AI 빠른 뉴스

arXiv:2512.09757v1 Announce Type: cross Abstract: Transformers generate valid and diverse chemical structures, but little is known about the mechanisms that enable these models to capture the rules of molecular representation. We present a mechanistic analysis of autoregressive transformers trained on drug-like small molecules to reveal the computational structure underlying their capabilities across multiple levels of abstraction. We identify computational patterns consistent with low-level syntactic parsing and more abstract chemical validity constraints. Using sparse autoencoders (SAEs), we extract feature dictionaries associated with chemically relevant activation patterns. We validate our findings on downstream tasks and find that mechanistic insights can translate to predictive performance in various practical settings.

arXiv:2512.07801v3 Announce Type: replace-cross Abstract: LLM-based agents are increasingly deployed for expert decision support, yet human-AI teams in high-stakes settings do not yet reliably outperform the best individual. We argue this complementarity gap reflects a fundamental mismatch: current agents are trained as answer engines, not as partners in the collaborative sensemaking through which experts actually make decisions. Sensemaking (the ability to co-construct causal explanations, surface uncertainties, and adapt goals) is the key capability that current training pipelines do not explicitly develop or evaluate. We propose Collaborative Causal Sensemaking (CCS) as a research agenda to develop this capability from the ground up, spanning new training environments that reward collaborative thinking, representations for shared human-AI mental models, and evaluation centred on trust and complementarity. Taken together, these directions shift MAS research from building oracle-like answer engines to cultivating AI teammates that co-reason with their human partners over the causal structure of shared decisions, advancing the design of effective human-AI teams.

arXiv:2512.09610v1 Announce Type: cross Abstract: People living with Motor Neuron Disease (plwMND) frequently encounter speech and motor impairments that necessitate a reliance on augmentative and alternative communication (AAC) systems. This paper tackles the main challenge that traditional symbol-based AAC systems offer a limited vocabulary, while text entry solutions tend to exhibit low communication rates. To help plwMND articulate their needs about the system efficiently and effectively, we iteratively design and develop a novel multimodal text generation system called ImageTalk through a tailored proxy-user-based and an end-user-based design phase. The system demonstrates pronounced keystroke savings of 95.6%, coupled with consistent performance and high user satisfaction. We distill three design guidelines for AI-assisted text generation systems design and outline four user requirement levels tailored for AAC purposes, guiding future research in this field.

arXiv:2509.15872v2 Announce Type: replace-cross Abstract: Prediction of complete step-by-step chemical reaction mechanisms (CRMs) remains a major challenge. Whereas the traditional approaches in CRM tasks rely on expert-driven experiments or costly quantum chemical computations, contemporary deep learning (DL) alternatives ignore key intermediates and mechanistic steps and often suffer from hallucinations. We present DeepMech, an interpretable graph-based DL framework employing atom- and bond-level attention, guided by generalized templates of mechanistic operations (TMOps), to generate CRMs. Trained on our curated ReactMech dataset (~30K CRMs with 100K atom-mapped and mass-balanced elementary steps), DeepMech achieves 98.98+/-0.12% accuracy in predicting elementary steps and 95.94+/-0.21% in complete CRM tasks, besides maintaining high fidelity even in out-of-distribution scenarios as well as in predicting side and/or byproducts. Extension to multistep CRMs relevant to prebiotic chemistry, demonstrates the ability of DeepMech in effectively reconstructing 2 pathways from simple primordial substrates to complex biomolecules such as serine and aldopentose. Attention analysis identifies reactive atoms/bonds in line with chemical intuition, rendering our model interpretable and suitable for reaction design.

arXiv:2510.17281v3 Announce Type: replace-cross Abstract: Scaling up data, parameters, and test-time computation has been the mainstream methods to improve LLM systems (LLMsys), but their upper bounds are almost reached due to the gradual depletion of high-quality data and marginal gains obtained from larger computational resource consumption. Inspired by the abilities of human and traditional AI systems in learning from practice, constructing memory and continual learning frameworks for LLMsys has become an important and popular research direction in recent literature. Yet, existing benchmarks for LLM memory often focus on evaluating the system on homogeneous reading comprehension tasks with long-form inputs rather than testing their abilities to learn from accumulated user feedback in service time. Therefore, we propose a user feedback simulation framework and a comprehensive benchmark covering multiple domains, languages, and types of tasks to evaluate the continual learning abilities of LLMsys. Experiments show that the effectiveness and efficiency of state-of-the-art baselines are far from satisfying, and we hope this benchmark could pave the way for future studies on LLM memory and optimization algorithms.

arXiv:2511.06316v2 Announce Type: replace Abstract: Reliable geospatial information on road accidents is vital for safety analysis and infrastructure planning, yet most low- and middle-income countries continue to face a critical shortage of accurate, location-specific crash data. Existing text-based geocoding tools perform poorly in multilingual and unstructured news environments, where incomplete place descriptions and mixed language (e.g. Bangla-English) scripts obscure spatial context. To address these limitations, this study introduces ALIGN (Accident Location Inference through Geo-Spatial Neural Reasoning), a vision-language framework that emulates human spatial reasoning to infer accident location coordinates directly from available textual and map-based cues. ALIGN integrates large language and vision-language model mechanisms within a multi-stage pipeline that performs optical character recognition, linguistic reasoning, and map-level verification through grid-based spatial scanning. The framework systematically evaluates each predicted location against contextual and visual evidence, ensuring interpretable, fine-grained geolocation outcomes without requiring model retraining. Applied to Bangla-language news data source, ALIGN demonstrates consistent improvements over traditional geoparsing methods, accurately identifying district- and sub-district-level crash sites. Beyond its technical contribution, the framework establishes a high accuracy foundation for automated crash mapping in data-scarce regions, supporting evidence-driven road-safety policymaking and the broader integration of multimodal artificial intelligence in transportation analytics.

arXiv:2503.19804v2 Announce Type: replace-cross Abstract: Low-light image enhancement is crucial for a myriad of applications, from night vision and surveillance, to autonomous driving. However, due to the inherent limitations that come in hand with capturing images in low-illumination environments, the task of enhancing such scenes still presents a formidable challenge. To advance research in this field, we introduce our Low Exposure Night Vision (LENVIZ) Dataset, a comprehensive multi-exposure benchmark dataset for low-light image enhancement comprising of over 230K frames showcasing 24K real-world indoor and outdoor, with-and without human, scenes. Captured using 3 different camera sensors, LENVIZ offers a wide range of lighting conditions, noise levels, and scene complexities, making it the largest publicly available up-to 4K resolution benchmark in the field. LENVIZ includes high quality human-generated ground truth, for which each multi-exposure low-light scene has been meticulously curated and edited by expert photographers to ensure optimal image quality. Furthermore, we also conduct a comprehensive analysis of current state-of-the-art low-light image enhancement techniques on our dataset and highlight potential areas of improvement.

arXiv:2512.09629v1 Announce Type: new Abstract: We present an end-to-end framework for planning supported by verifiers. An orchestrator receives a human specification written in natural language and converts it into a PDDL (Planning Domain Definition Language) model, where the domain and problem are iteratively refined by sub-modules (agents) to address common planning requirements, such as time constraints and optimality, as well as ambiguities and contradictions that may exist in the human specification. The validated domain and problem are then passed to an external planning engine to generate a plan. The orchestrator and agents are powered by Large Language Models (LLMs) and require no human intervention at any stage of the process. Finally, a module translates the final plan back into natural language to improve human readability while maintaining the correctness of each step. We demonstrate the flexibility and effectiveness of our framework across various domains and tasks, including the Google NaturalPlan benchmark and PlanBench, as well as planning problems like Blocksworld and the Tower of Hanoi (where LLMs are known to struggle even with small instances). Our framework can be integrated with any PDDL planning engine and validator (such as Fast Downward, LPG, POPF, VAL, and uVAL, which we have tested) and represents a significant step toward end-to-end planning aided by LLMs.

arXiv:2512.09727v1 Announce Type: new Abstract: Monte Carlo Tree Search is a cornerstone algorithm for online planning, and its root-parallel variant is widely used when wall clock time is limited but best performance is desired. In environments with continuous action spaces, how to best aggregate statistics from different threads is an important yet underexplored question. In this work, we introduce a method that uses Gaussian Process Regression to obtain value estimates for promising actions that were not trialed in the environment. We perform a systematic evaluation across 6 different domains, demonstrating that our approach outperforms existing aggregation strategies while requiring a modest increase in inference time.

