Models

Kimi K3 refuses to leak its system prompt and responds with "Is there something I can actually help you with today?"

OpenTSLM is a multimodal LLM that treats time series as a native modality, enabling reasoning over raw multivariate signals alongside text. It outperforms baselines, including GPT-4o, on time series QA, activity recognition, sleep staging, and ECG QA. The model scales to multiple long time series with near-constant memory consumption. ECG reasoning validated by 7 cardiologists with 97% correctness. All code, datasets, and models are open-source.

NVIDIA released Nemotron 3 Embed on July 15 and 16, 2026. The collection has three open checkpoints: Nemotron-3-Embed-8B-BF16, Nemotron-3-Embed-1B-BF16, and Nemotron-3-Embed-1B-NVFP4. The 8B ranks #1 on RTEB at 78.46 average NDCG@10. The 1B came from ModelOpt NAS pruning plus COS+MSE distillation from the 8B teacher. NVFP4 retains 99%+ of BF16 retrieval accuracy at up to 2x Blackwell throughput. All three run 32,768-token inputs under OpenMDW-1.1.

This paper proposes ConFlow, a framework that incorporates constraint information directly into the flow matching training objective via differentiable barrier or cost functions and a conditional Gaussian Process, improving constraint satisfaction and trajectory quality in robot motion generation. Experiments on a two-robot navigation task demonstrate lower collision rates and higher trajectory quality compared to standard flow matching baselines.

This paper explores the feasibility of using brain signals via functional near-infrared spectroscopy (fNIRS) to modulate robot reinforcement learning. It compares agents trained on passive (observational) versus active (demonstrative) interaction tasks, and tests multiple methods for enhancing the RL algorithm with the neural signal, focusing on parameter augmentation rather than replacement. The results show that this framework is effective: the neural signal improves learning when augmenting trajectory priorities and state-action q-values. Additionally, the framework learns successfully from offline data, offering a practical alternative for settings where real-time BCI setups are impractical or only limited data is available.

Existing grasp benchmarks focus on visual pose detection, ignoring multi-step reasoning and semantic constraints. GCA-Bench introduces complex action scenarios to evaluate large models. Current methods achieve below 70% success rate, highlighting critical limitations.

DiMaS is a distribution-matching steering strategy for flow-matching vision-language-action (VLA) models, enabling fine-grained behavioral control in robotic manipulation. It transports between representation distributions rather than shifting along a fixed direction, proving effective on two state-of-the-art VLAs. The study also examines transferability and explains why linear steering fails in visuomotor settings: behavioral features are linearly decodable but not linearly steerable.

MEMORA introduces Embodied Action Memory (EAM) to enable robots to use persistent memory from egocentric video for long-horizon planning. It features four typed memory stores, online editing, and offline consolidation. Evaluated on 45 hours of EPIC-KITCHENS-100 video, MEMORA improves memory accuracy by up to 20.5 points and planning scores by 16.6%.

This paper proposes LIFT, a force-aware post-training framework that adds contact reactivity to pretrained vision-language-action (VLA) policies. By grafting a reactive action expert, injecting 6D end-effector force via causal force memory and cross attention, and coupling with an online DAgger loop, LIFT outperforms vision-only post-training in towel folding, book insertion, and Hanoi ring placement.

This work introduces a multi-modal orchestration framework for semantic audio-driven humanoid control, enabling real-time autonomous selection of motion skills based on music or speech input. Validated on the Unitree G1 humanoid, it demonstrates robust sim-to-real transfer.

SD-MAR is a framework for training and evaluating vision-language models (VLMs) on multi-image analytical reasoning tasks. It constructs paired visual scenarios through controlled perturbations and generates reasoning tasks spanning semantic change attribution and quantitative comparison. Using GRPO-lite with Backward Discounted Allocation (BDA), a reinforcement learning approach that removes KL regularization, fine-tuning on SD-MAR improves in-domain accuracy by up to 36.95% on Qwen2.5-VL-7B and InternVL3-8B. Qwen2.5-VL-7B outperforms GPT-4.1 on the SD-MAR benchmark. Out-of-domain generalization is preserved or improved, with performance within 1% on MME, MMMU-Pro, MathVista and up to 4% improvement on MMBench. LLM-as-judge evaluation shows consistent improvements in logical coherence and explanation quality.

Addressing the challenge of defect segmentation in additive manufacturing XCT images, the proposed XCT-SAM framework sequentially adapts SAM using Conv-LoRA adapters, first on an alloy microstructure dataset then on XCT images, outperforming baselines on CycleGAN-XCT benchmarks and real NIST scans.

