论文库

Structured Electronic Health Records (EHR) models typically treat each clinical concept as a dataset-specific token, discarding the natural-language semantics associated with the event. As a result, relationships between concepts must largely be learned from EHR co-occurrence, without directly exploiting any prior semantic knowledge. We investigate how pretrained causal language models can be adapted for longitudinal EHR prediction by developing tEHR. Here, we represent events with their natural language description and use this text stream to apply EHR-adaptive continued pretraining, followed by task-specific discriminative fine-tuning using LoRA, utilising the LLM as a longitudinal patient-encoder. We evaluate this approach across four prediction tasks on the MIMIC-IV dataset and pancreatic cancer prediction on the CPRD dataset at 1 to 12 month horizons, with tEHR using the Mistral-7B-v0.3 backbone achieving the strongest overall performance. Ablations show that performance declines when clinical descriptions are removed or misaligned, when numerical values are omitted, and when either stage of adaptation is removed, while explicit elapsed-time descriptions provide little additional benefit. Analysis of the learned patient representations further reveals clinically coherent structure associated with distinct patterns of disease risk and clinical history.

LLM-based peer review is now used at production scale, but its evaluation remains limited. We introduce a benchmark that injects objective, localisable errors of four types (general issues, mathematics, numerics, citations) at three severity levels into real papers and measures each reviewer's detection rate by error type and severity, together with its false-positive rate on clean copies. We also present Skepthical, a multi-agent reviewer that verifies citation-backed claims by retrieving the cited papers, mathematics with computer algebra, and numerical claims with executed code. We apply the benchmark to two corpora of 15 recent arXiv preprints each, one in astrophysics and one in machine learning (ML), injecting a total of 357 errors, and evaluate twelve reviewer systems: six single-LLM reviewers, two Claude Code configurations, three publicly available reviewers, and Skepthical. The choice of model has the largest impact on detection rate: given the same prompt, three frontier LLMs differ by about 50 percentage points in total detection on both corpora. Tools and scaffolding help when they contribute what the model cannot supply on its own. In particular, on astrophysics, Claude Code improves on a single Claude call in mathematics and numerics detection, and enforced citation retrieval gives Skepthical the highest citation error detection rate of all systems by a large margin. This citation lead, however, is not observed on ML, where the models appear to know the literature better and Skepthical's citation detection rate is statistically indistinguishable from that of the best single LLM. False-positive rates are below 10% for most systems, but vary with prompt, model, and scaffolding, with mathematics flags the main source of false positives. The evaluation pipeline, the injected errors, and the adjudication logs are released as a benchmark.

Most AI assistants respond only when prompted, remaining silent even when events in the video stream offer meaningful opportunities for conversation. To move beyond purely user-driven interaction, we present AURA 2.0, a streaming video large language model that introduces three core upgrades to AURA. The key advance is Model-Initiated Conversation (MIC), which enables the model to identify topics worth discussing in the evolving visual stream and initiate conversations at appropriate moments without a user query. Stream-Aware Tool Calling (SATC) enables the model to invoke external tools using the current visual and conversational context and incorporate returned information into the ongoing interaction. Interactive Visual Grounding (IVG) supports visual grounding in both user inputs and model outputs. The model can identify objects marked by users on video frames and produce frame-aligned bounding boxes to link its responses to specific objects and regions. To develop these capabilities, we construct a dedicated streaming interaction dataset and adopt a three-stage training scheme to integrate the new behaviors with AURA's existing streaming interaction abilities. We further incorporate Long-Horizon Interaction Memory (LHIM) to maintain contextual continuity over extended video streams. Evaluations on public streaming benchmarks show that AURA 2.0 achieves leading performance among open-source streaming models, while targeted evaluations further validate its new interaction capabilities. We open-source the model weights and inference code.