arXiv:2511.03092v5 Announce Type: replace Abstract: The proliferation of 100B+ parameter Large Language Models (LLMs) with 100k+ context length support have resulted in increasing demands for on-chip memory to support large KV caches. Techniques such as StreamingLLM and SnapKV demonstrate how to control KV cache size while maintaining model accuracy. Yet, these techniques are not commonly used within industrial deployments using frameworks like vLLM or SGLang. The reason is twofold: on one hand, the static graphs and continuous batching methodology employed by these frameworks make it difficult to admit modifications to the standard multi-head attention algorithm, while on the other hand, the accuracy implications of such techniques on modern instruction-following and reasoning models are not well understood, obfuscating the need for implementing these techniques. In this paper, we explore these accuracy implications on Llama-3.1-8B-Instruct and DeepSeek-R1, and develop SnapStream, a KV cache compression method that can be deployed at scale. We demonstrate the efficacy of SnapStream in a 16-way tensor-parallel deployment of DeepSeek-671B on SambaNova SN40L accelerators running at 128k context length and up to 1832 tokens per second in a real production setting. SnapStream enables $4\times$ improved on-chip memory usage and introduces minimal accuracy degradation on LongBench-v2, AIME24 and LiveCodeBench. To the best of our knowledge, this is the first implementation of sparse KV attention techniques deployed in a production inference system with static graphs and continuous batching.

arXiv:2503.11901v4 Announce Type: replace-cross Abstract: This study characterizes GPU resilience in Delta, a large-scale AI system that consists of 1,056 A100 and H100 GPUs, with over 1,300 petaflops of peak throughput. We used 2.5 years of operational data (11.7 million GPU hours) on GPU errors. Our major findings include: (i) H100 GPU memory resilience is worse than A100 GPU memory, with 3.2x lower per-GPU MTBE for memory errors, (ii) The GPU memory error-recovery mechanisms on H100 GPUs are insufficient to handle the increased memory capacity, (iii) H100 GPUs demonstrate significantly improved GPU hardware resilience over A100 GPUs with respect to critical hardware components, (iv) GPU errors on both A100 and H100 GPUs frequently result in job failures due to the lack of robust recovery mechanisms at the application level, and (v) We project the impact of GPU node availability on larger-scales and find that significant overprovisioning of 5% is necessary to handle GPU failures.

arXiv:2510.15843v2 Announce Type: replace-cross Abstract: Accurately detecting sentiment polarity and intensity in product reviews and social media posts remains challenging due to informal and domain-specific language. To address this, we propose a novel hybrid lexicon-fuzzy-transformer framework that combines rule-based heuristics, contextual deep learning, and fuzzy logic to generate continuous sentiment scores reflecting both polarity and strength. The pipeline begins with VADER-based initial sentiment estimations, which are refined through a two-stage adjustment process. This involves leveraging confidence scores from DistilBERT, a lightweight transformer and applying fuzzy logic principles to mitigate excessive neutrality bias and enhance granularity. A custom fuzzy inference system then maps the refined scores onto a 0 to 1 continuum, producing expert)like judgments. The framework is rigorously evaluated on four domain-specific datasets. food delivery, e-commerce, tourism, and fashion. Results show improved alignment with user ratings, better identification of sentiment extremes, and reduced misclassifications. Both quantitative metrics (distributional alignment, confusion matrices) and qualitative insights (case studies, runtime analysis) affirm the models robustness and efficiency. This work demonstrates the value of integrating symbolic reasoning with neural models for interpretable, finegrained sentiment analysis in linguistically dynamic domains.

arXiv:2508.19366v4 Announce Type: replace-cross Abstract: Hallucinations in LLMs–especially in multimodal settings–undermine reliability. We present a rigorous information-geometric framework, grounded in diffusion dynamics, to quantify hallucinations in MLLMs where model outputs are embedded via spectral decompositions of multimodal graph Laplacians, and their gaps to a truth manifold define a semantic distortion metric. We derive Courant-Fischer bounds on a temperature-dependent hallucination profile and use RKHS eigenmodes to obtain modality-aware, interpretable measures that track evolution over prompts and time. This reframes hallucination as quantifiable and bounded, providing a principled basis for evaluation and mitigation.

arXiv:2512.09591v1 Announce Type: cross Abstract: Polysomnography (PSG), the gold standard test for sleep analysis, generates vast amounts of multimodal clinical data, presenting an opportunity to leverage self-supervised representation learning (SSRL) for pre-training foundation models to enhance sleep analysis. However, progress in sleep foundation models is hindered by two key limitations: (1) the lack of a shared dataset and benchmark with diverse tasks for training and evaluation, and (2) the absence of a systematic evaluation of SSRL approaches across sleep-related tasks. To address these gaps, we introduce Stanford Sleep Bench, a large-scale PSG dataset comprising 17,467 recordings totaling over 163,000 hours from a major sleep clinic, including 13 clinical disease prediction tasks alongside canonical sleep-related tasks such as sleep staging, apnea diagnosis, and age estimation. We systematically evaluate SSRL pre-training methods on Stanford Sleep Bench, assessing downstream performance across four tasks: sleep staging, apnea diagnosis, age estimation, and disease and mortality prediction. Our results show that multiple pretraining methods achieve comparable performance for sleep staging, apnea diagnosis, and age estimation. However, for mortality and disease prediction, contrastive learning significantly outperforms other approaches while also converging faster during pretraining. To facilitate reproducibility and advance sleep research, we will release Stanford Sleep Bench along with pretrained model weights, training pipelines, and evaluation code.

arXiv:2512.09892v1 Announce Type: cross Abstract: While modern language models and their inner workings are incredibly complex, recent work (Golowich, Liu & Shetty; 2025) has proposed a simple and potentially tractable abstraction for them through the observation that empirically, these language models all seem to have approximately low logit rank. Roughly, this means that a matrix formed by the model’s log probabilities of various tokens conditioned on certain sequences of tokens is well approximated by a low rank matrix. In this paper, our focus is on understanding how this structure can be exploited algorithmically for obtaining provable learning guarantees. Since low logit rank models can encode hard-to-learn distributions such as noisy parities, we study a query learning model with logit queries that reflects the access model for common APIs. Our main result is an efficient algorithm for learning any approximately low logit rank model from queries. We emphasize that our structural assumption closely reflects the behavior that is empirically observed in modern language models. Thus, our result gives what we believe is the first end-to-end learning guarantee for a generative model that plausibly captures modern language models.

arXiv:2512.09458v1 Announce Type: new Abstract: This chapter argues that the reliability of agentic and generative AI is chiefly an architectural property. We define agentic systems as goal-directed, tool-using decision makers operating in closed loops, and show how reliability emerges from principled componentisation (goal manager, planner, tool-router, executor, memory, verifiers, safety monitor, telemetry), disciplined interfaces (schema-constrained, validated, least-privilege tool calls), and explicit control and assurance loops. Building on classical foundations, we propose a practical taxonomy-tool-using agents, memory-augmented agents, planning and self-improvement agents, multi-agent systems, and embodied or web agents – and analyse how each pattern reshapes the reliability envelope and failure modes. We distil design guidance on typed schemas, idempotency, permissioning, transactional semantics, memory provenance and hygiene, runtime governance (budgets, termination conditions), and simulate-before-actuate safeguards.

arXiv:2509.11719v2 Announce Type: replace Abstract: Multi-agent trajectory prediction in autonomous driving requires a comprehensive understanding of complex social dynamics. Existing methods, however, often struggle to capture the full richness of these dynamics, particularly the co-existence of multi-scale interactions and the diverse behaviors of heterogeneous agents. To address these challenges, this paper introduces HeLoFusion, an efficient and scalable encoder for modeling heterogeneous and multi-scale agent interactions. Instead of relying on global context, HeLoFusion constructs local, multi-scale graphs centered on each agent, allowing it to effectively model both direct pairwise dependencies and complex group-wise interactions (\textit{e.g.}, platooning vehicles or pedestrian crowds). Furthermore, HeLoFusion tackles the critical challenge of agent heterogeneity through an aggregation-decomposition message-passing scheme and type-specific feature networks, enabling it to learn nuanced, type-dependent interaction patterns. This locality-focused approach enables a principled representation of multi-level social context, yielding powerful and expressive agent embeddings. On the challenging Waymo Open Motion Dataset, HeLoFusion achieves state-of-the-art performance, setting new benchmarks for key metrics including Soft mAP and minADE. Our work demonstrates that a locality-grounded architecture, which explicitly models multi-scale and heterogeneous interactions, is a highly effective strategy for advancing motion forecasting.

