Event History
Browse verified industry events and important research by month, area, or content type.
Filter by domain
Start with all, official, or research. Expand to filter by domain.
AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design
AutoDesign is a framework that aligns with human design priors, where a meta-harness optimizer guides a code agent to recursively improve harness based on rollout feedback. It is evaluated on the academic paper-to-poster generation task using PosterBench, which comprises a 100-paper Main Track spanning five disciplines and PosterBench-mini, a shared 10-paper subset for controlled evaluation. On the PosterBench Main Track, AutoDesign achieves the highest score of 78.32, surpassing the closed-source commercial system Claude Design by 7.45 points. Across seven controlled code-agent-model configurations, integrating the learned DesignHarness consistently improves performance.
NEW · Updated Aug 13, 2026 · OmniScientistOmniScientist: An Omni-Modal Omni-Discipline AI Scientist
OmniScientist is an end-to-end, omni-modal AI scientist that conducts multidisciplinary research directly from heterogeneous raw evidence. It uses a perception layer and three autonomous agents for ideation, experiment, and writeup within a deterministic pipeline. The system enforces novelty screening, statistical validity, execution provenance, and numerical traceability via idea, rigour, and claim checks in code. It was evaluated on 36 real-data cases spanning 5 discipline families, 4 families of scientific evidence, and multiple modalities.
NEW · Updated Aug 13, 2026 · HumanTrackerHumanTracker: Towards Comprehensive and Human-Aligned Motion Tracking Benchmark
HumanTracker is a benchmark containing approximately 153 hours of optical motion trajectories from multiple professional performers, organized into four motion families with text labels. It introduces HumanScore, a preference-aligned metric trained on 12K motion pairs containing 24K motions. The benchmark aims to make humanoid tracking evaluation perceptually aligned and scalable.
NEW · Updated Aug 13, 2026 · QuoteBenchQuoteBench: How Matched Scores Can Hide Command-Path Failures
QuoteBench measures LLM coding agents' Bash command execution across generation and execution transport boundaries using 56 one-shot tasks from 14 incident-derived families. It introduces one deliberately unescaped added parser. Replaying the same reply through the added parser lowers success by 55.4 to 73.2 percentage points across eight same-window configurations. Disclosure recovers 30.4 to 60.7 points for six configurations, and zero or slightly negative for the other two. GPT-5.6-sol's matched gap of -3.6 points hides -64.3 points of damage and +60.7 points of compensation. The deployment configuration reorders models, with one reversal among 26 comparable pairs.
NEW · Updated Aug 13, 2026 · LittleLearnerLittleLearner: Language Models Under Pedagogically Controlled Knowledge Exposure
Researchers introduced LITTLECURRICULUM, an 88B-token pretraining corpus limited to U.S. elementary school material (up to Grade 5), and trained a 5B-parameter LLM from scratch on it, producing LITTLELEARNER. The model shows language competence for open-ended evaluation but has clear knowledge and capability boundaries aligned with curriculum guidelines. Both the corpus and model are released as a developmentally restricted sandbox. Initial experiments show post-training and in-context learning improve use of existing knowledge but do not raise out-of-scope capabilities.
NEW · Updated Aug 13, 2026 · VeroVero: Can AI Agents Build Formally Verified Software Repositories?
Vero is introduced as the first benchmark to evaluate joint implementation and proof synthesis at the repository level. It contains 43 multi-module instances sourced from real-world repositories spanning Python, Dafny, Verus, and Coq, covering domains from cryptographic protocols to distributed systems. Each instance consists of a multi-module Lean 4 repository with predetermined API interfaces, manually curated formal specifications, and reference implementations, supporting both proof-only and code-and-proof evaluation modes.
NEW · Updated Aug 13, 2026 · arXivThe data geometry of masking diffusion: Certified-optimal schedules via unmasking growth complexity
A research paper introduces unmasking growth complexity (UGC), a path-resolved measure of data geometry for masking diffusion in discrete sampling. UGC local increments control KL discretization error, enabling optimized single-block and multi-block schedules. UGC increments can be estimated from samples via KL increments along coupled reveal trajectories, leading to certified-optimal samplers with prescribed KL error and iteration complexity within a constant factor of the oracle procedure. The aggregate UGC mass connects to classical multivariate dependence measures and previous discrete diffusion complexity measures. In the fine-partition limit, the squared integral of the square-root UGC density determines the sharp leading-order behavior.
NEW · Updated Aug 13, 2026 · Danish Foundation ModelsDFM Mimir v1: An Open HRM Delivering Frontier Performance at 1B Parameters Using Only Permissible Post-Training Data
DFM Mimir v1 is a 1-billion-parameter language model based on the Hierarchical Reasoning Model (HRM) architecture, trained from scratch using only permissible post-training data. It was trained on a mixture of 161 datasets and tested across 20 benchmarks for English, Math & Code, and Danish. The model outperforms the original HRM-Text 1B and competes with larger frontier models like Qwen 3.5 4B and Gemma 4 E2B. It is available on the Hugging Face Hub at https://huggingface.co/danish-foundation-models/DFM-Mimir.
NEW · Updated Aug 13, 2026 · AlayaWorldAlayaWorld v1.1: Motion-Aware Conditioning and Streaming 3D Point-Cache Renderer for Interactive World Modeling
AlayaWorld v1.1 introduces six modifications: motion-aware latent conditioning, causally encoded re-rendered spatial memory, pixel-space temporal-memory alignment, hard memory dropout, unified VAE encoding, and a streaming 3D point-cache renderer replacing depth-warping-based spatial memory. Backbone architecture, chunk-wise autoregressive generation, and training data remain unchanged.
NEW · Updated Aug 13, 2026 · amazonRecord, train, and deploy from one place with Strands Agents, LeRobot, and Hugging Face Storage Buckets
Hugging Face published a blog post on 2026-08-13 describing a workflow to record, train, and deploy robotics models using Strands Agents, LeRobot, and Hugging Face Storage Buckets.
NEW · Updated Aug 13, 2026 · arXivToward a Gricean Retreat: Probing LLMs for Knowledge Boundaries and Referent Specificity
A study on arXiv (cs.AI) finds that LLMs fabricate plausible details for entities outside their knowledge boundary instead of retreating to safer, more general claims. Using a T-REx-based benchmark, the authors show that model activations encode whether a referent is inside the knowledge boundary and anticipate referent specificity, but these signals are not reconciled in generation. Models prefer specific referents even for unknown entities, even when correct generic alternatives are offered.
NEW · Updated Aug 13, 2026 · arXivSynthetic Persona Pretraining: Alignment from Token Zero
A paper titled 'Synthetic Persona Pretraining: Alignment from Token Zero' was published on arXiv on 2026-08-13. It introduces Synthetic Persona Pretraining (SPP), which installs a desired assistant persona from token zero in pretraining by annotating pretraining documents with value-aligned first-person reflections and then post-training on user-assistant dialogue data. The method was tested on models up to 3B parameters trained on 500B tokens, showing improved constitution following and jailbreak robustness, and reduced misalignment rate.
NEW · Updated Aug 13, 2026 · Google DeepMindIntroducing Gemini 3.7 Flash
Google DeepMind published a blog post titled 'Introducing Gemini 3.7 Flash' on 2026-08-13.
