Research

As AI makes code generation cheap, costs shift from generation to ownership. To avoid technical debt, coding agents need an open-source intelligence layer that helps them reuse trusted components before generating new code.

This article contrasts learning from curated datasets with learning from raw experience. It shows that SGD and its variants absorb noise in online data streams, failing to learn only predictable components. The IDBD algorithm, however, can selectively assign credit and learn only useful associations. Extensions to neural networks (NetworkIDBD) demonstrate similar advantages on the NoisyMNIST task. The authors argue that better credit assignment algorithms are needed for online continual learning.

Mistral AI introduced Robostral Navigate, an 8B embodied navigation model. It moves robots from a plain-language instruction using only a single RGB camera, with no LiDAR or depth sensors. The model reaches 76.6% success on R2R-CE validation unseen through a pointing method, prefix-caching training, and CISPO online reinforcement learning.

Anthropic's new research reveals a hidden 'J-space' inside LLMs where words influence reasoning without appearing in output, offering insight into model decision-making but also raising questions about anthropomorphization.

A free benchmarking tool to evaluate your engineering team's AI agent maturity in 5 minutes. Based on hundreds of discussions with engineering leaders, it uses a 1-5 scale covering from suggestions only to fully autonomous multi-hour workflows.

This article debunks common lies from AI data center proponents, such as claims of innovation and job creation. It argues that these projects primarily bring pollution, water strain, and few local jobs, and criticizes media and corporate think tanks for misleading communities. The author warns that broken regulations make it difficult to hold tech companies accountable.

RQSHC V64I is a native Windows image compression research tool that uses a proprietary RQI format. It supports PNG, PPM, BMP input and achieves ~33% size reduction with very high SSIM. The core is built with C++17 and x64 assembly with AVX2 optimizations. Free for non-commercial use.

Copyright law is a key obstacle for AI companies investing in Australia. Creators accuse AI firms of using their work without permission, while tech groups argue the law blocks investment. The government considers multiple reform options but has not decided.

This survey systematically reviews the state of the art in embodied Vision-and-Language Navigation (VLN), organizing methods along two orthogonal dimensions: action paradigms (hierarchical vs. monolithic) and model paradigms (discriminative vs. generative). The authors conduct a real-world evaluation on a physical robotic platform across ten diverse scenes, revealing a significant simulation-to-real gap: a monolithic RGB-only method achieves 61% success in simulation but drops to 22% in the real world, while a hierarchical framework attains 51% real-world success. Key challenges in perception, decision-making, and control are highlighted.

This paper proposes a human-efficient post-training pipeline that enables a small number of human operators to supervise multiple robots through specialized division of labor and automatic trajectory segmentation using VLAC-CUT. Validated on four real-world manipulation tasks, the final policies achieve 80%-95% success rates and improve task throughput by 1.7x-4.2x over the base model.

This paper proposes a risk-field enhanced closed-loop digital twin framework for safety validation of autonomous driving systems. The framework integrates physical data acquisition, virtual reconstruction, risk-aware scenario generation, and algorithm evaluation, using a driving risk field as a unified intermediate representation to identify high-risk scenarios and provide safety guidance for reinforcement learning policies. Experiments show the method improves targeted validation and interpretability, but its effectiveness is bounded by model fidelity and sim-to-real transfer.

OmniSCS proposes an innovative system for generating photorealistic safety-critical scenarios (SCS) with high physical fidelity, enabling closed-loop simulation testing. It consists of a Fully Editable Driving World Construction module and an SCS Synthesis module that preserve data fidelity during scene editing. Experiments on nuScenes, Waymo, and KITTI datasets show that OmniSCS outperforms state-of-the-art methods in edited scene fidelity and supports real-time (13Hz) closed-loop testing, providing a safer and more cost-effective solution for autonomous driving development.

UAV swarms have potential in SAR and environmental monitoring but face limitations in situational awareness, connectivity, and cybersecurity. This paper proposes LAUS, an LLM-centric agentic AI framework integrating perception, memory, reasoning, and action for adaptive swarm behavior. It reviews enabling technologies, analyzes threats like Priority Manipulation Attacks, and identifies open challenges including hallucination-resistant reasoning, onboard LLM deployment under SWaP constraints, and standardized security benchmarks.

