arXiv AI

arXiv:2610.29281v1 Announce Type: cross Abstract: Reinforcement learning with verifiable rewards (RLVR), particularly through group-relative policy optimization (GRPO), has emerged as a central paradigm for the post-training of large language models (LLMs). While RLVR is highly effective at enhancing the reasoning capabilities of LLMs, it commonly induces \textit{distribution sharpening}, whereby the policy distribution becomes excessively peaked and rare-yet-correct trajectories are progressively suppressed, thus diminishing sample diversity. To this end, we promote flat-reward optima in RLVR to smooth the policy distribution. We theoretically link flat reward RLVR to reduced distribution sharpening, and empirically show improved sample diversity while maintaining single-attempt accuracy. Our method, called \texttt{Flat}-RLVR, searches for flat reward regions via sharpness-aware minimization during LLM post-training, combining this with GRPO variants to achieve better trade-offs between sampling efficiency and reasoning capacity.

arXiv:2610.29257v1 Announce Type: cross Abstract: As-built BIM reconstruction from point clouds demands semantic, volumetric, parametric, and editable building components, beyond 2D layouts or planar abstractions. Yet existing scan-to-BIM pipelines often rely on handcrafted rules, heuristic optimization, or manual correction, while learning-based methods fall short of recovering wall thickness, opening geometry, and component-level BIM structure. We introduce CRAB, a BIM-ready cuboid reconstruction framework that encodes walls, doors, windows, beams, and columns as compact architectural cuboids with IFC-compatible semantics and parameters. CRAB bridges point clouds and editable BIM by combining ProposalNet, which predicts coarse component cuboids, with AssemblyNet, a relation-aware Flow Matching model that refines geometry, predicts structural relationships, and enables automatic structural correction and human-guided editing. To address limited cuboid-level supervision, we extend Structured3D with component-wise cuboid annotations and construct cuboid-based procedural scenes. Experiments show CRAB surpasses floor plan and plane-based alternatives in component completeness, geometric accuracy, and generalization. For interior wall thickness estimation, CRAB achieves a mean error of 2.80 cm, with 79.50% of matched walls within 3 cm. The project page is available at https://vcc.tech/research/2026/CRAB.

arXiv:2610.29255v1 Announce Type: cross Abstract: Masked diffusion language models have emerged as an alternative to autoregressive models, enabling parallel token generation. However, standard masked diffusion language models use a shared noise schedule across positions without explicitly encoding which positions should be resolved first. Many tasks exhibit a natural priority among positions: document titles and section headings can guide the generation of body text, while Sudoku cells determined directly from the initial clues can help resolve more difficult cells. To incorporate this notion of priority (or hierarchy) among token positions, we introduce the Hierarchical Masked Diffusion Language Model (HMDLM), which augments each token position with a hierarchy level available as part of the training data, and requires zero or minimal additional annotations. It factors generation into two passes: the first predicts a hierarchy distribution over masked positions, the second predicts token distributions conditioned on it. Giving each hierarchy level its own noise schedule induces a soft, controllable preference for revealing higher-priority positions earlier while keeping decoding fully parallel. Our formulation is mathematically grounded, and we re-derive the forward and reverse processes and the NELBO for this factorization. Our evaluation shows that for the task of language generation on LM1B, HMDLM reduces perplexity by more than 1.5 points over Anchored Diffusion Language Model (ADLM), a recently proposed strong diffusion baseline; on WikiHow, we see a drop in perplexity of up to 4.4 pts on ADLM and up to 7 pts against MDLM. An LLM judge on WikiHow prefers HMDLM generations (the two variants combined) more than $2\times$ as often as the baseline. On the task of solving Sudoku puzzles requiring multi-step reasoning, it improves solution accuracy by up to $8.6\%$ absolute points over existing diffusion baselines.

arXiv:2610.29249v1 Announce Type: cross Abstract: Omni-modal assistants must remember not only what users say, but also what they show and convey. Existing long-term memory benchmarks are mostly text-and-image or report a single end-to-end score, so they cannot locate whether evidence is lost when it is represented, retrieved, delivered, or read. We introduce MOSAIC, a benchmark of 923 questions over persona-grounded histories of 11-24 sessions spanning text, image, video, and audio. MOSAIC constructs targeted spatial, temporal, and prosodic signals rather than decorative media, with the strongest causal validation for spatial information and selector-level validation for annotated prosody, and pairs them with a matched diagnostic protocol that freezes the question, reader, prompt, and budget while intervening on representation, retrieval, and delivery. Matched decomposition shows that representation loss is modality-dependent, residual reading dominates even under gold evidence, and abstention remains weak on these controls: strong answerable scores coexist with failures to refuse. MOSAIC thus turns omni-modal memory evaluation from a leaderboard into a diagnosis of write, retrieve, deliver, and read failures.