Perceptual quality enhancement of severely compressed videos remains challenging due to complex artifact patterns and substantial information loss. Recent diffusion models have demonstrated strong generative capability for visual restoration, but directly applying them to compressed video often ignores compression degradation characteristics and may introduce structure-inconsistent hallucinations. To address this issue, this paper presents a diffusion-based compressed video enhancement method, named DiffCVE. Coding Prior-enhanced Dual Conditioning (CPDC) branches are designed to jointly model compressed video and coding prior conditions, where coding priors including residuals and motion vectors provide complementary structural and motion guidance during the diffusion denoising process. To make the diffusion process aware of compression severity, a Compression Degradation Semantic Prompting (CDSP) mechanism is introduced to leverage QP-conditioned textual prompts together with LoRA fine-tuning. In addition, a Coding Prior-guided Weighted Fusion (CPWF) module is incorporated into the VAE decoder to fuse VAE encoder and coding prior encoder features with QP-predicted weights. Extensive experiments demonstrate the effectiveness of the proposed method in improving perceptual quality, especially under severe compression settings. The project page and code are available at https://wqmaker.github.io/projects/DiffCVE/.

Users of artificial intelligence services, including those powered by large language models and other transformer-based models, need assurance that inference and training follow the declared computation without exposing private models or data. Zero-knowledge proofs can provide this assurance, but proof generation remains expensive. Hardware acceleration research often reports gains for individual cryptographic operations, leaving their effect on the full cost of proving neural-network workloads, particularly large transformer models, unclear. We introduce a four-layer framework linking verifiable AI workloads, proof systems, cryptographic kernels and hardware platforms, use it to review the literature, and add original measurements of six provers and seven kernels to identify which costs matter in practice. Our measurements show that the dominant cost changes with the proof system and input length. In low-rank fine-tuning, the cost of opening commitments to fixed model weights remains nearly constant across input lengths from 64 to 2048, while the work associated with intermediate values grows. A 13-billion-parameter fine-tuning configuration exceeds the capacity of a 40-gigabyte graphics processor but completes on an 80-gigabyte device, showing why capacity must be assessed alongside execution time. Our kernel profiling further identifies instruction issue and synchronization limits that a bandwidth-only model misses. We use these findings to assess graphics processors, reconfigurable logic, custom chips and existing artificial intelligence accelerators. The resulting research agenda connects proof-friendly numeric formats, reuse of fixed-weight commitments, heterogeneous execution and shared benchmarks to the measured bottlenecks. Effective acceleration requires selecting hardware and proof-system optimizations around the cost profile of the target workload.

Multi-agent deep-research systems use delegation to parallelize search and keep long tool traces out of the main agent's context. Yet delegation helps only when task allocation and subtask execution are coordinated: the main agent must issue useful subtasks, while subagents must return reliable evidence. Existing training either optimizes only the main agent or broadcasts the same team advantage to all agents, which can reward misleading subagent reports and penalize useful ones. We introduce LEAD-RL (Learning Effective Agent Delegation with Reinforcement Learning), a joint main-agent-subagent training framework with contribution-aware credit assignment. A training-time rubric evaluates the research utility of each subagent report and gates its inherited team advantage, while final-answer correctness remains the only team reward; no token, step, or speedup reward is used. Across five benchmarks and two model families, LEAD-RL improves answer quality. On BrowseComp, it improves accuracy by 2.49 percentage points and reduces critical-path cost by 24.5% over joint training with shared team credit. During validation, accuracy increases by 6.78 points at approximately constant critical-path cost. These results show that contribution-aware credit assignment improves the quality-efficiency trade-off of multi-agent deep research.

Watermark detectors decide whether one document came from a watermarked language model. Auditors often need a different number: the share of a corpus that did. We study this share for Gumbel-max and green-list (KGW) watermarks when every document may have been edited by an unknown amount. Without restrictions on the edits, the upper end of the identified set is 1. Combining a known least-favourable next-token distribution with a sequential coupling, we derive a power envelope indexed by three budgets: a retention floor, an entropy budget, and a budget for in-place substitutions, which for Gumbel-max make edited positions look less watermarked than human text. The envelope gives an identified set that is sharp for the alternatives it dominates, finite-sample bounds on the realized share of a fixed corpus with either the exact null or a null calibrated on an independent human sample, and confidence sequences for streams. In Gumbel-max simulations where the envelope binds, the bounds cover, and the threshold that maximizes the gap between envelope and null is generally as tight as twelve thresholds. On text from OPT-1.3B and Qwen3-1.7B, 12,000 fresh human documents scored under three keys show no robust departure from the exact null, and all bounds covered in resampling trials from finite pools of documents with random edits, substitutions and paraphrases; at a share of 0.2 the exact-null bound is about [0.14,0.38]. These results are fragile in two ways we make precise. LLM-paraphrased KGW pools fall below the envelope, yet at 2000 documents the bounds still covered in these trials. A simulated null error of 0.01, comparable to the largest deviation we observed at our thresholds, is likewise harmless at 2000 documents but breaks exact-null coverage at 100,000; a null calibrated on an independent human sample keeps coverage there, but needs a large reference sample to stay narrow.