Researchers propose SWIFT, a unified framework integrating small-world networks with traffic flow theory for trajectory prediction in autonomous driving. It introduces structural inductive biases via a Small-World Interaction Network and a Flow Regime Encoder, outperforming baselines on nuScenes, MoCAD, and NGSIM datasets, with improved generalization and robustness.

A new differentiable physics framework for robust trajectory optimization of reusable launch vehicles introduces a Differentiable Particle Tube Control (DPTC) scheme that integrates actuator saturation constraints. Monte Carlo simulations show improved robustness over conventional methods by proactively managing performance trade-offs.

Proposed DecisionPerceiver architecture projects dynamic agent features into a fixed-size latent space, regulating granularity with latent queries, improving scalability. Evaluated across three driving scenarios shows consistent gains and generalization.

A new framework called RoboNav-Arm enables robotic manipulators to safely navigate and avoid obstacles in cluttered environments using agentic AI. It combines real-time obstacle detection, semantic reporting, central coordination, and adaptive motion planning, tested in Gazebo simulations.

EgoSteer is a full-stack system that enables steerable dexterous manipulation by pre-training a VLA model on 9.6K hours of egocentric human videos and post-training on robots. It achieves robust execution of free-form instructions across 40+ tasks, with failure recovery and few-shot adaptation to long-horizon tasks like box folding at 75%+ success.

Real-image diffusion inversion faces a quality-cost trade-off. This paper reveals two mechanisms: element-wise compression asymmetry and trajectory binding, leading to Noise-Anchored Reverse Correction (NARC), a training-free method that outperforms baselines with drastically reduced storage.

A paper accepted at ECCV 2026 presents a new approach to wearable motion capture that works with any combination of consumer devices like smartphones and smartwatches, introducing the WHIP model and a comprehensive dataset spanning 50 activities, along with a systematic study of sensor complementarity.

This paper proposes a Generalized Deep Non-negative Matrix Factorization (G-DNMF) method for SAR automatic target recognition. It overcomes the error accumulation and local optima problems of layer-by-layer decomposition in existing DNMF methods by deriving globally optimal update rules using the Lagrangian multiplier method. Experiments on MSTAR and OpenSARship datasets show improved stability and recognition performance over existing DNMF algorithms.

Multi-Modal Knowledge Graphs (MMKGs) enrich entities with modalities like text and images, but entities with highly similar multi-modal features remain hard to distinguish. Temporal information can serve as an additional modality for disambiguation, yet existing approaches rarely treat time as a separate modality due to sparse temporal semantics and noise from multiple timestamps. This paper proposes Time Imprint, a framework that treats time as an entity-level modality and aligns temporal, textual, and visual representations via a three-view contrastive objective. It also designs a compact timestamp subset selection with attention pooling to balance specificity and robustness. Experiments on three MMKG benchmarks show state-of-the-art link prediction, with Hits@1 improvements up to 6.07% overall and 58% on the top-1% ambiguous samples.

This paper proposes a knowledge-constrained shape optimization framework that translates expert knowledge and user intent into quantifiable parameters for DFFD-based deformation operators. A Mixture-of-Experts Neural Operator (MoE-NO) improves drag prediction and trend consistency on heterogeneous datasets. Experiments show MoE-NO achieves 1.16% MAPE and 94.34% trend accuracy, with CFD-validated drag reductions of 4-10%.

A new AI system called ReflectWorld-MM enables assistants to continuously process and remember open-ended video streams by organizing memory around persistent entities rather than frames, achieving state-of-the-art results on six benchmarks.

RSLoRA is a training-free, gradient-free method for allocating LoRA ranks based on activation-space geometry. It introduces virtual representational probing to identify high-sensitivity layers, outperforming state-of-the-art allocators like AdaLoRA and GoRA.