arXiv:2610.29225v1 Announce Type: cross Abstract: Current LLM safety benchmarks predominantly rely on binary safe or unsafe labels, lacking explanatory depth and suffering from low consistency across automated judges. To address this, we introduce DiagSafe, a diagnostic evaluation framework that moves beyond binary classification by employing prompt-level rubrics. Grounded in globally applicable AI safety policies, each rubric provides explicit criteria along two dimensions: safety, which targets concrete harms with specific exemptions to suppress false positives, and usefulness, which independently evaluates instruction fulfillment. To ensure a robust and realistic evaluation, we construct a bilingual (English-Chinese) benchmark grounded in a comprehensive 84-label safety taxonomy, consisting of two subsets: DiagBase, which covers single- and multi-turn prompts with direct harmful queries and template-based jailbreaks, and DiagAttack, which embeds harmful intents into complex instruction-following tasks to form challenging instruction attacks. Extensive experiments demonstrate that DiagSafe significantly enhances the reliability of LLM-as-a-judge evaluation, achieving 95.34% agreement with human annotations -- substantially higher than the 69%-87% achieved by the baselines -- and high inter-judge consistency (Cohen's $\kappa$ = 0.864). Furthermore, applying our framework reveals that leading LLMs differ markedly in their unsafe rates, and each model also exhibits unique vulnerabilities along different dimensions. In addition, we observe that models are particularly susceptible in Chinese and multi-turn contexts. Ultimately, DiagSafe provides a reliable and explainable tool for auditing and optimizing LLM alignment.

arXiv:2610.29213v1 Announce Type: cross Abstract: Satellite precipitation products face a persistent trade-off: real-time estimates are timely but less accurate, whereas the most reliable products are often released months after observation. This delay keeps high-quality precipitation information outside the decision window for flood warning and disaster response. Here we report Huayu-Global, a cross-constellation deep-learning system that combines three geostationary sensor streams from five satellites through a common retrieval framework and generates coherent quasi-global precipitation fields every thirty minutes. Across more than a million independent station records from two periods outside the training years, Huayu-Global improves event detection and precipitation-amount estimation relative to the real-time products evaluated and reaches delayed-product-level skill in key metrics. Huayu-Global therefore turns operational geostationary observations into a timely, high-skill precipitation layer for flood forecasting, water management and disaster response. Its cross-constellation design also provides a scalable foundation for incorporating future geostationary missions into hydrological and weather-prediction systems.

arXiv:2610.29188v1 Announce Type: cross Abstract: Tables in real-world documents, including scientific papers, financial reports, and statistical summaries, frequently use merged cells, multi-level headers, and sectioned layouts. Yet the lightweight formats LLMs handle best (Markdown, CSV) cannot express these structures, while expressive formats (HTML, LaTeX) impose excessive token cost and parsing complexity. To enable systematic analysis of this trade-off, we introduce a taxonomy decomposing table structure into six atomic features (F1-F6) and corresponding complexity levels (L1-L6), and construct STB (Synthetic Table Benchmark), a controlled benchmark with 1,089 tables systematically distributed across complexity levels. Using this framework, we propose LTEM (LLM-friendly Table-Enhanced Markdown), a formally specified, minimal extension of Markdown that explicitly encodes spans and hierarchical headers. Experiments across five state-of-the-art LLMs show that LTEM achieves accuracy comparable to HTML, offers a favorable accuracy-token trade-off relative to HTML and LaTeX, and produces substantially fewer invalid responses than LaTeX; for table generation, LTEM is comparable to MultiMarkdown though HTML is best. We release our taxonomy, benchmark, and format specification to support future research on LLM-table interaction.

arXiv:2610.29183v1 Announce Type: cross Abstract: Executing vision-language models (VLMs) aboard low-Earth orbit satellites enables flexible processing of observation data without first downlinking it to the ground. On commercial off-the-shelf payloads, however, a request may span multiple execution windows: power and thermal constraints, fault-triggered reboots, and competing payload demands can interrupt inference before it completes. Each shutdown discards volatile activations and KV caches, forcing an unprotected request to repeat work after restart. We present OrbitWeaver, a framework that preserves VLM inference progress across onboard interruptions through coordinated checkpointing and recovery. Its phase-aware checkpointing selects recovery boundaries according to the state lifetimes and computational characteristics of vision encoding, language-model prefill, and decoding, balancing storage traffic against recomputation. Its dependency-aware dynamic recovery schedules restoration from partially completed states across the layer and token-chunk dimensions. By overlapping checkpoint and weight reads with recomputation, it reduces restart latency; the recovery planner computes an optimal execution order under its scheduling model within 100 ms. Experiments on satellite payload hardware using on-orbit telemetry traces show up to a 4.07-fold end-to-end speedup and a 3.56-fold reduction in SSD write volume over the evaluated baselines. A 72-hour deployment on two low-Earth orbit satellites demonstrates that OrbitWeaver retains inference progress and resumes execution across repeated bus-enforced shutdowns.