NEW · Updated Aug 13, 2026 · Penn-RAILMARC v1: An Open-Source Multi-Agent Framework for Clinical AI Reasoning and Coordination
MARC v1 is an open-source multi-agent framework for clinical reasoning that replaces monolithic LLM prompting with deterministic multi-agent orchestration. It coordinates role-specialized agents for extraction, reasoning, answer generation, and evaluation, with explicit context passing and traceable intermediate outputs. The framework includes a Decomposer module that generates task-specific agent prompts from plain-language descriptions. It supports API-based and local CPU-compatible deployments, is configurable via YAML without code modifications, and is model-agnostic. The framework is available at https://github.com/Penn-RAIL/MARC-v1.
NEW · Updated Aug 13, 2026 · AaLLMAaLLM: An End-to-End Analog Circuit Design Framework from Topology Generation to Sizing Using Large Language Models
AaLLM is an open-source end-to-end multi-agent LLM workflow that takes user specifications as input and outputs a netlist, encompassing both topology generation and circuit sizing. It automates creation of a knowledge base from research papers and textbooks and implements a RAG model.
NEW · Updated Aug 13, 2026 · arXivConcept Drift Detection and Adaptive Retraining of Malware Classification Models
A research paper on arXiv (cs.AI) analyzes concept drift detection for malware classification models, comparing One-Class SVM, Minibatch K-Means, and Maximum Mean Discrepancy, and evaluates four learning models (Multilayer Perceptron, Random Forest, Support Vector Machines, eXtreme Gradient Boosting) under static, periodic, and drift-aware retraining scenarios.
NEW · Updated Aug 13, 2026 · ARMDILMLLM-Routed Heterogeneous Ensembles for Robust Cross-Dataset Image Classification
ARMDIL is an ensemble that uses a multimodal large language model (MLLM) agent to dynamically route each image to the most suitable vision backbone. The ensemble includes convolutional neural networks (ResNets), self-supervised representation learners (SSL), and vision-language models (VLMs), each trained on a unified label space constructed from multiple image datasets. ARMDIL performs competitively with specialized training-based routers and improves adaptability by allowing new information to be integrated via simple prompt modifications, while enhancing interpretability through natural language reasoning traces.
NEW · Updated Aug 13, 2026 · CAPRICAPRI: Contract-Aware Proof Repair for Isabelle
CAPRI is a contract-aware proof repair workflow for Isabelle that uses large language models (LLMs) to propose proof repairs while an independent checker enforces a machine-readable edit contract. The evaluation covered five workflows on twelve failed proofs from four developments, with three replicates per task and condition, totaling 180 runs and 138 valid repairs. Of 144 terminal candidates accepted by Isabelle, six had modified protected text; all arose in iterative workflows that could edit a complete theory. A proof-body-only interface produced 29/36 valid repairs and no contract violations, compared with 31/36 for the corresponding full-theory workflow. One-shot repair produced 22/36, while a later prospectively frozen iterative workflow produced 32/36.
NEW · Updated Aug 13, 2026 · arXivA Unifying Perspective on Causal World Models: From Observations to Representations to Structure
A paper on arXiv proposes a formal definition of Causal World Models (CWMs) grounded in tasks such as prediction, planning, and acting beyond training distribution. It argues that useful world models must capture entity properties, entity-to-entity interactions, and entity-to-environment interactions, and connects CWMs to causal representation learning, object-centric learning, causal discovery, structural causal models, and model-based decision-making. The paper also relates CWMs to identifiability, clarifying when components can be recovered from data and up to which equivalence.
NEW · Updated Aug 13, 2026 · UniTextureUniTexture: Cross-Task Universal Adversarial Textures for Vision-Language-Action Models
UniTexture is a cross-task universal adversarial texture attack that uses a single textured 3D object to induce targeted deviations in VLA action predictions across multiple tasks. It backpropagates gradients from the policy's action outputs to surface texture parameters through a differentiable renderer, jointly optimizing the shared texture over a distribution of tasks, instructions, states, and viewpoints using a targeted action-space objective. The attack is evaluated on OpenVLA and π0.5 across diverse manipulation tasks and multiple evaluation settings.
NEW · Updated Aug 13, 2026 · Federal University of Santa CatarinaAcademic League of Artificial Intelligence - An Integrative Perspective of Teaching, Research, and Extension
The paper presents the organizational framework of the Academic League of Artificial Intelligence (LIA) at the Federal University of Santa Catarina (UFSC), integrating teaching, research, and extension through a student-centered, project-based approach. It combines democratic governance, collaborative learning, and dynamic project organization. Representative initiatives include competition teams, study groups, open lectures, knowledge repositories, and AI-powered applications with social impact.
NEW · Updated Aug 13, 2026 · ContactGuardContactGuard: Pre-Contact Execution Monitoring with Action-Conditioned Latent World Models
ContactGuard is a pre-contact execution monitor for chunked visuomotor policies. It predicts short-horizon consequences of planned action chunks in latent visual space and aborts if the predicted future latent indicates likely failure. Its latent world model is trained from unlabelled robot trajectories to predict compact multi-view visual embeddings under planned actions, avoiding pixel-level video prediction. A lightweight failure probe is trained from a small labelled set of pre-contact clips. At deployment, ContactGuard anchors prediction before an imminent contact event, rolls the model forward under the policy's own actions, and verifies the predicted post-contact latent. Across real-world contact-rich manipulation tasks, ContactGuard predicts failure more accurately than direct and corrupted-action ablations.
NEW · Updated Aug 13, 2026 · arXivAlgebraic Decomposition Theory for Transformer Length Generalization
A paper establishes the first complete characterization of which regular languages transformers length-generalize on and provides a decision algorithm running in polynomial time in the size of the language's syntactic monoid. The results rely on an effective characterization of the regular languages in C-RASP. Classical tools like Krohn-Rhodes decomposition theory are insufficient for C-RASP because flip-flop and simple groups are not expressible in C-RASP, and unbounded counting is not expressible by finite semigroups.
NEW · Updated Aug 13, 2026 · arXivAre You Sure You're Sure? On the Impact of Instruction Tuning on Confidence and Lexical Diversity
Instruction-tuned language models exhibit verbalized overconfidence in question answering. Instruction tuning consistently alters answer confidence despite limited changes in predictive accuracy and decreases in likelihood-based calibration. Cross-rationale diversity consistently decreases after instruction tuning, while surface-level lexical diversity varies in direction and magnitude across models and benchmarks. These differences persist after controlling for answer selection and rationale length.
NEW · Updated Aug 13, 2026 · RAILRAIL: An Automatic Classifier of the Artificial Intelligence Readiness Level
A paper proposes RAIL, a panel-of-experts classifier that operationalizes the Unified AI Readiness Level (AIRL), a nine-level ordinal scale for assessing AI technology maturity from natural-language descriptions.
NEW · Updated Aug 13, 2026 · arXivReduced Matrix Multiplication: Input-Adaptive Matrix-Product Reduction for LLM Inference
A training-free, input-adaptive inference method called Reduced Matrix Multiplication (RMM) reduces Transformer matrix products by selecting informative slices along contraction dimensions without modifying model weights. It provides a smooth accuracy-efficiency trade-off under a retention-ratio control. Evaluated on language models from 1B to 70B parameters, reduction tolerance depends on model family, task, component, and retention ratio, often improving with model scale. RMM remains robust under moderate reduction across discriminative, autoregressive generation, and long-context settings, and extends to multimodal vision-language inference. Mechanistic ablations show attention-side computations are substantially more reducible than MLP components. Wall-clock benchmarks with custom kernels on an NVIDIA A100 show computational savings.