Electroencephalography (EEG) foundation models are pretrained on thousands of hours of recordings, yet their frozen features transfer poorly to downstream tasks, and this transfer does not consistently improve with larger models or more pretraining data. We show that a vision foundation model that has never seen EEG is a stronger frozen EEG encoder than these specialized models, once the signals are rendered as images that a vision transformer can read. Our method, SeeEEG, keeps the vision backbone frozen and learns only how EEG is displayed to it. Each channel occupies a dedicated row of patches and is drawn either as a waveform at a fixed gain or as a scalogram whose frequency range spans the patch height, so that task-relevant patterns fall within individual tokens. The montage, filtering, and contrast parameters of this display are initialized from a standard clinical montage and optimized with the task loss through the frozen backbone, which amounts to at most 1,113 trainable parameters per view. We compare SeeEEG with seven EEG foundation models on eight clinical, sleep, affective, and brain-computer interface datasets, with every encoder frozen and paired with identical readout heads. With a frozen DINOv3 ViT-B, SeeEEG outperforms every EEG foundation model in balanced accuracy on all eight datasets and reaches a mean rank of 1.3 in Cohen kappa or AUROC, compared with 2.9 for the strongest baseline, REVE. The learned displays transfer without further optimization to the video encoder V-JEPA 2.1, which outperforms every EEG foundation model on both metrics on six of the eight datasets. Replacing DINOv3 with a randomly initialized ViT-B lowers the score by 15.2 points on average, so the ability to read rendered EEG comes from visual pretraining.

High-granularity topic modelling extracts hundreds of niche topics from a corpus, yet this regime is rarely studied and is more commonly framed as a failure mode: "over-clustering," in which coherent themes fragment into redundant subtopics. We argue that this framing conflates genuine redundancy with topics that are semantically similar but meaningfully distinct, and that the central challenge in high-granularity modelling is separating the two. Recent post-hoc refinement methods address topic redundancy through either embedding-based similarity thresholding or LLM-guided merging. We show that neither suffices in high-granularity settings and propose embedding-gated iterative merge (EGIM), a hybrid approach that iteratively generates context-rich topic representations, uses embedding similarity to propose candidate merges across the topic space, then applies LLM judgement to accept only genuinely redundant pairs while preserving niche distinctions. Across two patent corpora and one legal corpus, LLM judgement rejects 83% of embedding-identified candidates as non-redundant. EGIM improves discriminability on merged topics by 25-55% over a prior iterative LLM baseline while better preserving the original topic distribution.

The ability of deep neural networks to learn hierarchical features is widely regarded as a key mechanism underlying their success in high-dimensional learning. Existing theory partially supports this view by establishing approximation rates based on parameter counts and sample complexity guarantees for compositional models without incurring the curse of dimensionality (CoD). To study overparameterized regimes, where the number of parameters exceeds the sample size, we develop a framework that measures complexity via the parameter norm. Within this approach, we establish approximation rates and excess risk bounds for learning sparse compositional functions whose compositional structure is represented by directed acyclic graphs (DAGs), using Frobenius norm-constrained deep neural networks. Our results have broad applicability since every function that is efficiently Turing computable admits sparse compositional representations. In particular, we cover a range of representative models, including multi-index models, binary tree structures, and general compositional architectures. The rates we derive show that deep networks can exploit the compositional structure of the target functions, effectively avoiding the CoD through hierarchical representations.

