The study investigates the appropriate authority for a reviewer in multi-agent language-model systems. It compares broad override authority with an evidence-gated approach that requires a mechanically valid record of constraint violation for replacement. Testing across various models and tasks showed that the evidence gate effectively reduces corruption of correct proposals while maintaining most violation repairs. However, the outcome is not universally better, as reviewer competence and interface design also affect the correction domain. Ultimately, reviewer competence, action interface, and binding jurisdiction jointly determine the scope of correction and the surface of corruption.
Execution harnesses possess vital state information, yet models often lack awareness of its health or suitability during action selection, creating operational blindness. Agent Harness Awareness introduces a design where the host provides a verified state snapshot to the agent before consequential actions. Testing with ToolRoute, the approach significantly reduced policy regret by 96.9% and improved single-dispatch completion rates. Further studies showed that verified state reduced mean regret to zero in live execution scenarios. These findings demonstrate that incorporating verified telemetry materially improves the agent's pre-dispatch operational choices.
Agentic AI systems exhibit unknown, dynamic resource demands because their execution paths depend on stochastic token sampling and observations. This unpredictability creates a dilemma for infrastructure orchestrators, as under-provisioning causes throttling and over-provisioning wastes capacity. The paper formalizes this as an uncertainty-aware orchestration problem aiming to allocate resources based on actual pipeline needs with probabilistic guarantees. Conformal Reservation for Orchestration (CRO) is proposed as an online algorithm that uses past execution traces to provide these allocations. Simulations show CRO effectively manages risk while minimizing resource over-reservation compared to existing methods.
Existing systems for medical multi-agent reasoning fail because they select experts and retrieve knowledge based only on the initial question, ignoring emergent reasoning gaps. The proposed method introduces a CHALLENGER to inspect the reasoning trace and identify specific factual or expertise gaps. These identified challenges trigger targeted knowledge retrieval or the recruitment of specialist agents during the discussion. This approach significantly improves accuracy on medical QA benchmarks by timing and signaling when to seek external resources. Ultimately, the study demonstrates that the timing and context of external resource recruitment are crucial design choices for effective multi-agent medical reasoning.
Tiered LLM inference distributes computation across various hardware tiers to achieve lower latency and privacy. The work derives a closed-form condition to determine when a tiered cascade is feasible compared to a single backstop query. This feasibility condition is bracketed by a signal-free floor and a perfect-signal ceiling, dependent on an architecture cost ratio $\kappa$. The condition can be inverted based on billing models to show when local compute cost binds or is sunk. The paper proposes three design requirements that existing primitives satisfy in isolation, and suggests future work on a co-designed architecture.
Agents require the ability to reuse and update experience as they operate in evolving environments. ReScopeBench is introduced to benchmark the experience lifecycle across 18 interactive environments, distinguishing between structural and rule experience. The research proposes an experience-first construction approach to study how experience is acquired, reused, and revised. Evaluation of 20 models shows that retained experience aids performance, but agents struggle to keep it current with environmental changes. This highlights the complementary challenges of selective maintenance and effective use for long-lived agents.
Large language models were evaluated across four stages of scientific literature review: finding, interpreting, synthesizing, and proposing future research using a 102-article benchmark. Six AI tools were tested against a corpus related to the Spanish plume weather pattern. While some tools correctly identified key articles, their syntheses often contained errors or limited insights when evaluated by domain experts. No single LLM proved superior across all review tasks. The findings suggest LLMs can assist with discovery and interpretation but cannot replace expert judgment in scientific evaluation.
SlackBench is a benchmark designed to evaluate AI agents on complex, real-world collaborative projects. It tests agents' ability to reason over workplace communications and evolving code repositories, which requires synthesizing information across multiple sources. The benchmark is grounded in privacy-preserving reenactments of academic research projects featuring multi-party dialogues and code changes. It includes 140 queries focused on tasks like agentic search, goal-directed summarization, and context inference from conversations and code. Performance evaluation on this benchmark showed that the best systems scored around 80%, indicating significant room for future improvement.
MA-SafeDiffuser extends SafeDiffuser to multi-agent planning using diffusion probabilistic models with finite-time diffusion invariance guarantees for joint safety. It formalizes joint safe sets as intersections of individual and pairwise barrier sets to derive constrained denoising procedures. The method addresses local-trap and deadlock issues by incorporating time-varying specifications and liveness constraints. Empirically, MA-SafeDiffuser reduces safety violations compared to unconstrained diffusion baselines while maintaining planning quality. The work includes algorithms, proofs, and a benchmarking suite for multi-agent environments.