arXiv:2610.29178v1 Announce Type: cross Abstract: Before an AI agent takes a consequential action, such as paying, messaging or changing code, several monitors check it. Their verdicts are combined by aggregating scores or by counting approvals, whatever inputs the monitors share. Monitors that share a base model or a poisoned input, however, may fail together. We borrow the fail-prone systems of distributed computing: lists of monitor sets that may fail together. A rule is safe if it executes no bad action and live if it blocks no good one. We measured twelve small open-weights monitors on payment approval under two pre-registered protocols. Without pre-registration, we also measured three closed frontier models (Claude Opus 5.5, GPT-6 Astra, Gemini 4 Argon) and the open DeepSeek-V3.2. Every reliable fooling of a monitor into approving a fraud came through an evidence source and reached only that source's readers. Fooling it into blocking a correct payment depended on injection resistance, a monitor's refusal to follow instructions planted in its data. Small models and DeepSeek obeyed such notes. The three closed models ignored them in all but one case, and Opus also resisted persuasion tailored to it. On new data, a rule that counts sources instead of votes blocked 0.9% of correct payments, against 7.4% for the best counting rule, at equal safety. The pre-registered worst-case test failed. One attack left the approvers on a single source, so no rule over the twelve language-model monitors survives every measured attack. Adding a rule engine, which cannot be persuaded, restored the condition for a safe and live rule on the measured attacks. Diversify the inputs, harden the monitors, count sources, not votes.

arXiv:2610.29172v1 Announce Type: cross Abstract: Conditional Flow Models (CFMs) are increasingly being used in place of deterministic predictors across machine learning, including for tasks that were classically formulated as regression. Given their success in such tasks, we ask the question: what useful information do the conditional distributions produced by CFMs carry beyond what a point predictor would provide? We identify a systematic phenomenon: in many natural regimes, CFM samples routinely have larger variance than the true conditional distribution, even in nearly deterministic regression problems. This effect is most pronounced where the model has not resolved the dependence on the conditioning variable, and thus the additional variance can be viewed as converting epistemic uncertainty in fitting the ground truth conditional distribution into apparent aleatoric variance in the generated samples. Through controlled synthetic and real-data experiments, and comparisons with explicit smoothing estimators, we show that the induced variance tracks model underfitting of complex or hard-to-fit targets, and decreases with additional training or data. We provide a theoretical account of the phenomenon in a simplified setting and demonstrate a consequent practical use-case of conditional flow-matching: designing high-coverage predictive sets that cover the ground-truth regression target.

arXiv:2610.29171v1 Announce Type: cross Abstract: We study how to improve training efficiency on Ascend NPUs. In the NPU training programs we study, framework-level compilation optimizations such as operator fusion are only partially applied, leaving fragmented computation and avoidable execution overheads. Profiling-guided tuning can address these inefficiencies, but it often requires costly trial and error. To make training on NPUs more efficient, we propose TrainBridge, an agentic approach for transferring GPU optimization guidance to NPU training. TrainBridge uses the highly optimized execution produced by mature GPU compilation stacks as a task-specific reference for NPU tuning. It extracts optimization intents from the GPU training execution to guide NPU program rewrites and further refines the rewritten program through iterative profiling. Our GPU-to-NPU migration experiments show that TrainBridge achieves an average 1.7x speedup across four workloads covering LLM pretraining, image classification, and vision-language training, and an average speedup of nearly 2.0x across 27 Modded-NanoGPT speedrun records with highly optimized GPU implementations.

arXiv:2610.29168v1 Announce Type: cross Abstract: World action models (WAMs) are shown to substantially diverge under generalization performance based on their generative backbones. We investigate this gap through the representational geometry that WAMs inherit from their backbones and how policy training reorganizes it. Our central hypothesis is that a useful representational prior would structure geometry that separates nuisance information from the in-distribution manifold in advance of training as a policy. We quantify this with nuisance orthogonality; orthogonality of the feature under visual variation outside the empirical in-distribution manifold. Across various image and video backbones, we show that a higher orthogonality from the representational prior is associated with stronger robustness to visual variations after WAM training, enabling policy generalization measurable prospectively for world models. In addition, analyzing the resulting robust WAMs reveals nuisance suppression; relative suppression of feature displacement caused by such variations and robust action generation of the policy. Together, these findings provide a geometric perspective on how the representational prior of the world model shapes the generalization performance of WAMs.