arXiv:2512.09566v1 Announce Type: new Abstract: Drug discovery is a time-consuming and expensive process, with traditional high-throughput and docking-based virtual screening hampered by low success rates and limited scalability. Recent advances in generative modelling, including autoregressive, diffusion, and flow-based approaches, have enabled de novo ligand design beyond the limits of enumerative screening. Yet these models often suffer from inadequate generalization, limited interpretability, and an overemphasis on binding affinity at the expense of key pharmacological properties, thereby restricting their translational utility. Here we present Trio, a molecular generation framework integrating fragment-based molecular language modeling, reinforcement learning, and Monte Carlo tree search, for effective and interpretable closed-loop targeted molecular design. Through the three key components, Trio enables context-aware fragment assembly, enforces physicochemical and synthetic feasibility, and guides a balanced search between the exploration of novel chemotypes and the exploitation of promising intermediates within protein binding pockets. Experimental results show that Trio reliably achieves chemically valid and pharmacologically enhanced ligands, outperforming state-of-the-art approaches with improved binding affinity (+7.85%), drug-likeness (+11.10%) and synthetic accessibility (+12.05%), while expanding molecular diversity more than fourfold.

arXiv:2410.08060v2 Announce Type: replace-cross Abstract: Many numerical and learning algorithms rely on the solution of the Monge-Kantorovich problem and Wasserstein distances, which provide appropriate distributional metrics. While the natural approach is to treat the problem as an infinite-dimensional linear programming, such a methodology limits the computational performance due to the polynomial scaling with respect to the sample size along with intensive memory requirements. We propose a novel alternative framework to address the Monge-Kantorovich problem based on a projection type gradient descent scheme. The dynamics builds on the notion of the conditional expectation, where the connection with the opinion dynamics is leveraged to devise efficient numerical schemes. We demonstrate that the resulting dynamics recovers random maps with favourable computational performance. Along with the theoretical insight, the proposed dynamics paves the way for innovative approaches to construct numerical schemes for computing optimal transport maps as well as Wasserstein distances.

arXiv:2512.09340v1 Announce Type: new Abstract: Understanding how humans and AI systems interpret ambiguous visual stimuli offers critical insight into the nature of perception, reasoning, and decision-making. This paper examines image labeling performance across human participants and deep neural networks, focusing on low-resolution, perceptually degraded stimuli. Drawing from computational cognitive science, cognitive architectures, and connectionist-symbolic hybrid models, we contrast human strategies such as analogical reasoning, shape-based recognition, and confidence modulation with AI’s feature-based processing. Grounded in Marr’s tri-level hypothesis, Simon’s bounded rationality, and Thagard’s frameworks of representation and emotion, we analyze participant responses in relation to Grad-CAM visualizations of model attention. Human behavior is further interpreted through cognitive principles modeled in ACT-R and Soar, revealing layered and heuristic decision strategies under uncertainty. Our findings highlight key parallels and divergences between biological and artificial systems in representation, inference, and confidence calibration. The analysis motivates future neuro-symbolic architectures that unify structured symbolic reasoning with connectionist representations. Such architectures, informed by principles of embodiment, explainability, and cognitive alignment, offer a path toward AI systems that are not only performant but also interpretable and cognitively grounded.

arXiv:2506.16659v2 Announce Type: replace-cross Abstract: Training large language models (LLMs) typically relies on adaptive optimizers such as Adam, which introduce extra operations and require significant more memory to maintain first- and second-order moments than SGD. While recent works such as GaLore, Fira and APOLLO have proposed state-compressed variants to reduce memory consumption, a fundamental question remains: What are the minimum modifications to plain SGD needed to match state-of-the-art pretraining performance? We systematically investigate this question using a bottom-up approach, and identify two simple yet highly (memory- and compute-) efficient techniques: (1) column-wise gradient normalization (normalizing the gradient along the output dimension), which boosts SGD performance without momentum; and (2) applying first-order momentum only to the output layer, where gradient variance is highest. Combining these two techniques lead to SCALE (Stochastic Column-normAlized Last-layer momEntum), a simple optimizer for memory efficient pretraining. Across multiple LLaMA models (60M-1B), SCALE matches or exceeds the performance of Adam while using only 35-45% of the total memory. It also consistently outperforms memory-efficient optimizers such as GaLore, Fira and APOLLO, making it a strong candidate for large-scale pretraining under memory constraints. For LLaMA 7B model, SCALE outperforms the state-of-the-art memory-efficient methods APOLLO and Muon, in terms of both perplexity and memory consumption.

arXiv:2512.09142v1 Announce Type: new Abstract: We present SDialog, an MIT-licensed open-source Python toolkit that unifies dialog generation, evaluation and mechanistic interpretability into a single end-to-end framework for building and analyzing LLM-based conversational agents. Built around a standardized \texttt{Dialog} representation, SDialog provides: (1) persona-driven multi-agent simulation with composable orchestration for controlled, synthetic dialog generation, (2) comprehensive evaluation combining linguistic metrics, LLM-as-a-judge and functional correctness validators, (3) mechanistic interpretability tools for activation inspection and steering via feature ablation and induction, and (4) audio generation with full acoustic simulation including 3D room modeling and microphone effects. The toolkit integrates with all major LLM backends, enabling mixed-backend experiments under a unified API. By coupling generation, evaluation, and interpretability in a dialog-centric architecture, SDialog enables researchers to build, benchmark and understand conversational systems more systematically.

arXiv:2510.06961v3 Announce Type: replace-cross Abstract: Despite rapid progress, ASR evaluation remains saturated with short-form English, and efficiency is rarely reported. We present the Open ASR Leaderboard, a fully reproducible benchmark and interactive leaderboard comparing 60+ open-source and proprietary systems across 11 datasets, including a dedicated multilingual track. We standardize text normalization and report both word error rate (WER) and inverse real-time factor (RTFx), enabling fair accuracy-efficiency comparisons. For English transcription, Conformer encoders paired with LLM decoders achieve the best average WER but are slower, while CTC and TDT decoders deliver much better RTFx, making them attractive for long-form and offline use. Whisper-derived encoders fine-tuned for English improve accuracy but often trade off multilingual coverage. All code and dataset loaders are open-sourced to support transparent, extensible evaluation.

arXiv:2512.09117v1 Announce Type: new Abstract: This paper presents a formal, categorical framework for analysing how humans and large language models (LLMs) transform content into truth-evaluated propositions about a state space of possible worlds W , in order to argue that LLMs do not solve but circumvent the symbol grounding problem.

arXiv:2512.01317v2 Announce Type: replace-cross Abstract: Measurement-induced entanglement (MIE) captures how local measurements generate long-range quantum correlations and drive dynamical phase transitions in many-body systems. Yet estimating MIE experimentally remains challenging: direct evaluation requires extensive post-selection over measurement outcomes, raising the question of whether MIE is accessible with only polynomial resources. We address this challenge by reframing MIE detection as a data-driven learning problem that assumes no prior knowledge of state preparation. Using measurement records alone, we train a neural network in a self-supervised manner to predict the uncertainty metric for MIE–the gap between upper and lower bounds of the average post-measurement bipartite entanglement. Applied to random circuits with one-dimensional all-to-all connectivity and two-dimensional nearest-neighbor coupling, our method reveals a learnability transition with increasing circuit depth: below a threshold, the uncertainty is small and decreases with polynomial measurement data and model parameters, while above it the uncertainty remains large despite increasing resources. We further verify this transition experimentally on current noisy quantum devices, demonstrating its robustness to realistic noise. These results highlight the power of data-driven approaches for learning MIE and delineate the practical limits of its classical learnability.

arXiv:2512.09872v1 Announce Type: cross Abstract: Generative Artificial Intelligence models, such as Large Language Models (LLMs) and Large Vision Models (VLMs), exhibit state-of-the-art performance but remain vulnerable to hardware-based threats, specifically bit-flip attacks (BFAs). Existing BFA discovery methods lack generalizability and struggle to scale, often failing to analyze the vast parameter space and complex interdependencies of modern foundation models in a reasonable time. This paper proposes FlipLLM, a reinforcement learning (RL) architecture-agnostic framework that formulates BFA discovery as a sequential decision-making problem. FlipLLM combines sensitivity-guided layer pruning with Q-learning to efficiently identify minimal, high-impact bit sets that can induce catastrophic failure. We demonstrate the effectiveness and generalizability of FlipLLM by applying it to a diverse set of models, including prominent text-only LLMs (GPT-2 Large, LLaMA 3.1 8B, and DeepSeek-V2 7B), VLMs such as LLaVA 1.6, and datasets, such as MMLU, MMLU-Pro, VQAv2, and TextVQA. Our results show that FlipLLM can identify critical bits that are vulnerable to BFAs up to 2.5x faster than SOTA methods. We demonstrate that flipping the FlipLLM-identified bits plummets the accuracy of LLaMA 3.1 8B from 69.9% to ~0.2%, and for LLaVA’s VQA score from 78% to almost 0%, by flipping as few as 5 and 7 bits, respectively. Further analysis reveals that applying standard hardware protection mechanisms, such as ECC SECDED, to the FlipLLM-identified bit locations completely mitigates the BFA impact, demonstrating the practical value of our framework in guiding hardware-level defenses. FlipLLM offers the first scalable and adaptive methodology for exploring the BFA vulnerability of both language and multimodal foundation models, paving the way for comprehensive hardware-security evaluation.

