Researchers demonstrate the first field validation of multi-robot coordination using only an external CCTV camera network and edge computing in a real warehouse. The system operates in image space on an uncalibrated, pixel-wise topological camera graph, using a hierarchical planner to assign camera sequences and plan motions. A prioritised-then-joint strategy treats overlapping camera regions as shared resources to prevent collisions and deadlocks. Tested with 4 robots, 30 cameras across 6 aisles.

Visuomotor policies struggle under distribution shifts due to poor action coordinate understanding. AxisGuide renders robot base axes into RGB images, providing explicit cues that significantly improve generalization and task success.

Autonomous cyber-physical systems (CPS) lie at the intersection of Model-Based Systems Engineering (MBSE) and data-driven ML/AI, but no integrated methodology spans both. IDDMBSE extends the MBSE V-process with a data-driven loop at every step, using SysML and a hybrid architecture. The open-source tool chain includes PERFECT, TRADES-X, and VERITAS. Demonstrated on a ground robot across development lifecycle tasks. Being reformulated on SysML v2/KerML for tighter integration.

SCOUT is an online semantic exploration framework that couples active traversal with probabilistic scene graph construction, enabling robots to progressively understand environments. It uses an uncertainty-guided planner balancing semantic certainty gain, geometric coverage, and travel cost for long-term autonomy.

This paper presents a model-based framework to control a 7-DoF manipulator performing non-prehensile ball juggling with a tool. A two-stage optimal control method computes feasible motion patterns, and offline trajectories enable real-time error correction. The approach is validated in simulation and on a Franka Emika Panda robot.

This paper presents a simple framework that settles the complexity of Multi-Agent Path Finding (MAPF) on trees across standard objectives—distance, makespan, and flowtime—for both labeled and colored variants. It proves NP-hardness on trees for both labeled and 2-colored MAPF under all three objectives, resolves the classical Pebble Motion problem on trees, and gives the first hardness result for colored Pebble Motion on any graph class. All results are established through the hardness of Stack Rearrangement.

This paper investigates key factors affecting transfer from everyday internet videos to robot manipulation policies. Using a new dataset of 532 human videos with 28 hours of high-quality hand labels, the authors find that hand pose quality matters, but the inherent motion gap hinders transfer unless networks specialize to each embodiment. Their cotraining recipe yields a 29.7% absolute success rate gain in low-robot-data regimes across six tasks.

Robot manipulation of physical puzzles is important for assembly/disassembly tasks. PhyRoGen leverages procedural content generation to automatically create synthetic datasets of manipulation puzzles, generating 24 puzzles with interlocking dependencies. All puzzles were solved in 1-300 seconds using sampling-based planning and manipulated by a KUKA robot in simulation.

This position paper argues that generalist robot intelligence requires more than scaling VLA models; it needs mechanisms to convert unstructured behavioral data into robot supervision. The authors propose four missing interfaces: data, embodiment, world-model, and reward, and advocate for learning from the broader physical world.

A new framework called Correlation-Optimized Fusion (COF) combines five uncertainty sources for deepfake detection, requiring only 42 seconds of optimization. It outperforms other methods under distribution shift but faces catastrophic failure across datasets, highlighting the need for domain-adaptive uncertainty quantification.

ViSAE uses neuroscience-inspired concept circuits to interpret and steer Vision Transformers, improving interpretability and worst-group accuracy.

DIRECT framework achieves pose-controllable object insertion by decomposing insertion conditions into appearance, geometry, and context guidance, outperforming prior methods in geometric controllability and visual quality.

This paper introduces DTG-FF, achieving state-of-the-art FF performance on multiple real-data benchmarks, but reveals three key limitations: FF lags behind BP on real data with a widening gap as class count increases; synthetic tasks overestimate FF scalability; and FF's memory advantage does not hold under fair hardware comparison.

WorldBench is a new multimodal reasoning benchmark emphasizing visual diversity. It builds a taxonomy of thousands of visual concepts, curates images from search engines and datasets, and designs questions that frontier MLLMs fail. Evaluation of 15 models shows the best achieves only 64.0% accuracy, highlighting the need for visual diversity in benchmarks.

AE-YOLO framework integrates lightweight bottleneck autoencoders, CBAM, and variance-maximizing regularization to address class imbalance, scale variation, and small defects in UAV insulator images. Achieves 95.10% [email protected], outperforming YOLO baselines by 5.0 points.

This paper proposes using U-net and DeepLabV3+ convolutional neural networks to segment foreground and background in cockpit images for mixed reality, achieving around 90% accuracy on images from a CAT793F truck simulator.

This paper introduces On-Policy Diffusion Language Model (OPDLM), which transforms autoregressive models into diffusion language models via on-policy distillation, addressing distribution shifts. OPDLM achieves strong performance with 15x to 7,000x fewer training tokens across various tasks, positioning DLM transformation as a form of ARLM post-training.

Web agents reading full DOM at every step suffer progressive context degradation. Signal-Driven Observation (SDO) proposes a sub-call that returns only task-relevant elements, re-invoked only when a lightweight signal detector fires. The paper calls for treating observation compression as a core architectural decision.