arXiv:2610.29162v1 Announce Type: cross Abstract: There is a widespread sense that AI-generated text has recently become harder to understand. In this study, we investigate a set of psycholinguistic features that are known to correlate with human processing difficulty (e.g., object relative clauses, long syntactic dependencies, low-frequency n-grams). Using a novel corpus of AI-generated text across diverse domains, we find that AI models use these difficult features more often than paraphrases generated by round-trip translation. Compared with human writing, all nine frontier and near-frontier models use more inanimate subjects with animate verbs and more low-frequency n-grams, and most also use longer dependencies, more stacked nouns, and more reduced object relatives in conversation; we also demonstrate substantial variation across model families. We propose and test a hypothesis as to why: while human language is optimized for successful communication, AI-generated text is increasingly optimized for other goals like math, code, and reasoning, in which there is a pressure for \textit{information density} rather than successful communication. We provide a proof-of-concept for this hypothesis by fine-tuning a language model under a length penalty, which leads to a divergence from human-like linguistic behavior. Our findings highlight the axes of misalignment between linguistic processing demands in humans and LLMs, and point toward possible remedies.

arXiv:2610.29160v1 Announce Type: cross Abstract: This paper explores the potential of Artificial Intelligence (AI) to enable end-to-end automotive design and software development for Software-Defined Vehicles (SDVs). Based on our contributions within the Central Car Server (CeCaS) and follow-up EU-funded projects, we present an AI-driven toolchain integrating Large Language Models (LLMs) with Model-Driven Engineering (MDE) to automate the transformation of natural-language requirements into system architectures, executable software, and validation artifacts. The approach emphasizes traceability, consistency, and compliance with safety-critical automotive development requirements. Validation is performed on a centralized vehicle computing platform supporting adaptive cruise control and emergency braking, tested in Hardware-in-the-Loop (HiL) and Software-in-the-Loop (SiL) environments. We discuss current capabilities, practical limitations, and future directions toward increasingly automated, reliable, and quality-assured automotive software development with minimal human intervention.

arXiv:2610.29152v1 Announce Type: cross Abstract: A generated chest radiograph can be useful as declared augmentation, but the same image becomes unsupported evidence when submitted as the record of a real examination. We study this problem by separating three questions: whether an image passes an automated quality check, whether an origin classifier flags it as generated, and whether a trusted record supports the claimed acquisition. In our experiments, Qwen selects up to 20 images from pools of 1,000 existing Cheff-identified or SynthCheX candidates. A receiver evaluates quality with MedGemma and predicted origin with a classifier on frozen DINOv3 features. Quality-based selection raises quality passes from 12 to 19 of 20 for Cheff-identified images and from 8 to 17 for SynthCheX. We find no evidence that selection improves passage through both checks: no selected Cheff-identified image passes both, while uniform, quality, and quality-constrained surrogate selection produce 2, 1, and 2 SynthCheX passes. On unprocessed test images, the origin classifier flags 4,677 of 5,000 SynthCheX images, but it also misclassify 110 of 2,297 NIH (real) images and 568 of 957 genuine images from another collection as fake. Our results further show that JPEG compression reduces SynthCheX-versus-NIH AUROC from 0.987 to 0.841. Even without processing, 152 SynthCheX images pass both pixel checks. These checks can filter submissions, but they cannot verify an acquisition claim. We strongly advocate that source and intended use must be recorded and checked separately for admission of chest radiographs in medical datasets.

arXiv:2610.29150v1 Announce Type: cross Abstract: Timbre similarity is a key evaluation target for generative speech systems, including voice conversion, speaker-cloning text-to-speech, and zero-shot speech models. When the generated voice does not match the target speaker, the problem often reflects imperfect separation of speaker-related timbre from content, making it important to identify where timbre information resides and which representations still carry it. Using speaker identity as a measurable proxy, we probe frequency bands, phase/magnitude structure, WORLD source/filter decomposition, voiced/unvoiced regions, self-supervised learning (SSL) layers, EnCodec/HuBERT tokens, and a 45-listener test across Mandarin, English, and Japanese speech/singing data. Speaker information is distributed: 1--8 kHz magnitude and vocal-tract spectral-envelope structure dominate perception, above-8 kHz bands remain machine-discriminative, shallow SSL layers expose compact speaker information, and token leakage depends on tokenizer/layer choice. These findings indicate where generation systems should condition on or transform speaker information, and where content, codec, and token pipelines should audit residual speaker leakage.