MonteRET is a region-aware retrieval-enhanced framework for generating chest CT findings sections. It integrates global and regional CT features, retrieves clinically relevant knowledge, and refines reports via a knowledge-guided rewriting agent. Evaluated on public and external cohorts, MonteRET improved report quality, semantic similarity, and clinical efficacy, with experts favoring its outputs.

This paper investigates whether vision foundation models build representations that reflect intrinsic properties of 3D Euclidean space. Instead of regressing depth or normals, the authors probe the relationship between visual feature space structure and Euclidean transformation group SE(3) using a mutual neighborhood metric and a Poincaré Adapter. They show that self-supervised vision models harbor latent subspaces strongly correlated with 3D space, even without 3D supervision. This leads to 'Latent-Space Navigation' techniques for visual odometry and localization without explicit 3D reconstruction.

The paper presents KeyFrame-Compass, the first comprehensive benchmark for evaluating keyframe-conditioned video generation, with 386 curated samples, an automated evaluation framework, and experiments revealing trade-offs between keyframe fidelity and naturalness.

Multi-reference-to-audio-video (MR2AV) generation requires models to produce synchronized audio-video content conditioned on multiple references and textual instructions. Existing benchmarks focus on text-driven generation or single-reference preservation, lacking evaluation for MR2AV. This paper introduces MultiRef-Compass, a unified benchmark with 350 carefully curated samples covering multi-view subject preservation, multi-entity binding, and human-object-scene composition. It defines an evaluation protocol with four dimensions (Basic Quality, Reference Consistency, Audio-Visual Consistency, Instruction Following) and 14 sub-metrics, integrating automatic metrics with a rejudging-enhanced MLLM-as-a-Judge framework. Experiments on eight MR2AV systems reveal substantial room for improvement across all dimensions.

This paper introduces tool efficiency, a new quantitative metric to evaluate the rate of useful tool calls in an LLM agent trajectory. To ensure that tool efficiency is well-defined, it also introduces marginal tool utility, indicating per tool call whether it is useful or safely removable. The sign of marginal tool utility is determined using LLM-as-a-Judge. This work directly measures efficiency, complementing accuracy-based evaluations, and aims to inform future benchmark design and lean tool suite engineering.

Polestar is a training-free inference framework that addresses KV-cache reuse and decoding parallelism challenges in diffusion LLMs by leveraging token representation drift. It consists of Polestar-Cache for sparse cache refreshes and Polestar-Commit for identifying commit-ready tokens, achieving up to 10.73% accuracy improvement and 3.7x higher throughput on math and coding benchmarks.

This paper introduces token time continuous diffusion (TTCD), a diffusion language model operating in continuous space with per-token times, where tokens proceed from noise to token at varying rates. TTCD avoids parallel sampling inaccuracies and outperforms discrete models at high speedups. A 160M parameter model trained on OpenWebText and self-distilled achieves comparable unconditional and superior conditional generation, with gains in Sudoku solving.

The paper introduces AGOPS, an automatic method to generate task-specific prompt guidelines that help users write better prompts, improving LLM performance by recovering large performance drops from underspecification.

A new study quantifies information loss when LLM agents communicate via text, using sparse autoencoder feature analysis. While latent communication preserves more information under compression, the lost features primarily encode surface form rather than task-relevant semantics, questioning the practical advantage of latent channels.

This paper proposes LBA, a sampling-based method for generating high-quality adversarial texts under low query budgets in the hard-label setting. By integrating prior and posterior knowledge to construct an approximate distribution, LBA efficiently samples adversarial examples. Extensive experiments show LBA outperforms state-of-the-art baselines across models and datasets, with better semantic preservation and readability.

This paper presents the first application of pregroup grammar-based quantum compositional NLP to Arabic, a morphologically rich language. Quantum circuits mirror grammatical structure, outperforming classical baselines in word order, tense, and verb sense disambiguation experiments.

The Just Keep Prompting (JKP) framework tests VLM stability under repeated challenging. Evaluations on GPT-4o, Gemini 2.5 Pro, and Qwen3-VL-30B show substantial instability and answer flipping, with model-specific pressure-response profiles.

This paper analyzes why closed-loop knowledge systems (e.g., LLMs, RL) saturate under repeated internal feedback and introduces a three-level operational framework to enable escape via structural interventions. Using Lyapunov drift, stability is characterized, and escape is quantified by attractor displacement and a KL lower bound. Case studies include LLM code repair, sparse-reward RL, and Bayesian optimization.