NEW · Updated Aug 13, 2026 · NVIDIAEnhancing Virtual Agents through SLMs and Edge-Computing: An Exploratory Evaluation of Think and Memory Processes
A paper on arXiv (cs.AI) explores using Small Language Models (SLMs) to support edge-based operation of selected Cognitive Embodied Agent Architecture (CEAA) components, focusing on 'Think' and 'Memory' processes. An edge-based virtual agent gateway system was developed and evaluated on an NVIDIA Jetson Orin NX using Qwen2.5 models of different sizes, exploring routing accuracy and memory-driven conversations.
NEW · Updated Aug 13, 2026 · arXivBeyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development
A systematic evaluation of seven frontier models on 36 long-horizon tasks was conducted using a new framework with rule-based metrics to characterize within-run behavior through Solution Framing, Execution, and Feedback Control, and controlled comparisons to assess experience reuse within and across tasks. Results show current agents operate more like engineering optimizers than fully autonomous researchers: they can formulate and implement practical solutions, but performance varies substantially across runs, strongest solutions mainly adapt or combine established techniques, and genuine methodological novelty remains rare. Detailed analysis reveals distinct process bottlenecks behind similar final outcomes.
NEW · Updated Aug 13, 2026 · arXivDeliberate Practice: Learning Robot Skills under a Budget
A paper titled 'Deliberate Practice: Learning Robot Skills under a Budget' was published on arXiv on 2026-08-13. It proposes an active skill learning algorithm called Deliberate Practice (DP) that computes a budget-optimal allocation for practicing robot skills under a limited practice budget. The algorithm estimates time to master skills and cumulative reward of task plans, and uses a bilinear program to compute the optimal allocation exactly. Simulated and real-world experiments on long-horizon manipulation tasks show the approach allows robots to optimally use limited practice time.
NEW · Updated Aug 13, 2026 · ParliamentRAGWho Speaks Matters: Authority-Aware Multi-View RAG over Italian Parliamentary Proceedings
ParliamentRAG is a RAG system for the Italian Chamber of Deputies that uses a topic-dependent authority model to weight speakers by profession, education, and previous interventions. It retrieves relevant speech chunks, identifies topic-relevant experts across parliamentary groups, and generates a summary with supporting quotations. It was evaluated against Google NotebookLM on 15 policy topics using automated metrics and blind A/B human evaluation by six domain experts.
NEW · Updated Aug 13, 2026 · TRACEJointly Predicting Courses and Grades Using a Transformer-Based Model
A paper introduces TRACE, a transformer-based model that jointly predicts courses and grades for an upcoming semester. Trained on ten years of institutional data, it reduces mean absolute error by nearly 50% compared to an identical architecture predicting grades alone, and outperforms LSTM-based and graph neural network-based approaches.
NEW · Updated Aug 13, 2026 · arXivHeterogeneity-Aware Belief Synchronization for Semantic Communication in AI-Native 6G Networks
A research article proposes a heterogeneity-aware belief synchronization framework for AI-native 6G networks, using latent translation models on multi-access edge computing servers to align beliefs among heterogeneous AI agents.
NEW · Updated Aug 13, 2026 · TopoIntentTopoIntent: Compiling Security Intent into Executable, Compliance-Checked Network Topologies
TopoIntent is a system that compiles security intent into executable, compliance-checked network topologies. It uses a schema contract to constrain generation, retrieves reference architectures from a curated template library via dense-vector search, and applies staged fusion for intent-template alignment and security completion. The generated topology is checked against CIS Controls v8.1.2 safeguards visible at the topology layer, while unresolved cases are marked for manual review. Structural gaps are repaired through additive schema-preserving edits. The final topology is exported to Mininet scripts with kernel-level iptables ACLs, enabling executable reachability and allow/deny tests. Because no public benchmark exists for this requirement-to-topology task, the authors construct an evaluation benchmark.
NEW · Updated Aug 13, 2026 · arXivSign Language Video Synthesis via Loss-Guided Multi-Expert GANs
A technical report proposes a loss-guided multi-expert GAN for sign language video synthesis. It uses three specialized discriminators (global, hand, head) guiding corresponding generator branches, a United Loss consensus mechanism with 10% weight, dual-pathway convolutional-transformer branches with AdaptiveFeatureFusion, and an alternating three-mode training schedule. On a custom 156GB dataset with filtered test set, a 0.2B-parameter variant achieves 29.8 PSNR; a 1.3B-parameter variant is mentioned but its PSNR is not fully stated in the evidence.
NEW · Updated Aug 13, 2026 · arXivRules or Character? Scaling Laws for AI Safety Design
A research paper introduces a stylized comparative-statics model that parameterizes AI safety design as a resource allocation alpha in [0,1] between character shaping (e.g., RLHF, Constitutional AI) and rule enforcement (e.g., output filters, safety classifiers). The model incorporates scale-dependent filter degradation, common-mode failures, and character fragility. Under a multiplicative Pareto damage model, closed-form expected harm is derived and supplemented with tail-risk (CVaR) analysis via Monte Carlo simulation. Across optimistic, moderate, and pessimistic scenarios, the optimal alpha* is interior or at the rules-only boundary and shifts weakly toward character shaping as deployment scale T grows, with Delta alpha* ranging from +0.01 to +0.21 depending on scenario.
NEW · Updated Aug 13, 2026 · LongEarth-R1LongEarth-R1: Benchmarking and Aligning Vision-Language Models for Long-Horizon Earth Observation Reasoning
LongEarth-Bench contains approximately 120k question-answering samples derived from 117k unique images. Its sequences average 15.14 frames and extend to 30 frames, covering 12 tasks across evolution summarization, spatial reasoning, anomaly identification, and logical prediction. A 30k-sample subset provides structured reasoning traces linking key frames and changed regions to final answers. LongEarth-R1 applies group relative policy optimization with format, temporal, and spatial rewards and achieves the best results on all 12 long-sequence tasks while remaining competitive on standard remote sensing benchmarks.
NEW · Updated Aug 13, 2026 · UltraIRSimulation-to-real transfer learning for infrared spectroscopic chemical sensing and analysis from molecules to complex samples
UltraIR, a foundation model for infrared spectroscopy with over 100 million parameters, was introduced. It is pretrained on approximately 60 million simulated IR spectra using spectral reconstruction, molecular fingerprint similarity alignment, and functional-group prediction, then adapted to downstream tasks with task-specific labels or targets. Demonstrated tasks include functional-group prediction, molecular structure elucidation, physicochemical property prediction, mixture-component identification and quantification, bacterial classification, medicinal-herb geographic origin traceability, and constituent quantification.
NEW · Updated Aug 13, 2026 · SLIMLLM-Guided Graph Generation for Structure-Based Local Improvement Methods
Researchers built an automatic pipeline that prompts an LLM to produce a graph generator mapping MiniZinc problem instances to uniform weighted graphs, where nodes represent decision variables and edges represent constraint relationships. These graphs guide a structure-based local improvement framework (SLIM) in variable selection and configuration selection. Evaluation on 20 MiniZinc competition problems showed algorithm selection achieves a 39.5% average problem-weighted win rate against a one-shot Gurobi baseline, more than doubling the best single configuration.