Chess board recognition is the task of mapping the image of a chess board to the information of which piece is on which square. So far this task has two established benchmarks: ChessCog is synthetic, and ChessReD comes from smartphone pictures of a single chess board setup. We introduce ChessQueries, a new method combining a ViT encoder with a DETR-style decoder, which outperforms existing methods. On the ChessReD benchmark, we improve the state of the art from 15.3% to 99.2%, and demonstrate strong capabilities on out-of-distribution datasets. Our method saturates the task on the two datasets, with an average 0.01 wrong squares per board (vs. SotA: 3.4 / 0.15 respectively). We also share a new, harder public dataset, parsed from broadcasted top-level chess tournaments. Code, model weights and the SLCC data have been released.

Large language models (LLMs) can supply autonomous systems with plans, predictions, and judgments, but these contributions are sometimes wrong. This review asks how a system can maximize an LLM's utility while managing its fallibility. Our answer is to adapt the system's reliance on the LLM as evidence about the LLM's performance accumulates. Systems built from imperfect components, such as noisy sensors and faulty software, have long worked in this way, and the classical methods they use can be applied to LLM-based systems as well. A system can change its reliance on an LLM in several ways, depending on the LLM's role and on the other components in the system. Each change needs an estimate of the LLM's reliability, and the LLM's role and the kind of contribution it makes determine what to estimate. Classical methods from decision theory, estimation, and online learning then turn these estimates into reliance decisions. These decisions can reduce how often LLM errors affect the system. However, some errors matter even when they are rare. An independently verified mechanism can prevent such errors from being executed. Without one, adapting reliance using reliability estimates can still keep their risk within a bound. Studies on adapting reliance on LLMs are emerging but limited, and the reviewed comparisons suggest that adapting reliance can improve system performance.

Layer-local self-supervised objectives remove the backward pass between layers, but almost all still compute their loss over a batch of samples and are therefore local in space but not in time. We introduce Temporal Regularized Learning (TeReL), a spatially and temporally local learning rule for self-supervised representation learning from sequential input streams. The key idea is to instantiate a VICReg-style objective with temporally adjacent positives and non-differentiated running population statistics, yielding a per-neuron online loss with bounded state. The objective is a soft form of Slow Feature Analysis, and its gradient is a Hebbian update of the feedforward weights and an anti-Hebbian update of one lateral matrix per layer. The aim of the experiments is to show that the rule works in a small network, not to compete with backpropagation. On a class-chunked MNIST stream, a two-layer fully connected network trained with batch size one outperforms random features (96.48% versus 95.09% linear-probe accuracy) and classical SFA, and a variant that relaxes both locality constraints is comparable to supervised backpropagation. Because training and inference use the same computation, the network can keep learning from unlabeled data during inference without loss of accuracy.

Federated learning (FL) is a promising framework for learning from distributed data while maintaining privacy. The development of efficient FL algorithms encounters various challenges, including heterogeneous data and systems, limited communication capacities, and constrained local computational resources. Recently developed FedADMM methods show great resilience to both data and system heterogeneity. However, they still suffer from performance deterioration if the hyperparameters are not carefully tuned. To address this issue, we propose an inexact and self-adaptive FedADMM algorithm, termed FedADMM-InSa. First, we design an inexactness criterion for the clients' local updates to eliminate the need for empirically setting the local training accuracy. This inexactness criterion can be assessed by each client independently based on its unique condition, thereby reducing the local computational cost and mitigating the undesirable straggle effect. The convergence of the resulting inexact ADMM is proved under the assumption of strongly convex loss functions. Additionally, we present a self-adaptive scheme that dynamically adjusts each client's penalty parameter, enhancing algorithm robustness by mitigating the need for empirical penalty parameter choices for each client. Extensive numerical experiments on both synthetic and real-world datasets are conducted. As validated by some numerical tests, our proposed algorithm can reduce the clients' local computational load significantly and also accelerate the learning process compared to the vanilla FedADMM. The source code is available at https://github.com/DCN-FAU-AvH/FedADMM-InSa.

Artificial intelligence (AI) is changing how biological research is done. Biological foundation models trained on millions of nucleotides, protein, and single-cell measurements predict protein structures, mutation effects, and cellular responses to genetic perturbations. Agentic systems, AI assistants built on large language models, can run multistep analyses from plain-language requests and increasingly call foundation models as tools. Yet it is often unclear when these models' outputs can be trusted. We frame this as a model's knowledge boundary: the range of inputs, tasks, and conditions over which its outputs are reliable. This review explains how foundation models are built and proposes that five design elements (data, architecture, training scheme, evaluation scheme, and post-analysis) define this boundary. We examine evidence on where current models succeed or fail and present standardized descriptions of model limits so that researchers, and the agents acting for them, can use these tools reliably.