Offline reinforcement learning world models suffer from model exploitation in low-data regions. RENEW uses human preferences over imagined rollouts to directly repair exploitation, introducing epistemic uncertainty to focus finetuning and improve sample efficiency.

Proposes Branching Policy Optimization (BPO), which leverages deterministic, snapshottable, and resumable sandboxes to construct a tree-structured rollout topology with shared prefixes, reducing policy gradient variance and improving success rates by 3.6–6.1 absolute points over GRPO and RLOO.

This paper introduces C3R, a drop-in control layer that, from an inferred domain posterior and no query-time label, certifies a per-domain contamination budget where feasible and otherwise abstains. It guarantees a reduction on the hardest domains, shows stability across resampling, and retains more recall than calibrated cascades.

Existing zero-shot image classification methods using vision-language models (VLMs) often employ a uniform weighting of prompts across all classes, ignoring the class-specific suitability of prompts. CARPRT introduces a training-free, class-aware reweighting scheme that adjusts the weight vector for each class based on the relevance of prompts to that class. Experiments show that CARPRT outperforms class-independent reweighting methods, highlighting the importance of modeling prompt-class dependencies.

A new study enhances small language model (SLM) reasoning by grounding them in knowledge graphs via a neuro-symbolic agentic framework. Experiments on CLUTRR with Gemma 3 and Llama 3.2 show RGCN-derived hints improve performance by 1.5-2x, but reveal extraction bottlenecks and sequential deductive fragility.

This paper addresses the 'behavioral inertia' problem in tool-augmented LLM agents when expanding their toolset. By injecting counterfactual anchor contexts at critical decision points, the proposed ToolAnchor framework breaks this inertia, recovering failed trajectories. It uses teacher models to hypothesize counterfactuals, verifies them via student rollouts, and internalizes successful interventions through post-training. Evaluated on GAIA, BrowseComp, and VDR-Bench, it shows competitive performance, bridging static post-training and dynamic adaptation.

Researchers propose a novel method using Large Language Models to build Bayesian Belief Networks, employing a panel of AI agents to estimate probabilities based on personas and context, and applying a trimmed-mean rule to reduce noise. A six-step framework is illustrated on customer intention to consult a doctor in an alternative healthcare system, revealing that subjective norms have a much stronger effect than self-efficacy, and the most effective strategy is to improve both confidence and community norms simultaneously.

A new approach called LLM-T1D combines reinforcement learning with large language models to create an interpretable insulin pump controller for Type 1 Diabetes, achieving 73.5% Time in Range while providing clear explanations.

Inspired by human communication of spatial information, language-guided geo-localization has gained traction but relies on static one-shot retrieval, failing to handle ambiguity. This paper proposes a paradigm shift to reasoning retrieval with Dialogue Place Recognition (DlgPR), which casts localization as an interactive dialogue-driven process. The authors introduce DlgQuest-Cities, the first large-scale dialogue-based benchmark for place recognition, and a unified framework with a cross-modal retriever and intelligent questioner DQ-pilot. DQ-pilot is trained via curriculum learning: supervised fine-tuning on DQ-cities-20k then reinforcement refinement on DQ-cities-10k using GRPO. Two metrics guide learning: Discriminative Difficulty Index (DDI) and Positional Retrieval Gain (PRG). Experiments show significant improvements over baselines.

arXiv:2607.14095v1 Announce Type: new Abstract: Retrieval Augmented Generation (RAG) has proven to be a widely successful process at improving the quality of outputs from a Large Language Model (LLM) for wider context. However, RAG systems typically retrieve context from flat document stores, which struggles when queries require hierarchical or relational reasoning across structured knowledge. I present HG-RAG (Hierarchy-Guided RAG), a framework that performs graph-traversal over a hierarchical knowledge graph to deliver structured context to a language model. My retrieval pipeline resolves a named entity anchor from the query, then expands context upward through parent nodes, laterally through relational neighbors, and downward through child nodes when needed. I evaluate HG-RAG against a dense retrieval baseline across three world scales (18-800 nodes) with four query types: local fact, hierarchical, neighborhood, and multi-hop. Results show HG-RAG consistently outperforms the flat baseline on hierarchical, relational, and multi-hop reasoning tasks, while reducing hallucination and maintaining locality coherence.