WiCAT, a multi-subject model using self-supervised pretraining, outperforms single-session models and enables zero-shot behavior decoding on unseen subjects in widefield calcium imaging.

Researchers propose DUNE, a training-free framework that refines diffusion models by detecting and suppressing early-stage fluctuations in deep latents, reducing artifacts and hallucinations while improving fidelity across both U-Net and Transformer backbones.

This paper investigates the feasibility of training a reasoning language model in Japanese. By applying GRPO to a Japanese continually pretrained model based on Qwen-3-Swallow-8B, the authors find that reasoning-language control is achievable, yet performance at best matches English-reasoning baselines. On Japanese cultural benchmarks, the model performs worse, indicating that reasoning in Japanese does not automatically improve culturally relevant tasks.

This work adapts an open-source spoken language model (SLM) to the Singaporean Home Team domain using LoRA fine-tuning, a surrogate text-QA dataset, and a multi-task objective with CoBa reweighting. The resulting model, HT-Moonstone (5B), matches or outperforms SLMs 7x its size on most tasks and achieves top accent and gender recognition with less than 2% loss in original speech QA ability.

Researchers present FindMyText, an open-source Python package to efficiently check if a given text appears in part or full within a corpus. It uses a novel fingerprint chain mechanism to reliably detect near-verbatim copies, ideal for copyright verification. The system scales to large web-crawled datasets via distributed disk-based indexing, outperforming alternatives on ArXiv, Wikipedia, and web content.

A new study shows that post-training quantization can silently alter how large language models reason even when task accuracy is preserved. Using a six-category failure taxonomy, the researchers classified 30,000 chain-of-thought outputs and found that hollow convergence exhibits a size-dependent shift under NF4 quantization, while shortcut collapse and confidence snowballing undergo qualitative changes. Hollow convergence cannot be reliably detected from surface-level text features, posing a deployment risk.

This report studies on-device English-to-Traditional-Chinese subtitle translation for Taiwan under short inputs, short outputs, batch-size-one inference, low latency, and privacy constraints. The authors replace the original 151k-token vocabulary with a 64k-token subtitle-domain tokenizer, perform embedding calibration and fine-tuning, achieving a 59.2% tie-excluded win rate against Google Translate on a subset of OpenSubtitles2024, and a 1.63x speedup on Apple M2.

A new benchmark framework evaluates the faithfulness of LLM-generated clinical trial summaries across three stakeholder audiences. Using 200 stratified trials from ClinicalTrials.gov and a six-dimension annotation schema, the study identifies 'Unsupported Claims' as the dominant failure mode. A knowledge-graph-augmented retrieval system shows statistically significant improvements in faithfulness scores.

A language-model forecasting system for merger arbitrage, utilizing long-context reasoning over technical documents, outperforms market-implied probabilities and frontier LLMs on a dataset of over 400 large deals across 42 countries.

Bilibili releases Index-1.9B, a series of open small language models with competitive performance on benchmarks.

CLIR-Bench is a benchmark for evaluating models on question answering over irregular clinical time series. It is constructed from de-identified ICU records using a principled four-stage pipeline, comprising 6,600 QA instances covering 11 clinical variables, organized into four capability dimensions and 11 tasks. Experiments reveal that current generalist models struggle to retrieve and reason over sparse clinical evidence, highlighting the need for stronger irregular time-series reasoning methods.

Researchers introduce a reference-based membership inference method to detect whether large language models are distilled from other models. By comparing a student model's preference for outputs from different candidate teachers against an earlier checkpoint, the method identifies the most likely teacher with near-perfect accuracy, handling unknown distillation pipelines and open-world settings.

A new study reveals that coding agents need minimal context when editing code: the signal is only in the code being edited, natural-language summaries fail to answer behavioral questions, surrounding context (UML skeletons) performs no better than deleting it, and compressed context matches full files at one-third the tokens. Temperature-0 inference introduces a ~9% noise floor. The authors release their instrument including gold-validated environments, deterministic patches, and pre-registered hypotheses.

