Pre-trained video LLMs struggle with auxiliary streams like audio or depth maps due to modality interference. UniMVU uses instruction-aware dynamic gating at two levels (inner-modality and modality-level) to adaptively balance importance, achieving gains up to 13.5 CIDEr across six benchmarks and aligning with human-interpretable relevance.

This study introduces EnterpriseMem-Bench, a multi-turn Text-to-SQL benchmark with 300 sessions and 1,400 turns. Evaluating five frontier models reveals: stateless models collapse to zero accuracy by Turn 3; memory complexity does not monotonically improve performance, with working memory dominating; Claude Sonnet 4.6 shows generational regression on SEC EDGAR; and under reasoning, Claude error distributions become mono-modal.

This paper proposes a generalizable framework for cultural evaluation and intervention in LLMs, using scenario-based behavioral probing and activation steering to shift internal alignments without retraining. Experiments reveal latent entanglement between cultural dimensions, limiting precise value alignment.

A new study investigates why large language models (LLMs) hallucinate on structured knowledge like graphs and tables, even when sufficient knowledge is provided. The research finds that hallucinations arise from systematic internal dynamics: attention disproportionately focuses on shortcut structural cues, and feed-forward representations fail to ground the provided knowledge, causing the model to revert to parametric memory. These patterns generalize across different structured knowledge formats, enabling effective hallucination detection.

This study reexamines retrieval-augmented generation (RAG) through the lens of gradient descent. It proves that a single linear self-attention layer can perform one gradient-descent step on a unified linearized RAG objective, establishing an exact equivalence between retrieval-augmented prediction and in-context optimization. Based on this insight, the authors propose a lightweight method that uses a forward-only update to optimize the evidence-use interface of frozen RAG large language models. Across seven QA benchmarks, the method improves baseline performance without modifying the retriever or backbone, approaching test-time gradient adaptation at significantly lower per-query cost.

A new framework called RICE-PO addresses credit assignment in training language agents for interactive retrieval by converting retrieval interactions into localized learning signals. It outperforms baselines on BRIGHT and BEIR benchmarks.

This paper describes The Daily Dose (TDD), an LLM-driven automated clinical summarization and trial identification system integrated into routine radiation oncology practice. A mixed-methods evaluation with 55 clinicians shows promising usability, satisfaction, and time savings.

CroCo extends contrastive preference tuning to multiple languages by leveraging an English reward model to rank self-generated responses across languages without language-specific annotations. Experiments on EuroLLM-9B and Aya-3B across 14 languages show improvements on both structured and open-ended tasks, with on-policy data being critical.

SPEAR (Sandboxed Prompt Engineer with Active Roll-back) is a free-form agentic optimizer that ports the code-as-action paradigm to automatic prompt engineering. It features four tools—evaluate, python, set_prompt, finish—and decides autonomously how to use them. The key innovation is a Python sandbox for structural error analysis on evaluation DataFrames. Two guardrails (auto-rollback and guard metric floor) ensure monotonic improvement. Evaluated on three industrial LLM-as-judge suites (13 tasks) plus 7 BBH tasks and GSM8K, SPEAR wins all industrial tasks on primary metrics and achieves 0.938 accuracy on BBH-7. Ablations show the Python tool is the largest single lever.

This paper presents the first unified survey of membership inference and data contamination under the Pretraining Data Exposure (PDE) framework, formalizing exposure levels, reviewing attack and defense methods, synthesizing empirical findings, and highlighting open challenges and future directions.

A new method called Self-Verified Distillation (SVD) enables LLMs to self-improve using only unlabeled prompts, without external feedback. The model generates candidate solutions, filters them through a three-stage verification cascade, and trains on the curated data. Experiments on Qwen3 models show significant gains across math, science, and coding benchmarks.

This paper proposes a neural rule evaluator that compiles logical constraints into directed acyclic graphs and introduces chimera training to address the scarcity of real anomaly examples. Experiments on CLEVRER, OpenImages, and VidOR show improved rule-level anomaly detection AUROC, especially for compositional and relational rules.

The paper proposes Lie group embedded dynamical neural networks (LieEDNN) that leverage adjoint action on Lie algebra to overcome incompatibility with addition and non-Euclidean dynamics, enabling stable learning on manifolds. Experiments on SE(3) for telescopic manipulators validate the approach.

Proposes PushCen-ADFL, a communication-efficient asynchronous decentralized federated learning framework that couples communication, aggregation, and local stabilization in a shared centroid representation space. It uses average-preserving push-sum mixing to correct aggregation bias and lightweight centroid regularization to mitigate drift under heterogeneity and staleness. Experiments on vision datasets show accuracy improvement up to 6% and communication cost reduction over 80%.

First work to study pretraining contamination auditing for time series foundation models (TSFMs). Proposes TSFMAudit, a method based on probe adaptation dynamics, detecting contamination via faster loss reduction and smaller backbone movement after fine-tuning probes. Evaluated on 6 TSFMs and 187 datasets, outperforming 10 baselines adapted from LLM literature.