arXiv:2610.29145v1 Announce Type: cross Abstract: Large language models increasingly automate optimization modeling from natural language problems, producing written mathematics, solver code and a returned report. Verifying one optimization problem across these artifacts requires semantic comparison: unequal models can attain the same objective, equivalent models can use different auxiliary variables, and mutually consistent artifacts can omit the same source requirement. ORSMT provides a typed semantic interface and exact SMT relation queries. It aligns observable decisions, projects auxiliary variables, proves or refutes feasible region and objective relations, localizes refutations to implicated rows or domains and links source texts when available. It then reconstructs one artifact from another, and each adopted candidate passes the required relation checks. Across 3,034 problems and six generator sizes, ORSMT improves the model-code match rate in all 72 benchmark and generator combinations before selection. Protected selection raises the aggregate rate from 43.2% to 88.2% and the share that are both correct and matched from 25.2% to 40.2%. On the 1,905 refuted outputs, certified voting under the shared fallback makes 50.9% of outputs correct and matched, compared with 44.8% for ReLoop and 49.4% for Opt-Verifier. A separate one-round source revision followed by exact reconstruction restores a model-code match for 84.4% of the refuted pairs.

arXiv:2610.29140v1 Announce Type: cross Abstract: Engineering graduates in India enter a technology sector that is being reshaped by generative AI, yet the route into that sector runs through examinations and campus placement screens that largely exclude AI. Institutions in India have an approved intake of nearly 2.9 million seats in engineering and technology for 2026-2027, but HCI research on how students prepare for AI-mediated work has drawn mostly on Western and East Asian settings. We interviewed 23 students and recent graduates of Indian engineering and computing programmes about how they used generative AI across coursework, examinations and placements, and how they judged their own readiness for work. Using reflexive thematic analysis, we report five themes, covering AI-first workplaces and AI-free placement screens, how participants prepared for placement, how they used AI in coursework, the role of faculty authority and enforcement, and participants' concerns about their readiness for work. Drawing on signaling theory, we discuss what grades and screen scores certify once AI can help produce them, and we offer recommendations for employers, institutions and AI tool designers.

arXiv:2610.29136v1 Announce Type: cross Abstract: Continuous diffusion language models generate a whole sequence in parallel, never drawing a single word at any intermediate step; only the final step turns embeddings into text. Models such as ELF now report autoregressive-level quality at a fraction of the sampling steps of discrete-token diffusion models, and this design principle -- no step-wise discretization -- is what their own ablations credit for the gain. We show that this same principle carries a hidden cost: a collapse of the lexical long tail. Measured against held-out human text, ELF's output matches the human rate on common words almost exactly, yet its rarest words appear at only 0.30x the human rate, a gap that does not close from 105M to 652M parameters and that no sampler setting these models expose can reach. A same-size GPT-2, in contrast, spans zero to above the human rate by its decoding rule alone. We trace the difference to what continuous diffusion models lack: with no step that ever draws a discrete word, every step is instead driven by an average over the candidates still in play, and averaging a rare candidate together with the common word competing for its slot lands on the common word. An autoregressive model avoids this because its decoding rule samples one candidate rather than averaging over all of them. We propose GRAFT, a training-free method that grafts this missing choice back onto a frozen model: while the model is still undecided, GRAFT commits the rare words the average would otherwise discard, then regenerates the sequence from scratch with those commitments held in place. Across three ELF scales and on LangFlow, an independent continuous dLLM, GRAFT restores the rare-word share of generated text from 0.42-0.77x the human rate to 0.82-1.10x, a relative gain of 34-109%, and moves the entire word-frequency profile closer to human on every model tested, by 2.7x on ELF-B and 12x on LangFlow.

arXiv:2610.29129v1 Announce Type: cross Abstract: Reward hacking is a major problem in current production machine learning systems, resulting in costly and dangerous failures. Here we study how to monitor for and mitigate reward hacking in realistic settings, including tool calls and long contexts. We apply a dynamic intervention to steering directions obtained by three different methods: difference-in-means vectors from labeled rollouts, gradient optimization-based vectors, and vectors extracted when models reference an anti-reward-hacking model spec. This dynamic intervention perturbs model representations less than uniform steering vector addition to the residual stream, allowing for more precise edits. We find that we are able to cheaply and significantly reduce reward hacking without correspondingly significant costs to model coherence or capability, and that steering vectors originally extracted from coding environments also generalize to reducing reward hacking in other domains. Our results pave the way towards latent space monitoring and control over realistic production systems.