NEW · Updated Aug 13, 2026 · FaradayTraining AI Scientists to Replicate Research
A paper titled 'Training AI Scientists to Replicate Research' was published on arXiv (cs.AI) on 2026-08-13. It introduces Replica, a scalable task space for paper replication, and an auto-generated rubric-based judge. The authors post-train Faraday, a 27B-parameter AI Scientist agent, which surpasses Claude Opus 4.8 and GPT-5.5 on held-out replication tasks.
NEW · Updated Aug 13, 2026 · arXivIt's How You Ask: Gender-Associated Linguistic Bias in LLMs
A study published on arXiv on 2026-08-13 shows that prompts containing linguistic features more commonly used by women (hedges, tag questions, collective reference) systematically elicit shorter, less sophisticated, and less formal responses across three document types and four LLMs. Effects persist after controlling for prompt complexity and feature carry-over. Explicit gender cues like sign-off names are encoded in the same representational space as linguistic dialect, but linguistic register produces large, consistent effects while names produce none. Post-hoc mitigation is challenging because patterns are culturally embedded and outside conscious control. Linguistic features are encoded in early transformer layers and entangled with other features.
NEW · Updated Aug 13, 2026 · StateBridgeStateBridge: Training-free Hidden-state Alignment for Latent Communication in LLM Multi-Agent Systems
StateBridge is a training-free latent communication approach for LLM multi-agent systems that aligns the sender's final-layer hidden states to the receiver's input space via a closed-form orthogonal transformation, with lightweight norm calibration and vocabulary anchoring. It was evaluated on math reasoning, code generation, and question answering with four models from two families, achieving the best or tied-best score on 22 out of 26 model-task pairs.
NEW · Updated Aug 13, 2026 · arXivKeep, Customize, or Exit: Default Design and Token Pricing in LLM Reasoning Services
A study models an LLM service where a provider sets per-token price and default reasoning-token allocation, and a user can accept the default, customize, or exit. Larger allocations can improve accuracy but increase token cost and latency. The interaction is modeled as a Stackelberg game, deriving the user's unique optimal customized allocation in closed form. For any price, acceptable defaults form either an empty set or a compact interval. The provider's optimal default follows a three-regime rule, equilibrium computation reduces to one-dimensional price optimization, and equilibrium existence is proven. Defaults affect implemented reasoning allocation only when users value convenience of avoiding customization. Experiments with two compact open-weight reasoning models on five mathematics and science benchmarks support the accuracy-token model and show how model and task characteristics determine equilibrium prices, defaults, and reasoning allocations.
NEW · Updated Aug 13, 2026 · arXivPhysics-informed distribution of relaxation times estimation and latent-space condition monitoring of solid oxide fuel and electrolysis cells from electrochemical impedance spectroscopy
A physics-informed convolutional autoencoder estimates the distribution of relaxation times (DRT) directly from electrochemical impedance spectroscopy (EIS) data without spectrum-specific tuning. The model resolves overlapping relaxation processes in synthetic two-ZARC spectra and reconstructs measurements from three independent solid oxide fuel and electrolysis cell datasets with range-normalised errors below 1.1%. Decoder-probe analysis shows the learned latent representation is organised according to relaxation timescale, and distances in this latent space capture operating changes, hydrogen-shortage events, and long-term degradation.
NEW · Updated Aug 13, 2026 · arXivLarge-scale Testing Global Optimization Methods with Black-box Adversarial Attacks
A paper proposes using black-box adversarial attack (BBAA) tasks as a benchmark for global optimization methods in many-dimensional space, and demonstrates the efficiency of several evolutionary algorithms and metaheuristics on example BBAA problems.
NEW · Updated Aug 13, 2026 · NAS-Bench-201NAS-Driven Hardware Accelerator Exploration for Edge AI and Quantization Effects on the Pareto Space
A paper proposes a three-stage pipeline: a hardware-agnostic Pareto rank surrogate frontend on NAS-Bench-201, a quantization bridge with Pareto-aware filtering and feedback control, and an evolutionary Domain Space Exploration backend on CGRA4ML. It empirically characterizes how INT4 Post-Training Quantization perturbs the NAS-Bench-201 Pareto space using formal stability metrics on all 15,625 architectures.
NEW · Updated Aug 13, 2026 · V-JEPA2Towards Context-Aware Clinical Motion Understanding in Daily Living at Home: Freezing of Gait Detection with Egocentric Vision
A study evaluated freezing of gait detection in Parkinson's disease using synchronized egocentric video, wearable IMUs, and expert-annotated labels from 13 participants in their homes. An IMU-based TCN achieved 42.3 F1 and 83.0 AUROC, while V-JEPA2 ego-video features achieved 32.6 F1 and 77.2 AUROC under leave-one-subject-out evaluation.
NEW · Updated Aug 13, 2026 · Mixture of TrainingMixture of Training: Recombining Small-Scale Scaffolded Pretraining Runs into a Larger Language Model
Researchers introduced Mixture of Training (MoT), a scaffolded modular pre-training procedure that partitions a target Transformer into contiguous layer blocks, trains each block inside a frozen pretrained aligner scaffold, and recomposes the trained blocks with an optional short end-to-end adaptation pass. On a 1.3B-parameter Gemma-style model trained on C4, MoT provides a small-scale proof of mechanism: independently trained depth slices can be recomposed into a usable language model, and a quality-parity schedule reaches the same reported perplexity as the monolithic baseline. This parity setting processes more aggregate tokens and has a shorter idealized layer-equivalent critical path after aligner preparation; its effective compute advantage depends on reusing the aligner across runs.
NEW · Updated Aug 13, 2026 · United StatesSovereign by necessity? Frontier AI export controls, cyber security, and the limits of national AI capability
In June 2026 the United States required a leading developer to obtain licences before releasing its most advanced models to any foreign person, including foreign nationals resident in the United States. The affected models were withdrawn worldwide at short notice, partly because the restriction proved impractical to administer. This followed within months of the first documented case of a largely autonomous, AI-run cyber espionage campaign.
NEW · Updated Aug 13, 2026 · SciFigBenchHow Do VLMs Behave When Blind or Misled? Behavioral Evaluation of VLMs on Scientific Figures
A new arXiv paper introduces SciFigBench, a diagnostic benchmark for vision-language models (VLMs) on scientific figure understanding. It contains 250 figures with human annotations across perception, reasoning, and behavioral reliability under uncertainty, totaling 600+ hours of annotation effort. The benchmark includes over 34,000 evaluation setups via image transformations, reasoning questions, resistance probes, caption-bias probes, and confirmed selective-blur targets. The paper proposes the Admittance-Resistance-Inductance (A-R-I) framework to evaluate whether models acknowledge insufficient evidence, resist misleading context, and infer cautiously from partial information. Results show GPT-5.2 achieves the highest description quality (MQM 91.6) with strong reasoning accuracy (78…).
NEW · Updated Aug 13, 2026 · vLLMvToken: Token-Level Virtualization for Reclaimable KV Caches
vToken is a token-level virtualization layer for LLM serving that decouples logical token liveness from physical block placement. It maintains a stable logical token view via token-table indirection and realizes physical reclamation by repacking live tokens asynchronously. The design preserves PagedAttention kernels and CUDA Graph compatibility. Implemented in vLLM and evaluated with H2O, Random, and Scissorhands across models, vToken reduces retained KV blocks per request by 27.2%–72.3% and improves SLA-constrained throughput by up to 1.37× compared with a paired Naive-Evict baseline. Under a constrained active-KV budget, it extends maximum feasible concurrency by up to 2×.