Neural Bayesian Sequential Routing (NBSR) is a framework that models neural inference as active evidence accumulation over a hierarchical Directed Acyclic Graph (DAG). It uses a Dirichlet-Categorical conjugate framework with a global knowledge oracle to extract positive evidence vectors, and Gumbel-Softmax Straight-Through estimator for hard path-dependent routing. It provides mechanisms for uncertainty quantification, early exiting, OOD abstention, and cost-aware evidence acquisition. The paper proves monotonic precision increase and bounded variance, and demonstrates competitive performance across various tasks.

SilIF enhances Isolation Forest with a silhouette-based scoring layer by clustering path-length fingerprints. On IEEE-CIS fraud benchmark, it improves AUC-PR by 0.0080 on average; no improvement on synthetic Sparkov dataset.

AirCast-SR is a foundation model that downscales global AI weather forecasts from 0.25-degree (~28 km) resolution to 1 km horizontal resolution at hourly intervals. It uses a three-dimensional U-Net within a Latent Consistency Model diffusion framework, trained on data over the contiguous United States. The model achieves near-zero bias and preserves fine-scale atmospheric structures, validated across multiple seasons and demonstrated zero-shot transferability to India and Germany without retraining.

This paper introduces 'constraint tax,' a metric for the accuracy loss caused by structured output constraints in small language models. Experiments show that enforcing schemas like JSON increases validity but reduces answer accuracy, advocating for a 'reason free, constrain late' approach. Production systems should report multiple metrics separately.

This paper proposes GEM (Geometric Entropy Mixing), a framework that reformulates data curation as a variational problem on the hypersphere with a mixing-balance regularizer. It overcomes cluster collapse to discover balanced semantic structures invisible to Euclidean heuristics. Using teacher-student distillation for scalability and introducing the Geometric Influence Score (GIS) for interpretable taxonomy generation, GEM integrated into mixing strategies like DoReMi and RegMix improves average downstream accuracy by up to 1.2% on 1.1B-parameter models.

JobBench is a new benchmark for AI agents that evaluates them on workflows experts prioritize for delegation, aiming to empower humans rather than replace them based on GDP value.

Standard evaluations of Theory of Mind (ToM) in LLMs rely on end-point question answering, which does not reveal whether models actually construct mental-state representations. OmniToM addresses this by requiring explicit modeling of belief structures for all actors in a narrative. The benchmark comprises two stages—Belief Extraction and Belief Labeling—using a seven-dimensional schema. Built from 895 stories and 22,343 labeled belief propositions via a human-calibrated LLM-assisted pipeline, zero-shot evaluations show that current LLMs struggle with belief-tracking bottlenecks.

AI agents are beginning to complete valuable, long-horizon business operations tasks, but training and evaluation environments for enterprise work still struggle to balance realism, verifiability, and scale. Environment and task creation frequently suffers from a failure mode called artifact drift: when instructions, environments, oracles, and verifiers are created by loosely coupled processes, they frequently disagree on what a task requires, producing environments that are unsolvable, reward-hackable, or inconsistent. We introduce Anchor, a task-generation pipeline that formalizes domain experts' specifications of business workflows into constraint optimization programs. From a single parametric specification, the pipeline jointly produces a natural-language instruction, environment con…

This paper presents two novel agentic AI frameworks—DeepTS/DeepCollector and DeepScribe—that leverage a hybrid local-remote architecture to automate scientific workflows, including time-series data curation and lecture-to-report conversion, and discuss extensions to knowledge graphs and high-energy physics.

A new benchmark called AgingBench reveals that deployed AI agents degrade over time through four aging mechanisms, requiring lifespan evaluation and targeted repair rather than just stronger base models.

Constraint Acquisition (CA) and related research on Mathematical Programming (MP) model validation and enhancement are limited by inadequate benchmarks. Existing benchmarks are designed for solver evaluation, lacking domain knowledge artifacts. This work presents MPMMine, a benchmark suite guided by consistency, standardization, completeness, extensibility, openness, and version control. It uses open formats (MiniZinc, CommonMark, JSON) and provides multiple models per problem, tens of instances per model, and thousands of solutions and non-solutions in integer and continuous domains, along with natural-language descriptions.

A new arXiv paper proposes GEM (Governed Evolving Memory), reframing long-term AI agent memory as a new data-management workload with state-level operations to overcome four failure modes of current record-level systems.

A new study questions whether large language models have genuine introspective abilities, arguing that current evidence may reflect pattern matching on surface cues rather than true metacognitive monitoring.

BrickAnything is a geometry-conditioned autoregressive framework that generates physically buildable brick structures from diverse 3D representations. Using point clouds as a unified interface, it predicts brick sequences under assembly constraints and introduces structure-aware tree tokenization to model dependencies, along with preference-based alignment, validity-constrained decoding, and adaptive rollback for improved stability and geometric fidelity.