In July 2026, the European Commission (EC) set out the anonymization measures Alphabet must implement to share Google Search data with competing search engines under Article 6(11) of the Digital Markets Act. As the first concrete illustration of how the EC operationalizes anonymization under the EU digital rulebook, this methodology could become a regulatory benchmark. In this paper, we show why it should not. We critically review the anonymization measures proposed by the EC, from the characterization of residual risk to the suppression thresholds and the reliance on contractual safeguards. For each of those issues, we highlight significant conceptual flaws, and, where applicable, we illustrate these flaws with empirical observations from real Google Search data provided by Alphabet and protected under the EC's measures. Importantly, we present examples of low-effort reidentification, record linkage, and confidential attribute inference achieved through straightforward and automatable attack patterns. Our findings show that the approach of the EC relies on unjustified heuristics rather than formal privacy guarantees, underestimates the risks inherent in free-text search queries, and sacrifices data utility without achieving meaningful protection. We also demonstrate that the proposed measures do not meet any of the three anonymization criteria established by the European Data Protection Board (EDPB): no record isolation, no linkage, and no inference. Practitioners and regulators are therefore cautioned against treating these measures as a sound anonymization practice.

Decision making often requires balancing exploitation of familiar options with exploration of uncertain alternatives. One strategy for exploration is random exploration, in which choices become more variable and less tied to the option that currently appears best. However, it remains unclear whether this variability reflects genuine stochasticity or instead a deterministic structure that traditional cognitive models fail to capture. We examined this question using the Horizon Task in an aggregated dataset of 1,392 participants completing over 1,270,695 trials. In this task, exploration becomes more valuable when participants have more future choices, and choice variability increases under longer horizons. We tested whether more expressive neural network models could recover deterministic structure from this additional variability. Although neural networks provided better overall fits to behavior than the traditional model, long-horizon choices remained consistently less predictable than short-horizon choices. These findings show that limited model expressiveness cannot fully account for the increased variability associated with random exploration, suggesting random exploration truly is substantially random.

Large language model (LLM) agents use external skills to reuse task procedures and experience with tools. General-purpose skills describe shared execution procedures, while specific tool choices, operating practices, and output requirements vary across users. In long-horizon tasks, subgoals, available inputs, and environmental states change as execution progresses. Agents must select currently applicable skills from an expanding skill pool and prepare candidates for subsequent steps. We propose Skill Adapter, which enables personalized modification and online routing of general-purpose skills. The system distills local adjustments from user requirements and execution feedback and stores them as Adapters linked to general-purpose skills. The routing agent selects general-purpose skills and applicable Adapters based on the current subtask, execution progress, and environmental feedback, incrementally maintaining a task-level online skill library that the main LLM reads on demand. The system preloads candidates for subsequent nodes and rechecks their applicability once actual inputs are determined. Reinforcement learning optimizes the two types of decisions using modification gains and routing outcomes, respectively. We evaluate Skill Adapter on ALFWorld, WebShop, and SkillsBench. Compared with SkillRL, it improves task success rates by an average of 11.43 percentage points and reduces mean execution time on all three benchmarks. Skill Adapter retains its performance advantage when the skill pool grows to 1,000 skills.

Federated Learning (FL) is a distributed learning paradigm that enables multiple clients to collaborate on building a machine learning model without sharing their private data. Although FL is considered privacy-preserved by design, recent data reconstruction attacks demonstrate that an attacker can recover clients' training data based on the parameters shared in FL. However, most existing methods fail to attack the most widely used horizontal Federated Averaging (FedAvg) scenario, where clients share model parameters after multiple local training steps. To tackle this issue, we propose an interpolation-based approximation method, which makes attacking FedAvg scenarios feasible by generating the intermediate model updates of the clients' local training processes. Then, we design a layer-wise weighted loss function to improve the data quality of reconstruction. We assign different weights to model updates in different layers concerning the neural network structure, with the weights tuned by Bayesian optimization. Finally, experimental results validate the superiority of our proposed approximate and weighted attack (AWA) method over the other state-of-the-art methods, as demonstrated by the substantial improvement in different evaluation metrics for image data reconstructions. The source code is available at https://github.com/DCN-FAU-AvH/FL-AWA-attack.