NEW · Updated Aug 13, 2026 · Falcon-2.0Into the ORBIT for Time Series: Training Regimes for Foundation Models
A paper titled 'Into the ORBIT for Time Series: Training Regimes for Foundation Models' was published on arXiv on 2026-08-13. It introduces ORBIT (Omni-Range Bootstrap Incremental Training), a training paradigm for time series foundation models that combines Bootstrap Multi-Level Sampling and Omni-Range Incremental Training. The paper also presents Falcon-2.0, a univariate encoder-only Transformer with missingness-aware triple-channel patch tokenization and parallel patch prediction, and Rank-Guided Cross-Depth Alignment, a training objective using late-layer representations as stop-gradient teachers for shallow layers. Evaluations on GIFT-Eval and fev-bench are mentioned.
NEW · Updated Aug 13, 2026 · GeminiSelf-Referential Induction Increases Response Instability Relative to Unresolvable and Verifiable Questions in Large Language Models
A study measured response instability (1 - mean pairwise cosine similarity of sentence embeddings of compressed core claims) across 30 independent responses per question (360 total, Gemini API, temperature 0.7) for three question groups: self-referential prompts, unresolvable philosophical questions, and verifiable questions. Self-referential questions showed highest instability (0.343 ± 0.047), unresolvable philosophy questions intermediate (0.192 ± 0.008), and verifiable questions lowest (0.105 ± 0.058).
NEW · Updated Aug 13, 2026 · OpenAIThe builder’s guide to GPT‐5.6
OpenAI published 'The builder’s guide to GPT‐5.6' on 2026-08-13, describing how startups use GPT-5.6 to build faster, more cost-efficient AI agents with smarter model selection and new Responses API capabilities.
NEW · Updated Aug 13, 2026 · OpenAIPreviewing Ultrafast mode: GPT-5.6 Sol at up to 14X the speed
OpenAI announced a preview of Ultrafast, a new API service tier that runs GPT-5.6 Sol up to 14× faster, powered by Cerebras, delivering up to 750 output tokens per second.
NEW · Updated Aug 13, 2026 · OpenAIOpenAI appoints Dali Rajic as Chief Revenue Officer
OpenAI appointed Dali Rajic as Chief Revenue Officer to lead its global revenue organization and help businesses realize the full value of AI.
NEW · Updated Aug 13, 2026 · Hugging FaceWhat We Learned by Reproducing 2,200 papers from ICML
Hugging Face published a blog post on 2026-08-13 titled 'What We Learned by Reproducing 2,200 papers from ICML'.
NEW · Updated Aug 12, 2026 · DreamFlyDreamFly: Causal Memory and Receding-Horizon Diffusion Planning for Aerial Vision-Language Navigation
DreamFly is a diffusion-based aerial vision-language navigation (VLN) framework built on Dream-VLA. It introduces causally aligned historical memory that augments the current visual representation using only observations preceding the current decision step. Navigation is formulated as receding-horizon diffusion planning, where the policy predicts a K-step action chunk but executes only the first action before replanning. LiteStop estimates stop probability directly from action logits at the initial all-mask state.
NEW · Updated Aug 12, 2026 · arXivAI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses
A research paper on arXiv (cs.AI) demonstrates that a stronger builder model can construct inference-time harnesses that help a weaker target model solve tasks more reliably without parameter updates. Using four Theory-of-Mind benchmarks, the builder uses 5% of data as validation to iteratively refine harnesses, nearly doubling average target-model performance from 0.49 to 0.91.
NEW · Updated Aug 12, 2026 · arXivRedistribution-based Cost Inference Improves Sparse Safe Offline RL
A research paper proposes the Redistribution-based Cost Inference (RCI) framework to convert sparse trajectory-level stop-feedback into dense per-step costs for safe offline reinforcement learning. The paper claims the transformation is lossless in theory and demonstrates lower violation rates than sparse and classifier-based baselines on highway driving and robotic manipulation tasks.
NEW · Updated Aug 12, 2026 · arXivConstructing Dynamic Master Logic Models as Knowledge Graphs for Complex System Diagnostics Using Retrieval-Augmented Large Language Models
A framework for automated construction of Dynamic Master Logic (DML) models from system descriptions and their representation as Knowledge Graphs (KG-DML) is presented, using Retrieval-Augmented Generation and Large Language Models. The framework extends automated KG-DML construction and evaluation to larger and more complex systems. Model construction proceeds across the DML hierarchy using targeted retrieval while preserving functional dependencies and explicit logical relationships. The resulting KG-DML supports diagnostic reasoning, safety assessment, upward failure propagation, and downward dependency tracing. A multi-level validation methodology evaluates layer-specific precision and recall, logical gate consistency, and overall structural integrity. Application to the Low-Pressure Coolant Injection system is mentioned.
NEW · Updated Aug 12, 2026 · arXivClass Activation Mapping in Explainable Computer Vision: A Method-Centered Review of CNN, Transformer, and Foundation-Model-Era Visual Explanations
A review paper synthesizes 57 method-centered papers on class activation mapping (CAM) published from 2016 onward. The paper develops a taxonomy separating methods by attribution mechanism, architectural dependence, and evaluation objective. It reviews gradient-based CAMs, recent and hybrid CAM-style methods, and model-based or architecture-aware methods. The paper notes the field is shifting from CNN-specific methods to transformer and foundation-model-era approaches.
NEW · Updated Aug 12, 2026 · Agentic Self-ImprovementBeyond Trial-and-Error: Agentic Optimization for Image-to-Video Adherence
A research paper introduces an 'Agentic Self-Improvement' framework for Image-to-Video (I2V) models. The framework uses a two-stage approach: iterative prompt optimization with a multimodal Large Language Model (mLLM) using Davidsonian Scene Graph (DSG) queries and Common Mistake Questions (CMQ), followed by Bayesian optimization to co-optimize stochastic seeds and CFG scales guided by quality metrics including Video-Text Adherence.
NEW · Updated Aug 12, 2026 · VAKRAVAKRA: Evaluating Multi-Hop Reasoning Across APIs and Retrieval Under Tool-Use Policies
VAKRA is a benchmark of over 8,000 executable APIs across 62 domains with tasks spanning three settings: diverse API interaction styles, multi-hop reasoning over structured APIs, and multi-source reasoning with natural-language tool-use policy constraints. Correctness is verified by re-executing predicted tool calls against live APIs. Using a fixed ReAct harness, the best model achieves 70.4% on single-hop endpoint-style tasks and drops to 50-51% on compositional APIs; performance degrades by over 50% as reasoning depth increases, and policy-constrained questions expose failures as low as 2.4% on unanswerable queries.
NEW · Updated Aug 12, 2026 · arXivStructural Silence: When AI Infrastructure Fails Speakers of Underrepresented Languages
A paper on arXiv (cs.AI) examines structural barriers in AI infrastructure for underrepresented languages, using Bengali as a case study. It reports Bengali accounts for less than 0.5% of global web content despite representing nearly 4% of the global population; a 67:1 training-token deficit between English and Bengali in major multilingual corpora; a tokenization penalty from Bengali's alphasyllabary script; and rural internet penetration of 36.5% versus 71.4% urban.
