MemSlides proposes a hierarchical memory framework separating long-term memory (user profile and tool memory) from working memory, combined with scoped slide-local revision, to maintain user preferences across tasks and reliably carry out localized edits over multiple turns. Experiments show improvements in persona alignment, modification behavior, and preference carryover.

This research addresses concurrency anomalies in multi-agent LLM systems by formalizing four anomaly types using TLA+ and building a mechanically verified consistency hierarchy L0-L4. With 274 Verus proof obligations, the detectors are proven sound and complete. Three Rust runtimes implement L0-L1, and the work reproduces real-world anomalies in ByteDance's deer-flow and LangGraph, providing verified fixes.

This paper proposes an end-to-end graph neural surrogate for predicting CO2 plume migration in geological storage. Evaluated on the SPE11A benchmark, the model uses anisotropic message passing and autoregressive residual formulation to achieve competitive forecasts with low cumulative error over long horizons.

This research reveals that grokking—the abrupt delayed onset of generalization—is consistent with hysteresis in first-order L2 phase transitions in linear deep neural networks. Models trapped in low-accuracy metastable states escape only when SGD noise provides enough energy to cross barriers, with escape times following Arrhenius scaling. The number of metastable states equals the number of learnable features, offering insights for more efficient learning.

Mixture-of-Experts Multimodal Large Language Models (MoE-MLLMs) offer remarkable performance but incur high GPU memory costs, making compression essential. Existing expert-level mixed-precision quantization degrades on MoE-MLLMs due to two biases: vision tokens dominate cross-modal expert selection, and redundant tokens skew intra-vision frequency. MODE addresses this by decomposing selection frequency by modality, filtering redundant tokens, and evaluating per-modality quantization sensitivity, using integer linear programming to assign bit-widths per expert. At W3A16, average performance loss is limited to 2.9%.

This study systematically evaluates foundation model (FM) representations for computational pathology tasks using two real-world commercial cohorts, IH-BC and IH-NSCLC. Findings show that image and omics representations carry complementary predictive signals, and multimodal fusion yields gains primarily when no single modality dominates. Conformal prediction reveals that in most cases where point predictions fail, the true diagnosis remains recoverable, emphasizing the value of uncertainty-aware inference for clinical support.

Distinguishing causal adverse drug events (ADEs) from spurious correlations is a central challenge in pharmacovigilance. The InferBERT framework integrates transformer models with Do-calculus, but its success hinges on the underlying classification model. This study compares XGBoost, ALBERT, BioBERT, and Med-LLaMA on two benchmarks, finding that domain-specific pre-training (BioBERT) achieves the highest accuracy, while larger LLMs like Med-LLaMA underperform. Results indicate that investing in manageable, domain-aware models is more effective than simply scaling model size.

A study reveals that during prefill, LLMs write field-conditioned conclusions onto downstream 'notes', making KV cache editable and composable. Editing a field with chain-of-thought recovers decisions with ~1% compute, while precompiled skills can be position-portable and spliced into any context, achieving near-identical outputs to full recompute with up to 14.9x lower latency. The approach is validated across scales and cache types, achieving 98.5% prefix cache hit rate in production benchmarks.

This paper presents a diffusion-based approach for generating clinical time series that jointly models laboratory values and their observation patterns. Using the DACMI benchmark from MIMIC-III, it aligns chart times into 4-hour intervals and segments admissions into 7-day windows. The method extends TimeDiff to learn continuous lab values and discrete missingness patterns, demonstrating that generated data closely matches real patient trajectories, capturing clinically meaningful dependencies under MNAR-like missingness. This work serves as an initial component for developing clinical foundation models.

This study addresses the shape-prior shortcut problem in learning-based long-range single-shot fringe projection profilometry (FPP). Using mechanistic interpretability and conformal uncertainty quantification, the authors diagnose that UNet baselines rely on object-boundary shape priors rather than fringe-phase decoding. They propose PhiCalNet, which outputs wrapped phase and applies a differentiable calibration layer, reducing object MAE 3.3x to 4.46 mm and confirming architecture as the key factor.

A new study identifies 'editing decoupling failure' in Multimodal Large Language Models: knowledge updates succeed under multimodal inputs but revert when inputs are split into unimodal. The authors propose DECODE, which disentangles and localizes modality-specific neuron groups for consistent knowledge editing across modalities.

Researchers propose an online adaptive framework integrating treatment effect estimation, patient digital twin, and reinforcement learning for clinical decision support. The system ensures safety via rule-based monitoring and outperforms baselines on synthetic and real ovarian cancer datasets, demonstrating potential for personalized medicine.

This paper studies distributed general-purpose agent networks, proposing a layered architecture centered on a protocol adaptation layer, and identifies three core mechanism problems: semantic announcement propagation, verifiable identity and multi-topic reputation, and semantic-gradient mechanism design. Prototype overhead and simulations validate feasibility, providing a system-level foundation for open, trustworthy, and scalable agent collaboration.

SpeechDx is a large-scale benchmark spanning 12 datasets and 27 tasks for evaluating clinical speech AI. It organizes tasks by stages of speech production and tests generalization. Evaluation of 12 audio encoders shows large-scale speech models are strongest overall, but no universal representation generalizes reliably.

MemTrace is a benchmark that evaluates LLM long-term memory at the knowledge-point level along three dimensions: memory age, question type, and evidence condition. It reveals that similar pooled accuracy masks different failures, with the main bottleneck being evidence use rather than retrieval.

A new study investigates whether language models can independently discover the mathematical concept of zero. It finds that GPT-2 scale models fail at test-time generalization but improve substantially with few-shot training, and language pretraining reduces the required examples by about 50%, suggesting language abilities scaffold mathematical discovery in neural models.

In skill-constrained production-inventory systems, the qualified human capacity available tomorrow depends on training decisions made today: production requires certified workers, certifications decay unless maintained, and training consumes the same scarce worker hours that production needs now. We study a closed-loop skill-constrained model predictive controller that, at every shift, solves a finite-horizon mixed-integer program over production, inventory, backlog, and training, with binary predicted certification, hard production eligibility, and an interpretable terminal value that prices certified-capacity gaps at the horizon boundary; only the first-period action is applied before replanning. On synthetic, seed-controlled SkillChain-Gym scenarios - announced and surprise new-skill s…

Researchers propose a self-evolving framework where an LLM-based agent automatically generates query rewriting rules to enhance BM25 for legal case retrieval, without parameter training. Evaluated on LeCaRD-v2, it outperforms non-evolutionary baselines, especially with a high-capacity LLM.

This paper introduces DivInit, a training-free method that diversifies first-turn queries to overcome diminishing returns in standard parallel sampling for agentic search. It improves multi-hop QA by 5-7 points on average across models and benchmarks.

The article introduces the Agentic CDP, a new type of Customer Data Platform built for the era of AI agents. It argues that traditional CDPs are obsolete because buyers now use agents that make decisions in milliseconds, demanding speed, hyper-personalization, and richer context (Golden Context). The Agentic CDP is embedded in the data foundation, powers Infinity Campaigns (always-on, AI-driven personalization), and is designed from the ground up for agents and humans to work together. Databricks' CustomerLake is presented as an implementation.