Presented by Oracle NetSuiteWhen any company tells you it is their biggest product release in almost three decades, it’s worth listening. When the person saying it founded the world’s first cloud computing company, it’s time to take note. At SuiteWorld 2025, Evan Goldberg, founder and EVP of Oracle NetSuite, did just that when he called NetSuite Next the company’s biggest product evolution in nearly three decades. But behind that sweeping vision lies a quieter shift — one centered on how AI behaves, not just what it can do. “Every company is experimenting with AI,” says Brian Chess, SVP of Technology and AI at NetSuite. “Some ideas hit the mark, and some don’t, but each one teaches us something. That’s how innovation works.”For Chess and Gary Wiessinger, SVP of Application Development at NetSuite, the challenge lies in governing AI responsibly. Rather than reinventing its system, NetSuite is extending the same principles into the AI era that have guided its strategy for 27 years — security, control, and auditability. The goal is to make AI actions traceable, permissions enforceable, and outcomes auditable.The philosophy underpins what Chess calls a “glass-box” approach to enterprise AI, where decisions are visible and every agent operates within human-defined guardrails.Built on Oracle’s foundationNetSuite Next is the result of five years of development. It is built on Oracle Cloud Infrastructure (OCI), which is relied on by many of the world’s most important AI model providers, and has AI capabilities integrated directly into its core rather than added as a separate layer.“We are building a fantastic foundation on OCI,” Chess says. “That infrastructure provides more than compute power.” Built on the same OCI foundation that powers NetSuite today, NetSuite Next gives customers access to Oracle’s latest AI innovations along with the performance, scalability, and security of OCI’s enterprise-grade platform.Wiessinger emphasizes the team’s approach as “needs first, technology second.”“We don’t take a technology-first approach,” he says. “We take a customer-needs-first approach and then figure out how to use the latest technology to solve those needs better.”That philosophy extends across Oracle’s ecosystem. NetSuite’s collaboration with Oracle’s AI Database, Fusion Applications, Analytics, and Cloud Infrastructure teams helps NetSuite deliver capabilities that independent vendors can’t match, he says — an AI system that is both open to innovation and grounded in Oracle’s security and scale.The data structure advantageAt the heart of the platform is a structured data model that serves as a critical advantage.“One of the great things about NetSuite is, because the data comes in and it gets structured, the connections between the data are explicit,” Chess explains. “That means the AI can start exploring that knowledge graph that the company has been building up.”Where general LLMs sift through unstructured text, NetSuite’s AI works from structured data, identifying precise links between transactions, accounts, and workflows to deliver context-aware insights. Wiessinger adds, “The data we have spans financials, CRM, commerce, and HR. We can do more for customers because we see more of their business in one place.” Combined with built-in business logic and metadata, that scope allows NetSuite to generate recommendations and insights that are accurate and explainable.Oracle’s Redwood design system provides the visual layer for this data intelligence, creating what Goldberg described as a “modern, clean and intuitive” workspace where AI and humans collaborate naturally.Designing for accountabilityOne downside of enterprise AI is that many systems still function as a black box — they produce results but offer little visibility into how they reached them. NetSuite is different. It is designing its systems around transparency, making visibility a defining feature.“When users can see how AI reached a decision — tracing the path from A to B — they don’t just verify accuracy,” Chess says. “They learn how the AI knew to do that.”That visibility turns AI into a learning engine. As Chess puts it, transparency becomes a “fantastic teacher,” helping organizations understand, improve, and trust automation over time.But Chess cautions against blind trust: “What’s disturbing is when someone presents something to me and says, ‘Look what AI gave me,’ as if that makes it authoritative. People need to ask, ‘What grounded this? Why is it correct?’” NetSuite’s answer is traceability. When someone asks, “Where did this number come from?” the system can show them the full reasoning behind it.Governance by designAI agents inside NetSuite Next follow the same governance model as employees: roles, permissions, and escalation rules. Role-based security embedded directly into workflows helps ensure that agents act only within authorized boundaries.Wiessinger puts it plainly: “If AI generates a narrative summary of a report and it’s 80% of what the user would have written, that’s fine. We’ll learn from their feedback and make it even better. But booking to the general ledger is different. That has to be 100% correct and is where controls and human review really matter.”Auditing the algorithmAuditing has always been part of ERP’s DNA, and NetSuite now extends that discipline to AI. Every agent action, workflow adjustment, and model-generated code snippet is recorded within the system’s existing audit framework. As Chess explains, “It’s the same audit trail you might use to figure out what the humans did. Code is auditable. When the LLM creates code and something happens in the system, we can trace back.”That traceability transforms AI from a black box into a glass box. When an algorithm accelerates a payment or flags an anomaly, teams can see exactly which inputs and logic produced the decision — an essential safeguard for regulated industries and finance teams.Safe extensibilityThe other half of trust is freedom — the ability to extend AI without risking data exposure.The NetSuite AI Connector Service and SuiteCloud Platform make that possible. Through standards like the Model Context Protocol (MCP), customers can connect external language models while keeping sensitive data secure inside Oracle’s environment.“Businesses are hungry for AI,” Chess says. “They want to start putting it to work. But they also want to know those experiments can’t go off the rails. The NetSuite AI Connector Service and governance model give partners the freedom to innovate while maintaining the same audit and permission logic that govern native features.”Culture, experimentation, and guardrailsGovernance frameworks only work if people use them wisely. Both executives see AI adoption as a top-down and bottom-up process.“The board is telling the CEO they need an AI strategy,” Chess says. “Meanwhile, employees are already using AI. If I were a CEO, I’d start by asking: what are you already doing, and what’s working?”Wiessinger agrees that balance is key: “Some companies go all-in on a centralized AI team while others let everyone experiment freely. Neither works by itself. You need structure for major initiatives and freedom for grassroots innovation.”He offers a simple example: “Write an email? Go crazy. Touch financials or employee data? Don’t go crazy with that.”Experimentation, both emphasize, is imperative. “No one should wait for us or anyone else,” Wiessinger says. “Start testing, learn quickly, and be intentional about making it work for your business.”Why transparent AI winsAs AI moves deeper into enterprise operations, governance will define competitive advantage as much as innovation. NetSuite’s approach — extending its heritage of ERP controls into the age of autonomous systems, built on Oracle’s secure cloud infrastructure and structured-data foundation — positions it to lead in both.In a world of opaque models and risky promises, the companies that win won’t just build smarter AI. They’ll build AI you can trust.Sponsored articles are content produced by a company that is either paying for the post or has a business relationship with VentureBeat, and they’re always clearly marked. For more information, contact [email protected].

arXiv:2512.09736v1 Announce Type: new Abstract: Multi-Agent Path Finding (MAPF) algorithms are increasingly deployed in industrial warehouses and automated manufacturing facilities, where robots must operate reliably under real-world physical constraints. However, existing MAPF evaluation frameworks typically rely on simplified robot models, leaving a substantial gap between algorithmic benchmarks and practical performance. Recent frameworks such as SMART, incorporate kinodynamic modeling and offer the MAPF community a platform for large-scale, realistic evaluation. Building on this capability, this work investigates how key planner design choices influence performance under realistic execution settings. We systematically study three fundamental factors: (1) the relationship between solution optimality and execution performance, (2) the sensitivity of system performance to inaccuracies in kinodynamic modeling, and (3) the interaction between model accuracy and plan optimality. Empirically, we examine these factors to understand how these design choices affect performance in realistic scenarios. We highlight open challenges and research directions to steer the community toward practical, real-world deployment.