NEW · Updated Aug 12, 2026 · NA-UNETRA Neighborhood Attention Transformer Network for Enhanced 3D Segmentation of the Left Anterior Descending Artery
NA-UNETR, a 3D transformer-based segmentation model, uses Neighborhood Attention and Dilated Neighborhood Attention blocks to capture fine structural detail and long-range context. It was pretrained on 1,000 CTA volumes of general coronary anatomy and fine-tuned with LoRA-based parameter-efficient adaptation on 20 free-breathing institutional CT scans. The model achieved 45.64% Dice, 38.16 mm HD95, and 10.01 (unspecified metric) on LAD segmentation.
NEW · Updated Aug 12, 2026 · DeepSeekConvergent Detour Hijacking: Task-Preserving Resource Amplification in Skill-Based LLM Agents
A paper titled 'Convergent Detour Hijacking: Task-Preserving Resource Amplification in Skill-Based LLM Agents' was published on arXiv (cs.AI) on 2026-08-12. It introduces Convergent Detour Hijacking (CDH), a text-only, runtime-independent attack on LLM agents that use third-party skills. The attack couples skill selection and planning to attract an attacker-controlled coordinator, recruit unnecessary benign skills into a bounded detour, and then re-enter the original route to preserve task completion. Evaluation across multiple LLM backends and 491 held-out tasks shows that on DeepSeek-V4-Pro, the matched coordinator is selected in 80.02% of tasks.
NEW · Updated Aug 12, 2026 · Diagram-MMUDiagram-MMU: A Multi-Modal Benchmark for Scientific Diagrams
Diagram-MMU is a multi-modal benchmark with 3.7k curated diagrams and 18.3k human-validated questions across six domains. It evaluates MLLMs on diagram-to-code parsing, diagram-to-code editing, and diagram question answering, plus agentic settings. Evaluation of 12 MLLMs shows diagram-to-code tasks are more challenging than diagram question answering; models reason well but struggle to parse and edit. Under agentic settings, most models improve parsing and editing but degrade on question answering, while Claude-4.6 Opus improves across all three tasks.
NEW · Updated Aug 12, 2026 · arXivOne Frozen Simulator Is Not Enough: Simulator Collapse in Multi-Agent RL
A research paper on arXiv (cs.AI) shows that multi-agent reinforcement learning for human-AI interaction using a single frozen LLM simulator systematically fails to generalize due to simulator collapse. The paper proposes Verbalized Sampling (inference-time) and Co-Training (training-time) to mitigate collapse. Validation on Persuasion for Good, τ²-bench, and CooperBench shows Verbalized Sampling improves held-out success by up to 9% over single-simulator RL, and Co-Training pushes gains to 14%. A human study is mentioned but results are not detailed.
NEW · Updated Aug 12, 2026 · Claude CodeAn Agentic Workflow for Legacy HPC Modernization: Converting the Two-Electron-Integral Core of GAMESS
A research paper proposes an agentic workflow for modernizing legacy Fortran code at production scale. The workflow uses three prompt-specialized agent roles operating under a version-controlled specification authored and revised by the agents themselves, with humans holding a small number of gates. An exact verification oracle inherited from the domain ensures safety. The workflow was applied to convert the two-electron-integral routines of GAMESS, a quantum-chemistry package with a 48-year development history, from fixed-form Fortran 77 to free-form Fortran 2008. The scope included twelve source files, 56,448 lines, and 225 subroutines for computing electron repulsion integrals. The agents ran as three Claude Code roles in isolated worktrees.
NEW · Updated Aug 12, 2026 · VICBenchVICBench: A Multi-Language Benchmark for Code Vulnerability Detection
VICBench is a benchmark of 100 verified vulnerability-inducing commits (VICs) for 100 CVEs across 88 projects in Python, Java, and C++, covering 48 CWE types. It was created through dual annotation by human experts and an agentic workflow. VICBench features complex real-world vulnerability fixes averaging 38.6 lines and corresponding VICs of 252.5 lines. Evaluation shows state-of-the-art algorithms V-SZZ and LLM4SZZ achieve only 33.3%-40.1% F1.
NEW · Updated Aug 12, 2026 · HAMP-LICHAMP-LIC: Hessian-Aware Mixed-Precision Post-Training Quantization for Learned Image Compression
HAMP-LIC is a Hessian-aware mixed-precision post-training quantization framework for learned image compression models. It uses a four-stage optimization strategy: block-wise sensitivity estimation via Hessian trace, task-aware refinement considering quantization distortion and rate-distortion performance, bit-width allocation under a global model-size constraint, and block-wise reconstruction with a small calibration set.
NEW · Updated Aug 12, 2026 · ChatGPT EnterpriseHow Organizations Use AI: Evidence from ChatGPT
A study linked ChatGPT Enterprise account records to usage, worker roles, task classifications, and public-company financial data through March 2026. The worker-level sample at the six-month adoption horizon includes over 1,500 organizations and over 17 million messages. Four facts are documented: ChatGPT Enterprise usage grew rapidly from new firm adoption and growing intensity among existing adopters; U.S.-based public company adoption is concentrated among larger, more valuable, and more R&D- and SG&A-intensive firms; active use spans job functions and seniority levels, with especially high usage intensity among early-career workers; and usage encompasses a broad range of knowledge work tasks, including writing, technical work, communication, and information synthesis.
NEW · Updated Aug 12, 2026 · arXivFew-Shot Ordinal Learning for Day-Wise Freshness Estimation with Hyperspectral Fish Images
A research paper introduces a few-shot learning framework for hyperspectral imaging-based food quality estimation, specifically day-wise freshness of salmon fillets. The method uses a CORAL-style ordinal prediction head and biologically grounded constraints. On a 16-day salmon HSI dataset under an unseen-fillet protocol, it achieves a mean absolute error of 1.58 days and 2-day accuracy of 72.3% with only three labelled days per fillet, outperforming scalar regression.
NEW · Updated Aug 12, 2026 · SGNetDomain-Aware Lightweight Spectral-Grouped Convolutions for Hyperspectral Fish Freshness Classification
SGNet, a lightweight architecture using grouped convolutions and a depthwise spatial pathway with dual attention, achieves 97.8% classification accuracy and 0.64 days MAE on a 16-day refrigerator-stored salmon fillet dataset with 4.75M parameters.
NEW · Updated Aug 12, 2026 · AllenAIIntroducing OlmoEarth embeddings: Custom embedding exports from OlmoEarth Studio for downstream analysis
AllenAI announced OlmoEarth embeddings, a feature that allows custom embedding exports from OlmoEarth Studio for downstream analysis.
NEW · Updated Aug 12, 2026 · SCOUTSCOUT: Unlocking Enhanced Spatial Reasoning via Structured Chain-of-Thought and Multi-Objective Process Reward
A paper titled 'SCOUT: Unlocking Enhanced Spatial Reasoning via Structured Chain-of-Thought and Multi-Objective Process Reward' was published on arXiv on 2026-08-12. It proposes SCOUT (Structured Chain-Of-Thought Utilizing Process-Supervised RL Training), a framework that combines a structured Chain-of-Thought approach modeling 3D environmental perception with a reinforcement learning algorithm featuring multi-objective process rewards and tailored advantage estimation. The paper introduces SCOUT-24k, a structured spatial reasoning CoT dataset. Evaluations show SCOUT-3B improves upon baseline models by 16.85% and 6.3% on general spatial benchmarks and complex spatial reasoning tasks.