This paper introduces a hybrid search framework that combines Thompson sampling with parallel self-avoiding walks to efficiently allocate computational resources in the Low Autocorrelation Binary Sequences (LABS) problem. The method, modeled as a multi-armed bandit, dynamically prioritizes promising search space partitions, achieving new best-known results for 35 sequence lengths and a longest sequence with merit factor exceeding 8.0.

A new paper introduces MawForge, a system that enables practical local inference of Sparse Mixture-of-Experts (MoE) language models on memory-constrained unified-memory machines by storing the model on disk and materializing expert tensors on demand into a bounded cache. The system is effective as a measurement substrate but not as a cache-maximization policy.

A new study uses macroeconomic forecasting as a stress test to evaluate five model families (ARIMA, LSTM, NODE, PINN, UDE) across 23 countries with sparse annual data. Results show no model consistently performs well, but less-constrained models (ARIMA, NODE) consistently outperform more-constrained heuristic-prior models (PINN, UDE). The study finds that structural priors can act as misregularizers when they do not match the data-generating process, and identifies failure modes including prior misalignment, regime shifts, structural breaks, and optimization instability.

This paper systematically compares Turbo-Quant and SpectralQuant KV-cache compression methods using a statistical validation methodology that separates systematic codec differences from implementation variance. Key findings reveal that eigenbasis-based methods fail on heavy-tailed data due to covariance instability but excel in structured regimes, with the effective semantic dimension adapting to calibration budgets rather than true data rank.

This position paper argues that ground truth datasets in machine learning are not neutral objective measurements but are constructed through human and technical arrangements. It advocates for recognizing the contingent, context-dependent nature of these datasets and promoting 'situated reliability' to enhance transparency, accountability, and interdisciplinary work.

This paper proposes a novel two-level taxonomy for GNN-based knowledge graph technologies, covering construction, embedding, reasoning, and applications, and reviews various GNN models, discussing their strengths, limitations, and future directions.

The paper presents a continuous-time instantiation of Feedback-Coupled Memory Systems (FCMS) by defining the agent update operator via Mechanism-Based Intelligence (MBI) and the environment update operator via Coupled Memory Graph Process (CMGP). It achieves Lyapunov global dissipativity with a computable threshold that generalizes previous discrete FCMS and CMGP stability conditions, establishing memory dissipation exceeding feedback gain as a universal organizing principle. Numerical simulations confirm the threshold and a self-reinforcing coordination cascade when violated.

A learning-based graph edge sparsification method for efficient large-scale Euclidean TSP solving. By integrating geometric structure and combinatorial optimization, it adaptively generates sparse graphs, pruning up to 95% of edges on MATILDA dataset with solution gap under 1%, and demonstrates strong generalization on TSPLIB.

Recent latent reasoning methods like CODI and COCONUT lack interpretability because they maintain multiple superimposed traces. Researchers model these as trajectories in representation space and apply dynamical systems analysis, revealing that CODI behaves as a stable attractor while COCONUT behaves as an unstable expanding system. SIM-CoT supervision tightens both behaviors without changing underlying dynamics.

This paper proposes interpreting the MapReduce reduce operation as a partition function in statistical mechanics. Under local asymptotic normality (LAN), the confidence density emitted by a worker is a Gibbs–Boltzmann measure with inverse temperature equal to sample size. This leads to precision-weighted pooling and frequentist consistency as the zero-temperature limit.

A new study investigates how message format affects information fidelity in multi-hop LLM agent relays, finding that effects are tier-dependent. Under strong relays with faithful instructions, loss is minimal, while weak relays show large inter-format variability. Structured formats provide a faithful, error-localizing channel, not an error-correcting code.

This study introduces the Format Sensitivity Index (FSI) and Parseability Sensitivity Index (PSI) to measure how prompt wrappers affect LLM accuracy and answer parseability. Experiments on 140,000 generations show mean FSI varies by over 30x across models, largely explained by compliance failures. Parseability remains a strong predictor of accuracy even after controlling for task, model, and wrapper. Recommendations for robust benchmarking and structured-output deployments are provided.