arXiv:2512.09678v1 Announce Type: cross Abstract: In this article, we explore the use of various matrix norms for optimizing functions of weight matrices, a crucial problem in training large language models. Moving beyond the spectral norm underlying the Muon update, we leverage duals of the Ky Fan $k$-norms to introduce a family of Muon-like algorithms we name Fanions, which are closely related to Dion. By working with duals of convex combinations of the Ky Fan $k$-norms with either the Frobenius norm or the $l_\infty$ norm, we construct the families of F-Fanions and S-Fanions, respectively. Their most prominent members are F-Muon and S-Muon. We complement our theoretical analysis with an extensive empirical study of these algorithms across a wide range of tasks and settings, demonstrating that F-Muon and S-Muon consistently match Muon’s performance, while outperforming vanilla Muon on a synthetic linear least squares problem.

arXiv:2412.11927v5 Announce Type: replace Abstract: Procedural mistake detection (PMD) is a challenging problem of classifying whether a human user (observed through egocentric video) has successfully executed a task (specified by a procedural text). Despite significant recent efforts, machine performance in the wild remains nonviable, and the reasoning processes underlying this performance are opaque. As such, we extend PMD to require generating visual self-dialog rationales to inform decisions. Given the impressive, mature image understanding capabilities observed in recent vision-and-language models (VLMs), we curate a suitable benchmark dataset for PMD based on individual frames. As our reformulation enables unprecedented transparency, we leverage a natural language inference (NLI) model to formulate two automated metrics for the coherence of generated rationales. We establish baselines for this reframed task, showing that VLMs struggle off-the-shelf, but with some trade-offs, their accuracy, coherence, and efficiency can be improved by incorporating these metrics into common inference and fine-tuning methods. Lastly, our multi-faceted metrics visualize common outcomes, highlighting areas for further improvement.

arXiv:2512.09824v1 Announce Type: cross Abstract: Visual concept composition, which aims to integrate different elements from images and videos into a single, coherent visual output, still falls short in accurately extracting complex concepts from visual inputs and flexibly combining concepts from both images and videos. We introduce Bind & Compose, a one-shot method that enables flexible visual concept composition by binding visual concepts with corresponding prompt tokens and composing the target prompt with bound tokens from various sources. It adopts a hierarchical binder structure for cross-attention conditioning in Diffusion Transformers to encode visual concepts into corresponding prompt tokens for accurate decomposition of complex visual concepts. To improve concept-token binding accuracy, we design a Diversify-and-Absorb Mechanism that uses an extra absorbent token to eliminate the impact of concept-irrelevant details when training with diversified prompts. To enhance the compatibility between image and video concepts, we present a Temporal Disentanglement Strategy that decouples the training process of video concepts into two stages with a dual-branch binder structure for temporal modeling. Evaluations demonstrate that our method achieves superior concept consistency, prompt fidelity, and motion quality over existing approaches, opening up new possibilities for visual creativity.

arXiv:2512.09829v1 Announce Type: new Abstract: The massive scale of modern AI accelerators presents critical challenges to traditional fault assessment methodologies, which face prohibitive computational costs and provide poor coverage of critical failure modes. This paper introduces RIFT (Reinforcement Learning-guided Intelligent Fault Targeting), a scalable framework that automates the discovery of minimal, high-impact fault scenarios for efficient design-time fault assessment. RIFT transforms the complex search for worst-case faults into a sequential decision-making problem, combining hybrid sensitivity analysis for search space pruning with reinforcement learning to intelligently generate minimal, high-impact test suites. Evaluated on billion-parameter Large Language Model (LLM) workloads using NVIDIA A100 GPUs, RIFT achieves a \textbf{2.2$\times$} fault assessment speedup over evolutionary methods and reduces the required test vector volume by over \textbf{99\%} compared to random fault injection, all while achieving \textbf{superior fault coverage}. The proposed framework also provides actionable data to enable intelligent hardware protection strategies, demonstrating that RIFT-guided selective error correction code provides a \textbf{12.8$\times$} improvement in \textbf{cost-effectiveness} (coverage per unit area) compared to uniform triple modular redundancy protection. RIFT automatically generates UVM-compliant verification artifacts, ensuring its findings are directly actionable and integrable into commercial RTL verification workflows.

arXiv:2512.09912v1 Announce Type: cross Abstract: In-context learning with attention enables large neural networks to make context-specific predictions by selectively focusing on relevant examples. Here, we adapt this idea to supervised learning procedures such as lasso regression and gradient boosting, for tabular data. Our goals are to (1) flexibly fit personalized models for each prediction point and (2) retain model simplicity and interpretability. Our method fits a local model for each test observation by weighting the training data according to attention, a supervised similarity measure that emphasizes features and interactions that are predictive of the outcome. Attention weighting allows the method to adapt to heterogeneous data in a data-driven way, without requiring cluster or similarity pre-specification. Further, our approach is uniquely interpretable: for each test observation, we identify which features are most predictive and which training observations are most relevant. We then show how to use attention weighting for time series and spatial data, and we present a method for adapting pretrained tree-based models to distributional shift using attention-weighted residual corrections. Across real and simulated datasets, attention weighting improves predictive performance while preserving interpretability, and theory shows that attention-weighting linear models attain lower mean squared error than the standard linear model under mixture-of-models data-generating processes with known subgroup structure.

arXiv:2407.01630v2 Announce Type: replace-cross Abstract: Recommendation systems and assistants (in short, recommenders) influence through online platforms most actions of our daily lives, suggesting items or providing solutions based on users’ preferences or requests. This survey systematically reviews, categories, and discusses the impact of recommenders in four human-AI ecosystems — social media, online retail, urban mapping and generative AI ecosystems. Its scope is to systematise a fast-growing field in which terminologies employed to classify methodologies and outcomes are fragmented and unsystematic. This is a crucial contribution to the literature because terminologies vary substantially across disciplines and ecosystems, hindering comparison and accumulation of knowledge in the field. We follow the customary steps of qualitative systematic review, gathering 154 articles from different disciplines to develop a parsimonious taxonomy of methodologies employed (empirical, simulation, observational, controlled), outcomes observed (concentration, content degradation, discrimination, diversity, echo chamber, filter bubble, homogenisation, polarisation, radicalisation, volume), and their level of analysis (individual, item, and ecosystem). We systematically discuss substantive and methodological commonalities across ecosystems, and highlight potential avenues for future research. The survey is addressed to scholars and practitioners interested in different human-AI ecosystems, policymakers and institutional stakeholders who want to understand better the measurable outcomes of recommenders, and tech companies who wish to obtain a systematic view of the impact of their recommenders.

arXiv:2510.19788v3 Announce Type: replace Abstract: Model-learning agents should gather information to learn world models that support many downstream tasks and inferences, such as predicting unobserved states, estimating near- and far-term consequences of actions, planning action sequences, and detecting changes in dynamics. Current methods for learning and evaluating world models diverge from this goal: training and evaluation are anchored to next-frame prediction, and success is scored by reward maximization in the same environment. We propose WorldTest, a protocol to evaluate model-learning agents that separates reward-free interaction from a scored test phase in a different but related environment. WorldTest is open-ended $\unicode{x2014}$ models should support many different tasks unknown ahead of time $\unicode{x2014}$ and agnostic to model representation, allowing comparison across approaches. We instantiated WorldTest with AutumnBench, a suite of 43 interactive grid-world environments and 129 tasks across three families: masked-frame prediction, planning, and predicting changes to the causal dynamics. We compared 517 human participants and three frontier models on AutumnBench. We found that humans outperform the models, and scaling compute improves performance only in some environments but not others. WorldTest provides a novel template $\unicode{x2014}$ reward-free exploration, derived tests, and behavior-based scoring $\unicode{x2014}$ to evaluate what agents learn about environment dynamics, and AutumnBench exposes significant headroom in world-model learning.

