4B model that outperforms 7B grounding LLMs and SAM3 on several benchmarks.

This research proposes a fully training-free open-vocabulary 3D point cloud segmentation method that uses frozen vision-language models (RegionPLC) and a promptable concept segmenter (SAM3) with cross-view consistency to achieve generalized few-shot segmentation, significantly improving novel class performance on ScanNet200 and ScanNet++ benchmarks without any training or few-shot support.