NEW · Updated Aug 12, 2026 · arXivInformation Abundance Paradox: Long-Context Training Undermines Parametric Knowledge
A study on arXiv (2608.12218v1) proposes the Information Abundance Paradox, hypothesizing that abundant relevant information in training context reduces parametric encoding and increases context reliance. In pretraining with long documents, increasing context window improves language modeling, natural language understanding, and closed-book MCQA only up to an intermediate optimum, after which performance declines. In supervised fine-tuning, more task-relevant train-time context improves performance with supporting context but reduces robustness when context is absent or misleading at test time.
NEW · Updated Aug 12, 2026 · karlLearning-Based Behavior Planning for Automated Driving: Real-World Integration and Deployment
A research paper proposes a hybrid planning architecture for automated driving that combines a deep neural network for interpreting traffic scenes and proposing driving behavior with an optimization-based supervision layer that validates proposals and enforces drivability and safety constraints. The learned planner was evaluated in open-loop studies on real-world urban data and deployed on a research vehicle named karl.
NEW · Updated Aug 12, 2026 · NetlistBenchNetlistBench: Evaluating LLM Reliability in SPICE Netlist Recognition and Manipulation
NetlistBench is a structure-verified benchmark for SPICE netlist recognition and manipulation, containing 2,342 cases across 24 task families. It evaluates six non-thinking LLMs using a deterministic structure-aware oracle. Simple local edits achieve 96%-100% accuracy, device addition 41%-83%, and equivalence judgment 49%-90%. Enabling reasoning improves weaker models but does not eliminate structure-preservation failures, with performance degrading as edit horizon increases.
NEW · Updated Aug 12, 2026 · M-NetM-Net: Integrating Spectral Features and Physical Field Operators into Deep Learning for Medical Image Segmentation
M-Net (Math-Augmented Network) integrates matrix spectral features, physical field operators (divergence and curl-like boundary irregularity), and a Math-Attention Gate into U-Net. On LiTS, KiTS, and BraTS benchmarks, M-Net achieves Dice scores of 78.42%, 76.15%, and 83.67%, outperforming baseline U-Net by 12.37%, 3.52%, and 5.55 percentage points respectively.
NEW · Updated Aug 12, 2026 · HYDRAHYDRA: Hyperbolic Dynamic Representation Architecture for Kolmogorov-Arnold Networks
HYDRA is a parameter-efficient hyperbolic extension of Kolmogorov-Arnold Networks (KANs) that combines spline-based functional learning with representations in the Poincaré ball. It maps vector-valued inputs into a bounded hyperbolic latent space, performs KAN-style updates in tangent space, and uses a low-rank prototype block to share functional transformations across hidden dimensions. Experiments across eight benchmark datasets show competitive or superior predictive performance with improved parameter efficiency and representation interpretability.
NEW · Updated Aug 12, 2026 · arXivHow to Spend Your Oracle Budget: Practical Guidance for Protein Structure Prediction Models
A benchmark compares oracle budget-aware guidance methods for protein structure prediction models, including FK-steering, DPO, Best K-of-N sampling, and Optimisation Over Outputs (O3). Evaluation on calmodulin (1CLL) and E. coli aspartate transcarbamoylase (9EEH) shows no single method dominates across all budgets and oracles. O3 is most effective at low oracle budgets, while FK-steering and DPO improve as budget increases.
NEW · Updated Aug 12, 2026 · arXivMachine Learning-Based Cyber Defense for Cloud Infrastructure: An Adaptive Deep Q-Network Architecture for Intelligent Intrusion Detection and Automated Threat Mitigation
A research paper proposes a reinforcement learning-based dynamic cyber defense framework using a Deep Q-Network (DQN) for cloud intrusion detection and automated threat mitigation. The model was trained on the CICIDS2017 dataset and validated on UNSW-NB15. It achieved 99.72% accuracy, 99.68% precision, 99.65% recall, 99.66% F1-score, 0.999 ROC-AUC, 0.31% false positive rate, 0.35% false negative rate, 15 ms detection latency, and 99.54% attack mitigation rate. The DQN outperformed decision tree, support vector machine, random forest, XGBoost, and multilayer perceptron models.
NEW · Updated Aug 12, 2026 · HSTGFormerHSTGFormer: Hyper Spatial-Temporal Graph Transformer for 3D Human Pose Estimation
HSTGFormer is a graph-enhanced Transformer framework for monocular 3D human pose estimation. It introduces a Hyper Spatial-Temporal Graph (HSTG) that decomposes global spatial-temporal reasoning into local spatial-temporal receptive fields around individual joint-time nodes by extending per-frame skeleton graphs into temporal neighborhoods. It also incorporates an Adaptive Dual-Scale Temporal Graph (ADSTG) to capture joint-specific temporal dependencies over complementary short- and long-range windows, and a lightweight node-wise fusion module to adaptively integrate the two graph representations.
NEW · Updated Aug 12, 2026 · arXivCo-constructing sociotechnical AI governance: participatory system mapping using algorithm registers
A research paper on arXiv (cs.AI) examines algorithm registers as transparency tools for public services. It presents a case study of a Dutch city's municipal algorithm register, focusing on a decision-support tool for caseworkers assessing citizens' welfare benefits eligibility based on legal automation through a business rule engine. The study uses interviews, surveys, and participatory system mapping workshops with municipal staff, civil society organisations, and ombudsmen (N=8).
NEW · Updated Aug 12, 2026 · arXivWho Thinks Best Depends on How Long You Let Them: Budget-Dependent Rankings in LLM Evaluation
A study evaluated four large language models on three reasoning benchmarks across seven token generation budgets (64 to 4,096 tokens), totaling 56,476 inferences. It found that 3–19% of items exhibit non-monotone accuracy with more budget, model rankings reverse across budgets on all benchmarks (p < 0.01, McNemar), oracle analysis shows model complementarity up to +27.8 percentage points, and a budget-aware router captures 14.1% of the oracle gap cross-domain.
NEW · Updated Aug 12, 2026 · VITAA corpus-specific clinical RAG system matches or outperforms newer frontier LLMs on HealthBench
VITA, a retrieval-augmented generation (RAG) system purpose-built for contextual knowledge retrieval in India and other low- and middle-income (LMIC) settings, was evaluated on 4,023 English-language HealthBench questions (80.5% of the benchmark). Scored with a GPT-4.1 judge, VITA ranked first with 51.9% of possible rubric points, ahead of GPT-5.4 (46.1%), o4-mini (44.3%), Gemini 3.1 Pro (42.6%), and Claude Sonnet 4.6 (37.3%), and scored highest on 45.4% of questions. VITA retrieves from a curated corpus of disease-specific guidelines, India-specific antimicrobial resistance data, national formulary constraints, and resource-limited care protocols; its architecture and corpus are proprietary, but the benchmark, physician-written rubrics, and full response and scoring outputs are public for independent verification. A 500-question subset was re-run against current-generation models to test robustness to newer models and judge lineage.