arXiv:2512.09831v1 Announce Type: new Abstract: This paper develops a geometric framework for modeling belief, motivation, and influence across cognitively heterogeneous agents. Each agent is represented by a personalized value space, a vector space encoding the internal dimensions through which the agent interprets and evaluates meaning. Beliefs are formalized as structured vectors-abstract beings-whose transmission is mediated by linear interpretation maps. A belief survives communication only if it avoids the null spaces of these maps, yielding a structural criterion for intelligibility, miscommunication, and belief death. Within this framework, I show how belief distortion, motivational drift, counterfactual evaluation, and the limits of mutual understanding arise from purely algebraic constraints. A central result-“the No-Null-Space Leadership Condition”-characterizes leadership as a property of representational reachability rather than persuasion or authority. More broadly, the model explains how abstract beings can propagate, mutate, or disappear as they traverse diverse cognitive geometries. The account unifies insights from conceptual spaces, social epistemology, and AI value alignment by grounding meaning preservation in structural compatibility rather than shared information or rationality. I argue that this cognitive-geometric perspective clarifies the epistemic boundaries of influence in both human and artificial systems, and offers a general foundation for analyzing belief dynamics across heterogeneous agents.

arXiv:2511.20321v2 Announce Type: replace Abstract: We seek to clarify the concept of active inference by disentangling it from the Free Energy Principle. We show how the optimizations that need to be carried out in order to implement active inference in discrete state spaces can be formulated as constrained divergence minimization problems which can be solved by standard mean field methods that do not appeal to the idea of expected free energy. When it is used to model perception, the perception/action divergence criterion that we propose coincides with variational free energy. When it is used to model action, it differs from an expected free energy functional by an entropy regularizer.

arXiv:2506.07040v3 Announce Type: replace-cross Abstract: We present a non-asymptotic convergence analysis of $Q$-learning and actor-critic algorithms for robust average-reward Markov Decision Processes (MDPs) under contamination, total-variation (TV) distance, and Wasserstein uncertainty sets. A key ingredient of our analysis is showing that the optimal robust $Q$ operator is a strict contraction with respect to a carefully designed semi-norm (with constant functions quotiented out). This property enables a stochastic approximation update that learns the optimal robust $Q$-function using $\tilde{\mathcal{O}}(\epsilon^{-2})$ samples. We also provide an efficient routine for robust $Q$-function estimation, which in turn facilitates robust critic estimation. Building on this, we introduce an actor-critic algorithm that learns an $\epsilon$-optimal robust policy within $\tilde{\mathcal{O}}(\epsilon^{-2})$ samples. We provide numerical simulations to evaluate the performance of our algorithms.

arXiv:2502.17801v3 Announce Type: replace-cross Abstract: Cloud computing environments are increasingly vulnerable to security threats such as distributed denial-of-service (DDoS) attacks and SQL injection. Traditional security mechanisms, based on rule matching and feature recognition, struggle to adapt to evolving attack strategies. This paper proposes an adaptive security protection framework leveraging deep learning to construct a multi-layered defense architecture. The proposed system is evaluated in a real-world business environment, achieving a detection accuracy of 97.3%, an average response time of 18 ms, and an availability rate of 99.999%. Experimental results demonstrate that the proposed method significantly enhances detection accuracy, response efficiency, and resource utilization, offering a novel and effective approach to cloud computing security.

arXiv:2512.09882v1 Announce Type: new Abstract: We present the first comprehensive evaluation of AI agents against human cybersecurity professionals in a live enterprise environment. We evaluate ten cybersecurity professionals alongside six existing AI agents and ARTEMIS, our new agent scaffold, on a large university network consisting of ~8,000 hosts across 12 subnets. ARTEMIS is a multi-agent framework featuring dynamic prompt generation, arbitrary sub-agents, and automatic vulnerability triaging. In our comparative study, ARTEMIS placed second overall, discovering 9 valid vulnerabilities with an 82% valid submission rate and outperforming 9 of 10 human participants. While existing scaffolds such as Codex and CyAgent underperformed relative to most human participants, ARTEMIS demonstrated technical sophistication and submission quality comparable to the strongest participants. We observe that AI agents offer advantages in systematic enumeration, parallel exploitation, and cost — certain ARTEMIS variants cost $18/hour versus $60/hour for professional penetration testers. We also identify key capability gaps: AI agents exhibit higher false-positive rates and struggle with GUI-based tasks.

arXiv:2505.14295v2 Announce Type: replace-cross Abstract: Data encoding plays a fundamental and distinctive role in Quantum Machine Learning (QML). While classical approaches process data directly as vectors, QML may require transforming classical data into quantum states through encoding circuits, known as quantum feature maps or quantum embeddings. This step leverages the inherently high-dimensional and non-linear nature of Hilbert space, enabling more efficient data separation in complex feature spaces that may be inaccessible to classical methods. This encoding part significantly affects the performance of the QML model, so it is important to choose the right encoding method for the dataset to be encoded. However, this choice is generally arbitrary, since there is no “universal” rule for knowing which encoding to choose based on a specific set of data. There are currently a variety of encoding methods using different quantum logic gates. We studied the most commonly used types of encoding methods and benchmarked them using different datasets.

arXiv:2510.06249v5 Announce Type: replace-cross Abstract: The 2025 Multimodal Models for Low-Resource Contexts and Social Impact (MMLoSo) Language Challenge addresses one of India’s most pressing linguistic gaps: the lack of resources for its diverse low-resource languages (LRLs). In this study, we investigate whether enforcing cross-lingual similarity in specific internal layers of a decoder-only multilingual large language model (LLM) can improve translation quality from LRL to high-resource language (HRL). Specifically, we combine Centered Kernel Alignment (CKA), a similarity metric that encourages representations of different languages to align, with REPINA, a regularization method that constrains parameter updates to remain close to the pretrained model, into a joint method we call TRepLiNa. In this research project, we experiment with zero-shot, few-shot, and fine-tuning settings using Aya-23 8B with QLoRA across MMLoSo shared task language pairs (Mundari, Santali, Bhili) with Hindi/English pivots. Our results show that aligning mid-level layers using TRepLiNa (CKA+REPINA) is a low-cost, practical approach to improving LRL translation, especially in data-scarce settings.

arXiv:2508.06783v2 Announce Type: replace-cross Abstract: Alignment is a key step in developing Large Language Models (LLMs) using human feedback to ensure adherence to human values and societal norms. Dependence on human feedback raises privacy concerns about how much a labeler’s preferences may reveal about their personal values, beliefs, and personality traits. Existing approaches, such as Differentially Private SGD (DP-SGD), provide rigorous privacy guarantees by privatizing gradients during fine-tuning and alignment but can provide more privacy than necessary as human preferences are tied only to labels of (prompt, response) pairs and can degrade model utility. This work focuses on LLM alignment with preference-level privacy, which preserves the privacy of preference labels provided by humans. We propose PROPS (PROgressively Private Self-alignment), a multi-stage privacy preserving alignment framework where privately aligned models in previous stages can serve as labelers for supplementing training data in the subsequent stages of alignment. We present theoretical guarantees for PROPS as well as comprehensive validation using multiple models (Pythia and GPT) and datasets (AlpacaEval, Anthropic HH-RLHF, truthy-dpo-v0.1) to demonstrate the utility of PROPS over existing methods while still providing high privacy. For the same privacy budget, alignment via PROPS can achieve up to 3x higher win-rates compared to DP-SGD, and 2.5x higher win-rates compared to Randomized Response (RR) based alignment.

arXiv:2512.09895v1 Announce Type: new Abstract: Metadata vocabularies are essential for advancing FAIR and FARR data principles, but their development constrained by limited human resources and inconsistent standardization practices. This paper introduces MatSci-YAMZ, a platform that integrates artificial intelligence (AI) and human-in-the-loop (HILT), including crowdsourcing, to support metadata vocabulary development. The paper reports on a proof-of-concept use case evaluating the AI-HILT model in materials science, a highly interdisciplinary domain Six (6) participants affiliated with the NSF Institute for Data-Driven Dynamical Design (ID4) engaged with the MatSci-YAMZ plaform over several weeks, contributing term definitions and providing examples to prompt the AI-definitions refinement. Nineteen (19) AI-generated definitions were successfully created, with iterative feedback loops demonstrating the feasibility of AI-HILT refinement. Findings confirm the feasibility AI-HILT model highlighting 1) a successful proof of concept, 2) alignment with FAIR and open-science principles, 3) a research protocol to guide future studies, and 4) the potential for scalability across domains. Overall, MatSci-YAMZ’s underlying model has the capacity to enhance semantic transparency and reduce time required for consensus building and metadata vocabulary development.