An analysis of 1,500 open-ended responses from 75 countries reveals that people's preferences for AI are highly diverse and contradictory. Only 'truthfulness' was requested by 49% of respondents, but it was defined in incompatible ways. Some capabilities, like human-likeness, are controversial. The study argues that current RLHF alignment methods flatten situated, contested signals into universal models, amounting to 'epistemic violence'.

The Piggyback Hypothesis proposes that chat-template tokens can piggyback finetuned behavior onto out-of-domain queries. Validated via prefix perturbations, leading to Token-Regularized Finetuning (TReFT) that mitigates emergent misalignment while preserving in-domain learning.

Formalizing complex reasoning from natural text is a central challenge in computational linguistics. Current Argument Mining techniques identify basic claims and premises but struggle with richer structures required by advanced schemas like the Carneades Argumentation Framework (CAF). We introduce CAF-Gen, an automated multi-agent framework that enriches shallow argument structures into CAF-compliant models using an iterative Creator-Reviewer pipeline. Experiments show the iterative feedback loop improves data quality and achieves strong alignment with original annotations.

This paper identifies two distinct processes of language model reasoning failures through token-level uncertainty signals: committed failure (early lock-in to incorrect paths) and persistent uncertainty (accumulating uncertainty). The framework is validated across 23 model-dataset configurations, with falsifiable predictions holding in 20 of 23 cases, and shows implications for self-consistency methods.

UnpredictaBench is a new benchmark that tests the ability of large language models (LLMs) to capture true underlying distributions. As LLMs are increasingly used as substitutes for other entities (e.g., for humans in economic simulations), many models tend to collapse towards a single plausible answer, failing to capture the unpredictability of real systems. The benchmark includes 448 problems involving canonical statistical distributions, distributions induced by stochastic programs, and natural-language scenarios describing random processes. It uses the KS@N metric, based on the Kolmogorov-Smirnov test, to quantify how well model outputs approximate target distributions. Experiments show a wide spread in performance, with KS@100 scores ranging from near 0 to over 20%, and no model exceeding 40%. Adding reasoning can slightly improve scores but does not solve the issue. UnpredictaBench demonstrates that even simple distributional simulation remains challenging, marking a necessary first step toward using LLMs as stand-ins for complex systems.

This study reveals significant gaps between synthetic and human data in evaluating LLM personalization, showing models struggle in attribute extraction, selection, and response generation. Two lightweight interventions improve automated evaluation, but human-aligned quality judgments remain challenging.

Large language models trained on English data often fail to express world knowledge reliably in other languages, known as cross-lingual factual inconsistency. This paper introduces PolyFact, a large-scale parallel multilingual factual QA dataset with 100K Wikidata-grounded facts across 12 languages. Comparing continual pretraining, supervised fine-tuning, and GRPO-based reinforcement learning on Qwen-2.5-7B and OLMo-2-1124-7B, GRPO consistently outperforms other methods, improving cross-lingual consistency and generalization to unseen languages. Mechanistic analyses show GRPO reduces language specialization in MLP layers and attention heads, promoting shared representations. Code, models, and dataset are released.

A new model called GPLFR combines compression and prediction by decoding high-dimensional outputs from a low-dimensional latent state with a Gaussian process prior, suited for low-data regimes. First demonstrated on building a spatially resolved emulator of global climate models for rocky exoplanets.

Researchers propose PoLar, a training-free dynamic program that skips or loops pretrained layers in LLMs per input. It achieves better accuracy with fewer layers on math reasoning benchmarks, suggesting fixed-depth inference underutilizes LLM capacity.

This paper argues that modern generative models, even below AGI capability, create structural risks for knowledge and cultural production. It defines Human Temporal Learning (HTL) as path-dependent knowledge accumulation through sustained engagement. Generative outputs increasingly resemble HTL-intensive work in surface features, making verification of genuine human learning costly relative to benefits. When verification loses economic justification, evaluators reward outputs indiscriminately, forcing human learners to compete on price with near-zero-cost AI outputs. The authors formalize this 'value collapse' with a costly-inspection framework. Cross-domain evidence from academia, law, content platforms, and software security maps four stages of verification erosion. Alignment success is orthogonal: better-aligned models narrow observable human-AI gaps, making source verification harder and intensifying competitive pressure on HTL-intensive work.

Symbolic regression systematically searches the space of mathematical functions to discover models that capture underlying data relationships. Despite recent advances, a lack of uncertainty quantification (UQ) support limits its real-world adoption. This first-of-its-kind survey introduces essential UQ concepts and reviews existing literature on UQ in SR, organized into frequentist, Bayesian, and model selection directions. The authors emphasize that UQ in SR remains underexplored, motivating further research.

A new method, WAV v1, proposes multi-resolution residual routing for deep decoder-only Transformers. It augments each block with directional detail bases (phase and split) to capture attention-vs-MLP and early-vs-late sublayer dynamics, routed together with block summaries. On character-level language modeling, it shows depth-dependent benefits, outperforming baselines at 48 layers with minimal parameter overhead.