Nonlinear time-optimal model predictive control (TOMPC) enables highly performant control of constrained systems, but its computational cost prohibits real-time use on fast dynamical systems. Neural network imitation of the TOMPC control law can reduce online computational costs, yet existing work mostly targets quadratic cost formulations in simulation and addresses the challenges of neural controllers in isolation. This work imitates a long-horizon nonconvex TOMPC with hard terminal constraints, whose high computational cost limits training-data generation. We present (i) \emph{self-informed sampling (SIFS)}, a data-efficient expert-sampling strategy that uses a learned time-to-go to query the expert only where the policy deviates from time-optimal behavior, (ii) a complete training and control architecture that jointly addresses dataset correlation, suboptimal expert labels, covariate shift, constraint satisfaction and model mismatch, and (iii) real-world validation on a Furuta pendulum. SIFS matches standard dataset aggregation performance in roughly one-third of the training time and with less than half the training data. The resulting controller matches the region of attraction of the expert TOMPC while raising the control frequency from 1.4~Hz to 1~kHz. On hardware, the controller swings up and stabilizes the pendulum along trajectories closely matching the expert TOMPC solution.

Robot agents built on vision-language models depend on both the planner and its harness: the tools, feedback, memory, and recovery logic around it. Can improvements to a large teacher's harness be carried into a compact agent with lower inference compute cost? We present TinyAstral, a two-stage approach. We first freeze an open-weight teacher and improve its harness through recursive self-improvement (RSI): frontier LLM agents read the teacher's trajectories and propose harness edits, and each edit is accepted only if it passes a paired statistical test. We then freeze the final harness, collect teacher trajectories with reasoning and failure recovery, and distill a small student that replaces the teacher in the same harness. On RoboTwin, harness evolution raises the teacher's success from 33.4% to 63.6%; the final 4B student reaches 62.8%. Across four harness versions, distilled 4B students score close to their corresponding teachers, and later versions yield stronger students, showing that harness evolution gains carry into compact models. The harness gains also transfer to other frozen planners without retraining. On identical episodes, the final 4B student outperforms open 4B agents, the open 27B harness-RSI model SenseNova-RoboRSI, and our own teacher model placed in three competitors' harnesses. It stays close to its teacher on all three training benchmarks, and on three benchmarks never used for RSI or training it scores close to a 552B planner running the same harness. On the measured deployment episodes, it uses about one fifth of its teacher's accelerator time. Project page: https://garfield-kh.github.io/TinyAstral/

Recent advances in neural density estimation have enabled amortized Bayesian inference for complex stochastic simulators. However, these methods rely on simulators accurately reflecting the true data-generating process and can degrade significantly under model misspecification. We consider the setting where multiple unlabeled observations are available and introduce robust variational neural posterior estimation (RVNP), an amortized Bayesian inference method that uses an importance-weighted autoencoder to jointly learn a misspecification-robust posterior proxy and an explicit error model that captures the simulation-to-reality gap. Our results show that RVNP can recover robust posterior inference in a data-driven and interpretable manner, outperforming previous methods on multiple metrics across misspecified benchmarks.

Coding agents have demonstrated strong capabilities in software development, yet their ability to tackle long-horizon engineering tasks governed by physical constraints remains less understood. We study this question through analog circuit layout, which demands domain knowledge, geometric reasoning, and iterative interaction with design and verification tools. We introduce \textbf{AnalogLayoutWorld}, an open-source agent benchmark with 40 expert-reviewed tasks for end-to-end layout generation from fixed transistor-level netlists. It combines an integrated design environment with independent physical and post-layout electrical verification, as well as objective layout quality metrics co-designed with analog layout experts. Across 13 model--harness configurations, 91.1\% of attempts produce GDS layouts, but only 23.0\% satisfy the benchmark's acceptance criteria. We also evaluate Novara, an internally developed agent system, which achieves the highest reported aggregate score. Our analysis links performance differences to test-time scaling under fixed budgets, suggesting that reasoning efficiency and effective use of verification feedback help sustain design quality as task complexity grows.