arXiv:2509.24125v3 Announce Type: replace-cross Abstract: In this technical note, we study the problem of inverse permutation learning in decoder-only transformers. Given a permutation and a string to which that permutation has been applied, the model is tasked with producing the original (“canonical”) string. We argue that this task models a natural robustness property across a variety of reasoning tasks, including long-context retrieval, multiple choice QA and in-context learning. Our primary contribution is an impossibility result: we show that an arbitrary depth, decoder-only transformer cannot learn this task. This result concerns the expressive capacity of decoder-only transformer models and is agnostic to training dynamics or sample complexity. We give a pair of alternative constructions under which inverse permutation learning is feasible. The first of these highlights the fundamental role of the causal attention mask, and reveals a gap between the expressivity of encoder-decoder transformers and the more popular decoder-only architecture. The latter result is more surprising: we show that simply padding the input with “scratch tokens” yields a construction under which inverse permutation learning is possible. We conjecture that this may suggest an alternative mechanism by which chain-of-thought prompting or, more generally, intermediate “thinking” tokens can enable reasoning in large language models, even when these tokens encode no meaningful semantic information (e.g., the results of intermediate computations).

arXiv:2512.01034v2 Announce Type: replace-cross Abstract: Neural networks have shown remarkable success in supervised learning when trained on a single task using a fixed dataset. However, when neural networks are trained on a reinforcement learning task, their ability to continue learning from new experiences declines over time. This decline in learning ability is known as plasticity loss. To restore plasticity, prior work has explored periodically resetting the parameters of the learning network, a strategy that often improves overall performance. However, such resets come at the cost of a temporary drop in performance, which can be dangerous in real-world settings. To overcome this instability, we introduce AltNet, a reset-based approach that restores plasticity without performance degradation by leveraging twin networks. The use of twin networks anchors performance during resets through a mechanism that allows networks to periodically alternate roles: one network learns as it acts in the environment, while the other learns off-policy from the active network’s interactions and a replay buffer. At fixed intervals, the active network is reset and the passive network, having learned from prior experiences, becomes the new active network. AltNet restores plasticity, improving sample efficiency and achieving higher performance, while avoiding performance drops that pose risks in safety-critical settings. We demonstrate these advantages in several high-dimensional control tasks from the DeepMind Control Suite, where AltNet outperforms various relevant baseline methods, as well as state-of-the-art reset-based techniques.

arXiv:2510.12604v3 Announce Type: replace-cross Abstract: With the rise of modern search and recommendation platforms, insufficient collaborative information of cold-start items exacerbates the Matthew effect of existing platform items, challenging platform diversity and becoming a longstanding issue. Existing methods align items’ side content with collaborative information to transfer collaborative signals from high-popularity items to cold-start items. However, these methods fail to account for the asymmetry between collaboration and content, nor the fine-grained differences among items. To address these issues, we propose SMILE, an item representation enhancement approach based on fused alignment of semantic IDs. Specifically, we use RQ-OPQ encoding to quantize item content and collaborative information, followed by a two-step alignment: RQ encoding transfers shared collaborative signals across items, while OPQ encoding learns differentiated information of items. Comprehensive offline experiments on large-scale industrial datasets demonstrate superiority of SMILE, and rigorous online A/B tests confirm statistically significant improvements: item CTR +1.66%, buyers +1.57%, and order volume +2.17%.

arXiv:2512.09897v1 Announce Type: new Abstract: Long-term planning in complex, text-based environments presents significant challenges due to open-ended action spaces, ambiguous observations, and sparse feedback. Recent research suggests that large language models (LLMs) encode rich semantic knowledge about the world, which can be valuable for guiding agents in high-level reasoning and planning across both embodied and purely textual settings. However, existing approaches often depend heavily on querying LLMs during training and inference, making them computationally expensive and difficult to deploy efficiently. In addition, these methods typically employ a pretrained, unaltered LLM whose parameters remain fixed throughout training, providing no opportunity for adaptation to the target task. To address these limitations, we introduce SCOPE (Subgoal-COnditioned Pretraining for Efficient planning), a one-shot hierarchical planner that leverages LLM-generated subgoals only at initialization to pretrain a lightweight student model. Unlike prior approaches that distill LLM knowledge by repeatedly prompting the model to adaptively generate subgoals during training, our method derives subgoals directly from example trajectories. This design removes the need for repeated LLM queries, significantly improving efficiency, though at the cost of reduced explainability and potentially suboptimal subgoals. Despite their suboptimality, our results on the TextCraft environment show that LLM-generated subgoals can still serve as a strong starting point for hierarchical goal decomposition in text-based planning tasks. Compared to the LLM-based hierarchical agent ADaPT (Prasad et al., 2024), which achieves a 0.52 success rate, our method reaches 0.56 and reduces inference time from 164.4 seconds to just 3.0 seconds.

arXiv:2511.18828v2 Announce Type: replace-cross Abstract: Over the last few months, AI models including large language models have improved greatly. There are now several documented examples where they have helped professional mathematical scientists prove new results, sometimes even helping resolve known open problems. In this short note, we add another example to the list, by documenting how we were able to solve a previously unsolved research problem in robust mathematical statistics with crucial help from GPT-5. Our problem concerns robust density estimation, where the observations are perturbed by Wasserstein-bounded contaminations. In a previous preprint (Chao and Dobriban, 2023, arxiv:2308.01853v2), we have obtained upper and lower bounds on the minimax optimal estimation error; which were, however, not sharp. Starting in October 2025, making significant use of GPT-5 Pro, we were able to derive the minimax optimal error rate (reported in version 3 of the above arxiv preprint). GPT-5 provided crucial help along the way, including by suggesting calculations that we did not think of, and techniques that were not familiar to us, such as the dynamic Benamou-Brenier formulation, for key steps in the analysis. Working with GPT-5 took a few weeks of effort, and we estimate that it could have taken several months to get the same results otherwise. At the same time, there are still areas where working with GPT-5 was challenging: it sometimes provided incorrect references, and glossed over details that sometimes took days of work to fill in. We outline our workflow and steps taken to mitigate issues. Overall, our work can serve as additional documentation for a new age of human-AI collaborative work in mathematical science.

arXiv:2512.09114v1 Announce Type: new Abstract: The deployment of AI systems faces three critical governance challenges that current frameworks fail to adequately address. First, organizations struggle with inadequate risk assessment at the use case level, exemplified by the Humana class action lawsuit and other high impact cases where an AI system deployed to production exhibited both significant bias and high error rates, resulting in improper healthcare claim denials. Each AI use case presents unique risk profiles requiring tailored governance, yet most frameworks provide one size fits all guidance. Second, existing frameworks like ISO 42001 and NIST AI RMF remain at high conceptual levels, offering principles without actionable controls, leaving practitioners unable to translate governance requirements into specific technical implementations. Third, organizations lack mechanisms for operationalizing governance at scale, with no systematic approach to embed trustworthy AI practices throughout the development lifecycle, measure compliance quantitatively, or provide role-appropriate visibility from boards to data scientists. We present AI TIPS, Artificial Intelligence Trust-Integrated Pillars for Sustainability 2.0, update to the comprehensive operational framework developed in 2019,four years before NIST’s AI Risk Management Framework, that directly addresses these challenges.

arXiv:2512.06813v2 Announce Type: replace-cross Abstract: High-performance concrete requires complex mix design decisions involving interdependent variables and practical constraints. While data-driven methods have improved predictive modeling for forward design in concrete engineering, inverse design remains limited, especially when some variables are fixed and only the remaining ones must be inferred. This study proposes a cooperative neural network framework for the partial inverse design of high-performance concrete. The framework integrates an imputation model with a surrogate strength predictor and learns through cooperative training. Once trained, it generates valid and performance-consistent mix designs in a single forward pass without retraining for different constraint scenarios. Compared with baseline models, including autoencoder models and Bayesian inference with Gaussian process surrogates, the proposed method achieves R-squared values of 0.87 to 0.92 and substantially reduces mean squared error by approximately 50% and 70%, respectively. The results show that the framework provides an accurate and computationally efficient foundation for constraint-aware, data-driven mix proportioning.

arXiv:2512.09908v1 Announce Type: new Abstract: Moralisation and Triangulation are transformations allowing to switch between different ways of factoring a probability distribution into a graphical model. Moralisation allows to view a Bayesian network (a directed model) as a Markov network (an undirected model), whereas triangulation addresses the opposite direction. We present a categorical framework where these transformations are modelled as functors between a category of Bayesian networks and one of Markov networks. The two kinds of network (the objects of these categories) are themselves represented as functors from a `syntax’ domain to a `semantics’ codomain. Notably, moralisation and triangulation can be defined inductively on such syntax via functor pre-composition. Moreover, while moralisation is fully syntactic, triangulation relies on semantics. This leads to a discussion of the variable elimination algorithm, reinterpreted here as a functor in its own right, that splits the triangulation procedure in two: one purely syntactic, the other purely semantic. This approach introduces a functorial perspective into the theory of probabilistic graphical models, which highlights the distinctions between syntactic and semantic modifications.