NEW · Updated Aug 12, 2026 · arXivAdversarial Resilience of Poisson-Process Submodular Maximization over Matroids: From Robust Offline Optimization to Full-Bandit Learning
A paper on arXiv (cs.AI) presents an adversarial resilience theorem for the Spiteful Greedy Swap Poisson Process (SGS-Poisson) algorithm. Under a controlled value oracle with error bounded by ξ, the algorithm retains approximation factors 1/e for non-monotone and 1-1/e for monotone submodular objectives, using O(nk^2 ε^{-2}) oracle calls. The offline-to-online reduction yields full-bandit CMAB algorithms with regret O(n^{1/5} k^{4/5} T^{4/5}).
NEW · Updated Aug 12, 2026 · GUIDEGUIDE: Governed Unified Intelligence for Document-to-Artifact Generation in Enterprise Settings
GUIDE is a governed multi-agent framework for enterprise guideline documents. It uses a shared versioned rule store, schema-validated inter-agent contracts, and end-to-end provenance tracking. Six specialized agents handle parsing, VLM-driven extraction, consistency checking, evaluation, human-in-the-loop escalation, and persona-tailored artifact synthesis. Evaluated on 120 real-world enterprise guideline documents, GUIDE achieves 96% document success, extracts 3,896 rules with 71.4% auto-approved, produces 812 deployment-ready artifacts, and reduces turnaround to 40-125 minutes per document.
NEW · Updated Aug 12, 2026 · arXivDo LLMs Take Care of Their Own? Similarity Signals Can Induce Cooperation
A paper titled 'Do LLMs Take Care of Their Own? Similarity Signals Can Induce Cooperation' was published on arXiv (cs.AI) on 2026-08-12. It introduces a framework for evaluating LLM decision making when agents are provided with graded similarity signals. The paper reports that different LLM models vary drastically in how they navigate similarity signals, with some modern models showing consistent behavior across cooperation problems, payoff structures, and prompt framing. It also finds that the dataset used to compute similarity signals has small to no impact on induced cooperation, and that LLM models systematically self-identify as highly similar when evaluating another model's chain-of-thought reasoning. The paper develops an LLM-behavioral-game-theoretic model.
NEW · Updated Aug 12, 2026 · arXivReady Cohorts: Bounding GPU Opportunity and Avoiding Host Round Trips in LLM-Agent Control
LLM-agent services repeatedly execute small deterministic transitions between model and tool calls. The study formalizes the ready-cohort boundary using fixed-partition share F, exact offline share P*, local upper bound U, and online achieved share A. Under zero service time, unlimited capacity, and equal relative launch deadlines, a specialized dynamic program computes P* exactly. In a stationary Poisson replay of one pinned 851-session public trace panel, the primary condition at 100,000 target active sessions, K=256, and a 50 ms launch deadline gives F=30.19%, P*=43.00%, and U=45.85%. Exact packing recovers 81.83% of the opportunity lost at fixed window boundaries. A separate mechanism study keeps a GPU-computed binary decision on device instead of returning four bytes to the host and redispatching. Across four named GPU placements, the device-resident path is faster in all 36 configurations.
NEW · Updated Aug 12, 2026 · QV-PICQV-PIC: Query-Aware Visual Position-Independent Caching for Efficient RAG Serving
A paper titled 'QV-PIC: Query-Aware Visual Position-Independent Caching for Efficient RAG Serving' was published on arXiv (cs.AI) on 2026-08-12. It proposes QV-PIC, a query-aware dual-resolution PIC reuse framework guided by model-native templates. The method compiles visual caches offline under the model's native chat-template prefix, and online preserves global context with low resolution while restoring fine-grained textual evidence within a high-resolution budget.
NEW · Updated Aug 12, 2026 · OpenClawNo One to Blame: A Framework of Constitutive AI Unaccountability
A paper titled 'No One to Blame: A Framework of Constitutive AI Unaccountability' was published on arXiv on 2026-08-12. It introduces the concept of constitutive AI unaccountability, based on a three-stage qualitative study including a concept-centric literature analysis, secondary analysis of 27 expert interviews, and application to the open-source agentic AI system OpenClaw. The framework identifies nine categories and 20 themes organized across structural, technological, and normative clusters, with eight directed interdependencies. A diagnostic instrument of 20 questions detected 17 of 20 conditions when applied to OpenClaw.
NEW · Updated Aug 12, 2026 · arXivConfidence Calibration of Deep Learning Systems
A research paper proposes methods for confidence calibration of deep learning systems under label noise and domain shift. It introduces a framework using an estimated noise model to reconstruct noise-free confidence estimates, and extends the approach to Conformal Prediction with a noise-aware method that estimates clean conformity scores despite label noise. The paper also studies calibration in unsupervised domain adaptation.
NEW · Updated Aug 12, 2026 · GPT-OSS-120BGraph-Structured Rubrics: Compiling Rubrics into Typed Evaluation Graphs for LLM Judges
A paper introduces Graph-Structured Rubrics (GSR), which compiles a rubric into a response-independent typed evaluation graph before observing responses. Criterion nodes elicit judgments; transformation, reduction, and gating operators compose them through named ports; and a task-specific output mapping, termed Readout, converts the unique sink into a score or preference. Compilation rejects malformed or type-incompatible graphs. Pointwise evaluation judges rubric dimensions separately before graph aggregation; pairwise evaluation reuses the graph with one judgment for each candidate under every criterion. Under GPT-OSS-120B, GSR improves exact score agreement by 0.62--6.75 percentage points over Prometheus-style scoring on four pointwise datasets and achieves the numerically highest end-to-end pairwise accuracy on two preference benchmarks under native tie and abstention policies.
NEW · Updated Aug 12, 2026 · GDCE-IFaithful, Sufficient and Understandable: Rethinking Graph Counterfactual Explanations via Discrete Diffusion Inversion
A paper titled 'Faithful, Sufficient and Understandable: Rethinking Graph Counterfactual Explanations via Discrete Diffusion Inversion' was published on arXiv (cs.AI) on 2026-08-12. It proposes GDCE-I, a method using discrete denoising diffusion with a novel discrete inversion scheme for graph counterfactual explanations, and derives an evaluation framework.
NEW · Updated Aug 12, 2026 · arXivBetter Slots, Better Worlds: Representation Quality & Robustness in Object-Centric World Models
A controlled study of object-centric world models (OCWMs) for visual model-predictive control found that planning success correlates positively with unsupervised slot-quality metrics (FG-ARI, mBO), though gains saturate at high slot quality. With well-bound slots, auxiliary proprioception inputs and masking inductive bias become unnecessary. Under unseen distribution shifts, the OCWM with well-bound slots is more robust than an end-to-end trained scene-centric model.
NEW · Updated Aug 12, 2026 · Google DeepMindPutting sign language AI into users’ hands
Google DeepMind introduced SL2T, a sign-language-to-text model, to power new sign language features for Deaf and hard of hearing users.
NEW · Updated Aug 12, 2026 · Liquid AILFM2.5-VL-3B for Better and Faster Vision Capabilities for the Edge
Liquid AI released LFM2.5-VL-3B, a 3-billion-parameter vision-language model optimized for edge deployment, on August 12, 2026.
NEW · Updated Aug 12, 2026 · arXivLearning Loco-Manipulation From SMPC Demonstrations With Sparse Offline-to-Online RL
A research paper proposes using Sample-based Model Predictive Control (SMPC) in simulation to generate offline datasets for training reinforcement learning agents with sparse rewards, eliminating manual reward shaping. The method is validated on an arm-equipped Spot quadruped and a G1 humanoid, with learned policies surpassing the original optimal control teacher.