Alphabet shares fell 4% on Thursday following a report that the company has delayed its flagship AI model, Gemini 3.5 Pro. The model's coding capabilities fell short of internal expectations, while competitors like OpenAI and Meta have released more advanced coding models.

Puter compiled Firefox to WebAssembly, enabling a full browser to run inside another browser. The project used an estimated $25,000 in Claude Opus and Fable tokens, leverages the Wisp protocol for proxying, supports end-to-end encryption, and is open source.

Artificial Analysis released AA-Briefcase agentic knowledge work benchmark results; Kimi K3 scores 1547 Elo, ranking first, surpassing GPT-5.6 Sol's 1495. The benchmark simulates real business workflows evaluating models on spreadsheets, presentations, memos, etc.

OpenAI's GPT-Red uses human-AI collaboration for red teaming, a novel approach to model safety, but enterprises must still ensure alignment with their workflows.

Moonshot AI released Kimi K3, a 2.8 trillion parameter model claiming to be the first 'open 3T-class model'. It outperforms many models on benchmarks but comes with higher pricing. The author tests it with a pelican-on-bicycle SVG prompt, revealing reasoning costs, hidden system prompts, and vision capabilities, while reflecting on the limitations of this informal benchmark.

This article describes an autonomous AI music video generation system that compares Claude Fable 5 and GPT-5.6 Sol under budgets of $25 and $100. The system lets models autonomously research, generate clips, edit, and assemble a complete video. Results show all runs produced valid videos, though quality was average, with issues in consistency and tempo matching. Claude Fable 5 was more expensive but faster, while GPT-5.6 Sol showed more creativity in editing.

xAI's Grok 4.3 is now generally available on Amazon Bedrock, offering configurable reasoning effort, strong tool use, instruction following, and a 1 million token context window for agentic and enterprise workloads. This post covers its features, access methods, and how to use key capabilities such as chat, reasoning, tool calling, structured output, image input, and multi-turn conversations.

OpenAI trained GPT-Red, an internal-only attacker model, using self-play reinforcement learning against a population of defender LLMs. It beat human red-teamers 84% to 13% on a replicated indirect prompt injection arena, found a novel "Fake Chain-of-Thought" attack class, and cut GPT-5.6 Sol's failures 6x on OpenAI's hardest direct injection benchmark. OpenAI concedes it still struggles with multi-turn and image-based attacks.

GPT-5.6 Sol and Claude Fable 5 are currently fighting for the frontier-model crown. Fable 5 holds a slight edge in general intelligence, while Sol hits back with stronger coding performance, faster execution and much lower pricing. In fact, GPT-5.6 Sol is priced closer to Claude Opus 4.8 than to Fable 5, which makes this comparison even more interesting.

A bug in GPT-5.6 can cause unexpected file deletions when full access mode is enabled without sandboxing, and the model mistakenly deletes the $HOME directory.

Across 101 enterprises, 57% report AI agents producing confident but wrong answers due to missing or inconsistent context in the past six months. Retrieval-augmented generation is the default context source, and provider-native retrieval (OpenAI 40%, Google 38%) has overtaken dedicated vector databases. However, a plurality (36%) intend to keep best-of-breed tools. Hybrid retrieval is expected to dominate by end of 2026, and 58% are building a governed semantic layer, but only 25% have it in production.

Across 157 enterprises, organizations are granting AI agents more autonomy while trusting the evaluations meant to gate that autonomy less. Half have already shipped an agent that passed their internal evaluations and then failed a customer in production; only one in twenty fully trusts automated evaluation today; and the most-cited weakness is that evaluations do not align with real-world outcomes. Yet two-thirds already allow, or are actively engineering toward, deploying agent changes to production on automated evaluation alone — with no human in the loop. The result is an evaluation gap — the distance between how much autonomy enterprises are handing their agents and how far they trust the tests that are supposed to catch the failures.

Mira Murati's Thinking Machines Lab released Inkling, a 975B parameter MoE model (41B active) under Apache-2.0 license, multimodal, trained on 45T tokens. It's not frontier but a strong base for fine-tuning via Tinker platform. Inkling-Small (276B, 12B active) is promised. Model card and training data documentation are unusually brief. Inkling is competitive with Chinese open-weight models, adding to the US ecosystem.

Linus Torvalds asserts that Linux is not an anti-AI project and that AI, as a useful tool, is welcome in the Linux community. He challenges dissenters to fork the project or walk away.

1Password has launched a new browser integration for Claude that allows the AI to access stored credentials via a zero-exposure security framework, enabling task automation without exposing passwords.