arXiv:2512.08518v2 Announce Type: replace-cross Abstract: Social robots must adjust to human proxemic norms to ensure user comfort and engagement. While prior research demonstrates that eye-tracking features reliably estimate comfort in human-human interactions, their applicability to interactions with humanoid robots remains unexplored. In this study, we investigate user comfort with the robot “Ameca” across four experimentally controlled distances (0.5 m to 2.0 m) using mobile eye-tracking and subjective reporting (N=19). We evaluate multiple machine learning and deep learning models to estimate comfort based on gaze features. Contrary to previous human-human studies where Transformer models excelled, a Decision Tree classifier achieved the highest performance (F1-score = 0.73), with minimum pupil diameter identified as the most critical predictor. These findings suggest that physiological comfort thresholds in human-robot interaction differ from human-human dynamics and can be effectively modeled using interpretable logic.

arXiv:2512.09673v1 Announce Type: cross Abstract: Equivariant neural networks encode symmetry as an inductive bias and have achieved strong empirical performance in wide domains. However, their expressive power remains not well understood. Focusing on 2-layer ReLU networks, this paper investigates the impact of equivariance constraints on the expressivity of equivariant and layer-wise equivariant networks. By examining the boundary hyperplanes and the channel vectors of ReLU networks, we construct an example showing that equivariance constraints could strictly limit expressive power. However, we demonstrate that this drawback can be compensated via enlarging the model size. Furthermore, we show that despite a larger model size, the resulting architecture could still correspond to a hypothesis space with lower complexity, implying superior generalizability for equivariant networks.

arXiv:2512.09586v1 Announce Type: cross Abstract: Quantum circuit design is a key bottleneck for practical quantum machine learning on complex, real-world data. We present an automated framework that discovers and refines variational quantum circuits (VQCs) using graph-based Bayesian optimization with a graph neural network (GNN) surrogate. Circuits are represented as graphs and mutated and selected via an expected improvement acquisition function informed by surrogate uncertainty with Monte Carlo dropout. Candidate circuits are evaluated with a hybrid quantum-classical variational classifier on the next generation firewall telemetry and network internet of things (NF-ToN-IoT-V2) cybersecurity dataset, after feature selection and scaling for quantum embedding. We benchmark our pipeline against an MLP-based surrogate, random search, and greedy GNN selection. The GNN-guided optimizer consistently finds circuits with lower complexity and competitive or superior classification accuracy compared to all baselines. Robustness is assessed via a noise study across standard quantum noise channels, including amplitude damping, phase damping, thermal relaxation, depolarizing, and readout bit flip noise. The implementation is fully reproducible, with time benchmarking and export of best found circuits, providing a scalable and interpretable route to automated quantum circuit discovery.

arXiv:2107.05664v1 Announce Type: cross Abstract: With the adoption of autonomous vehicles on our roads, we will witness a mixed-autonomy environment where autonomous and human-driven vehicles must learn to co-exist by sharing the same road infrastructure. To attain socially-desirable behaviors, autonomous vehicles must be instructed to consider the utility of other vehicles around them in their decision-making process. Particularly, we study the maneuver planning problem for autonomous vehicles and investigate how a decentralized reward structure can induce altruism in their behavior and incentivize them to account for the interest of other autonomous and human-driven vehicles. This is a challenging problem due to the ambiguity of a human driver’s willingness to cooperate with an autonomous vehicle. Thus, in contrast with the existing works which rely on behavior models of human drivers, we take an end-to-end approach and let the autonomous agents to implicitly learn the decision-making process of human drivers only from experience. We introduce a multi-agent variant of the synchronous Advantage Actor-Critic (A2C) algorithm and train agents that coordinate with each other and can affect the behavior of human drivers to improve traffic flow and safety.

arXiv:2512.09662v1 Announce Type: cross Abstract: Hate speech spreads widely online, harming individuals and communities, making automatic detection essential for large-scale moderation, yet detecting it remains difficult. Part of the challenge lies in subjectivity: what one person flags as hate speech, another may see as benign. Traditional annotation agreement metrics, such as Cohen’s $\kappa$, oversimplify this disagreement, treating it as an error rather than meaningful diversity. Meanwhile, Large Language Models (LLMs) promise scalable annotation, but prior studies demonstrate that they cannot fully replace human judgement, especially in subjective tasks. In this work, we reexamine LLM reliability using a subjectivity-aware framework, cross-Rater Reliability (xRR), revealing that even under fairer lens, LLMs still diverge from humans. Yet this limitation opens an opportunity: we find that LLM-generated annotations can reliably reflect performance trends across classification models, correlating with human evaluations. We test this by examining whether LLM-generated annotations preserve the relative ordering of model performance derived from human evaluation (i.e. whether models ranked as more reliable by human annotators preserve the same order when evaluated with LLM-generated labels). Our results show that, although LLMs differ from humans at the instance level, they reproduce similar ranking and classification patterns, suggesting their potential as proxy evaluators. While not a substitute for human annotators, they might serve as a scalable proxy for evaluation in subjective NLP tasks.

arXiv:2512.09806v1 Announce Type: cross Abstract: U-Net and other U-shaped architectures have achieved significant success in image deconvolution tasks. However, challenges have emerged, as these methods might generate unrealistic artifacts or hallucinations, which can interfere with analysis in safety-critical scenarios. This paper introduces a novel approach for quantifying and comprehending hallucination artifacts to ensure trustworthy computer vision models. Our method, termed the Conformal Hallucination Estimation Metric (CHEM), is applicable to any image reconstruction model, enabling efficient identification and quantification of hallucination artifacts. It offers two key advantages: it leverages wavelet and shearlet representations to efficiently extract hallucinations of image features and uses conformalized quantile regression to assess hallucination levels in a distribution-free manner. Furthermore, from an approximation theoretical perspective, we explore the reasons why U-shaped networks are prone to hallucinations. We test the proposed approach on the CANDELS astronomical image dataset with models such as U-Net, SwinUNet, and Learnlets, and provide new perspectives on hallucination from different aspects in deep learning-based image processing.

arXiv:2512.09729v1 Announce Type: cross Abstract: We present Ethics Readiness Levels (ERLs), a four-level, iterative method to track how ethical reflection is implemented in the design of AI systems. ERLs bridge high-level ethical principles and everyday engineering by turning ethical values into concrete prompts, checks, and controls within real use cases. The evaluation is conducted using a dynamic, tree-like questionnaire built from context-specific indicators, ensuring relevance to the technology and application domain. Beyond being a managerial tool, ERLs help facilitate a structured dialogue between ethics experts and technical teams, while our scoring system helps track progress over time. We demonstrate the methodology through two case studies: an AI facial sketch generator for law enforcement and a collaborative industrial robot. The ERL tool effectively catalyzes concrete design changes and promotes a shift from narrow technological solutionism to a more reflective, ethics-by-design mindset.

arXiv:2211.10585v1 Announce Type: cross Abstract: Autonomous vehicle (AV) navigation in the presence of Human-driven vehicles (HVs) is challenging, as HVs continuously update their policies in response to AVs. In order to navigate safely in the presence of complex AV-HV social interactions, the AVs must learn to predict these changes. Humans are capable of navigating such challenging social interaction settings because of their intrinsic knowledge about other agents behaviors and use that to forecast what might happen in the future. Inspired by humans, we provide our AVs the capability of anticipating future states and leveraging prediction in a cooperative reinforcement learning (RL) decision-making framework, to improve safety and robustness. In this paper, we propose an integration of two essential and earlier-presented components of AVs: social navigation and prediction. We formulate the AV decision-making process as a RL problem and seek to obtain optimal policies that produce socially beneficial results utilizing a prediction-aware planning and social-aware optimization RL framework. We also propose a Hybrid Predictive Network (HPN) that anticipates future observations. The HPN is used in a multi-step prediction chain to compute a window of predicted future observations to be used by the value function network (VFN). Finally, a safe VFN is trained to optimize a social utility using a sequence of previous and predicted observations, and a safety prioritizer is used to leverage the interpretable kinematic predictions to mask the unsafe actions, constraining the RL policy. We compare our prediction-aware AV to state-of-the-art solutions and demonstrate performance improvements in terms of efficiency and safety in multiple simulated scenarios.