Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
arXiv:2610.10438v1 Announce Type: new Abstract: Embodied and assistive agents must do more than recognize objects: they must reason about where an object belongs given the layout of an environment an...
arXiv:2610.09027v1 Announce Type: cross Abstract: Modern language model systems operate autonomously over increasingly long contexts containing untrusted text and images. Can an adversarial input con...
arXiv:2610.10125v1 Announce Type: cross Abstract: Household robots need to continually navigate to different objects in the same environment, many of which are small and portable, such as tools and t...
arXiv:2412.00112v5 Announce Type: replace Abstract: Generating natural and expressive human motions from textual descriptions is challenging due to the complexity of coordinating full-body dynamics a...
arXiv:2602.20531v2 Announce Type: replace Abstract: App ratings are among the most significant indicators of the quality, usability, and overall user satisfaction of mobile applications. However, exi...
arXiv:2604.01569v2 Announce Type: replace Abstract: Video multimodal large language models achieve strong results on existing benchmarks, but answer accuracy alone does not establish whether they can...
arXiv:2604.19710v2 Announce Type: replace Abstract: Vision-Language-Action (VLA) models bring rich world knowledge and reasoning capabilities to autonomous driving. However, most driving VLAs are tra...
arXiv:2610.10388v1 Announce Type: cross Abstract: Recent advances in multimodal foundation models have made them capable generalist physical agents for a range of manipulation tasks. However, success...
arXiv:2610.04432v2 Announce Type: replace Abstract: Building interactive simulators from real-world observations is a promising way to scale embodied data, but current pipelines still rely heavily on...
The paper introduces DAC (Divide and Cooperate), a role‑decomposed training framework that separates evidence acquisition (searcher) and response generation (generator) in agentic search systems. DAC uses task‑specific external verification signals to provide cross‑verification rewards, allowing the generator to abstain when evidence is insufficient and to assign credit appropriately to each role. Experiments on seven QA benchmarks and two model backbones show that DAC consistently outperforms single‑agent and multi‑agent baselines, with gains attributed to coordinated improvements in both search and generation.
The paper introduces random walks on graphs as a controllable sandbox for studying parallel sampling strategies in masked diffusion models (MDMs). It shows that common parallel unmasking methods, such as lowest‑entropy scoring, are not universally superior to random sampling and that performance depends on the graph’s conditional dependence structure. The authors also propose training‑free bisection samplers that are provably exact for random walks and demonstrate their effectiveness on both graph‑walk tasks and pretrained MDMs for language generation and reasoning benchmarks.
The paper introduces a kernel‑based operator learning method that preserves key physical properties—such as incompressibility, periodicity, and turbulence—of the incompressible Navier–Stokes equations. By mapping input functions to expansion coefficients in a property‑preserving kernel basis, the method guarantees that predicted velocity fields analytically maintain these properties. The authors provide theoretical convergence guarantees, develop efficient computational techniques for large‑scale training, and demonstrate significant accuracy and speed improvements over neural operators on 2D and 3D flow benchmarks.
CADReasoner is a model that iteratively refines its CAD reconstruction by comparing the input shape with the predicted mesh, outputting a runnable CadQuery Python program. It fuses multi‑view renders and point clouds and uses a scan‑simulation protocol to bridge realism gaps during training and evaluation. The approach achieves state‑of‑the‑art performance on DeepCAD, Fusion 360, and MCB benchmarks for both clean and scan‑sim tracks.
The study investigates how to adapt automatic speech recognition (ASR) models to child speech while preserving adult performance, focusing on Arabic and English data. It compares full fine‑tuning, LoRA, and post‑hoc weight‑space merging across encoder–decoder, encoder–CTC, and AudioLLM‑based ASR systems, using child and adult benchmarks. Results indicate that child adaptation is essential, especially for non‑native speakers, but direct adaptation can degrade adult accuracy; bilingual adaptation and weight‑space merging techniques help balance the trade‑off, with encoder–decoder fine‑tuning still yielding the best raw WER for bilingual models.
GeoPrior-Mamba is a new multi-directional Mamba framework that reconstructs fine‑resolution XCO2 fields from sparse satellite data by incorporating structured process priors derived from a language model. These priors encode biospheric uptake, ecosystem respiration, and anthropogenic emissions into deterministic tables, which are spatially instantiated with geographic, ecological, emission, and seasonal information and injected into the model via a lightweight knowledge adapter. On OCO‑2 data from 2018‑2020, the method achieves an RMSE of 0.81 ppm and an R² of 0.93, outperforming CAMS background interpolation and Trans‑XCO2, and shows faster convergence and consistency with ground‑based TCCON measurements.
The paper addresses the challenge of detecting plagiarism in music generated by AI, noting that traditional binary notions of plagiarism are insufficient due to nuanced musical elements. It evaluates whether existing music version identification models, originally designed for human covers, can transfer to human-to-AI plagiarism detection. The authors introduce COPYCAT, a large benchmark of 350,654 pairs derived from real plagiarism cases and augmented with generative re‑synthesis and DSP obfuscations, and demonstrate that a supervised embedding‑shift approach significantly improves detection performance over simple distance thresholding.
The paper introduces CanonQ, a unified quantization‑aware training framework that separates source canonicalization from task‑aware adaptation to enable extreme low‑bit LLM compression. By applying fixed rotations and energy normalization, CanonQ maps heterogeneous tensors to canonical coordinates, allowing frozen Gaussian‑reference codebooks to be reused across layers and models. Joint training then adapts the network to the coupled errors of weight, activation, and KV cache quantization, achieving significant performance gains—up to 14.28× lower perplexity and 57.9% higher zero‑shot accuracy on LLaMA models, and substantial improvements on instruction‑tuned MobileLLM‑Pro‑1B. "whyItMatters":"The framework demonstrates that carefully designed canonicalization and joint training can dramatically improve the efficiency and accuracy of low‑bit quantized large language models, offering a practical path to deploy powerful LLMs with reduced computational and memory footprints."
arXiv:2610.09737v1 Announce Type: cross Abstract: When does distribution alignment help a frozen foundation-model embedding generalize across acoustic domains? For cross-domain mosquito-species class...
arXiv:2610.09802v1 Announce Type: cross Abstract: Concept-based vision models represent images through an intermediate layer of human-inspectable concepts, so what a model relies on can be traced to...
arXiv:2610.10409v1 Announce Type: cross Abstract: General-purpose agents increasingly write code, use tools, and complete complex digital tasks, raising the question of how far these capabilities car...