Spectral graph generation uses the eigenvalues and eigenvectors of the Laplacian to represent graphs. Existing methods preserve eigenvector orthonormality but fail to enforce the Laplacian nullspace constraint. Laplacian-Constrained Stiefel Flow Matching (LaSFM) is introduced to preserve this nullspace condition along with other spectral properties during generation. The method is extended for property-guided generation, steering samples toward desired graph characteristics. Experiments show that LaSFM improves performance on spectral and graph-level metrics.
The Implicit Neural Operator Solver (INOS) learns the implicit solution and encodes a time integrator for high-fidelity dynamic simulations. INOS pretrains a neural operator from static solutions and uses it within Newmark-$\beta$ integration to advance dynamic states. This method achieves accurate full-field predictions in elastodynamics without requiring dynamic trajectory supervision. INOS provides over 4,600 times speedup in nonlinear simulations compared to sequential finite-element solvers. Consequently, INOS is presented as a reusable neural implicit solver for efficient and robust dynamic modeling.
Quantum machine learning models, specifically QSVM and VQC, were benchmarked against five classical models across six public clinical datasets. The comparison assessed differences using equivalence tests, finding that neither quantum model significantly outperformed the classical comparators. Results indicated that quantum methods generally approached but did not exceed the performance of tuned classical models. The study introduces the QML-Health-READY framework to establish reporting requirements for credible claims in QML applications. Ultimately, practical costs like finite-shot estimation were identified as the main hurdle rather than gate noise.
Aligning large language models with multiple objectives, like helpfulness and harmlessness, often requires costly preference judgments for each objective. Objective-Conditioned Active Preference Learning (OGC-DPO) is introduced as an active learning framework to reduce this cost. It selects preference comparisons that are expected to most influence the policy for each objective by measuring policy gradients. This method outperforms existing active preference learning techniques when aligning models across multiple objectives. Furthermore, informative conflicts among preferences are shown to provide valuable signals for policy training.
Textual information in time series datasets holds predictive signals that numerical history alone cannot reveal. This work proposes Context as Covariates (CoCo), a framework to bridge the gap between Large Language Models (LLMs) and Time Series Foundation Models (TSFMs). CoCo uses an LLM to generate numerically grounded covariates that guide a TSFM backbone. The process involves aligning the LLM and TSFM through staged fine-tuning using reward signals based on forecasting performance. Experiments demonstrate that this approach yields up to 12% MAE improvement over existing models on diverse benchmarks.
Kernel measures are used in machine learning to tune hyperparameters and select kernels, often by requiring kernels to capture all samples from the same class. The proposed "Confidence Score" is a new kernel measure designed to address the restrictive nature of existing measures. Instead of aligning to the ideal kernel, the Confidence Score measures the alignment between ground truth and its kernel projection. Experiments show that the Confidence Score has a higher correlation with classification accuracy than other kernel measures.
Visual Language Models can be recast as EEG foundation models for clinical neurodiagnostics. This approach aims to leverage generalist VLMs to interpret EEG signals in a manner similar to human clinicians. The research demonstrates that generalist VLMs perform competitively across clinical neuro-diagnostic tasks without extensive EEG pretraining. These models also show superior few-shot learning adaptability compared to specialist foundation models. Ultimately, VLMs can effectively learn to read and interpret EEG patterns for clinical applications.
Decentralized federated learning is vulnerable to data heterogeneity, Byzantine attacks, and channel noise, which accumulate during model propagation. FedSharp proposes a robust framework incorporating screened, hop-aware, ranked, proximal averaging to address these issues. This method ensures at least $b$ honest candidates and reduces channel-noise variance by a factor of $b$ through averaging. The framework also uses noise-adaptive regularization to mitigate accumulated noise and non-IID drift. Experiments demonstrate FedSharp achieves high accuracy on MNIST and CIFAR-10 even under Byzantine and noisy conditions.
The research investigates whether learned fluid representations organize their latent spaces around actual physical quantities. Experiments on 2D decaying Navier-Stokes flow tested three pre-registered studies to assess this. Findings indicated that joint-embedding predictive architectures and autoencoders exhibit varying degrees of organizing physical structure in their latent spaces. Attempts to steer these representations through learned directions yielded results where controlling spectral shape required more complex, effect-based rules than simple steering. Ultimately, the study concludes that what a latent contains and what can be controlled through it are not directly aligned with physical structure.
Multivariate conformal prediction often results in complex prediction sets that are hard to interpret. Generalized hyperrectangles provide a practical alternative by offering one interval per output. The work introduces a rectangularization framework to transform continuous nonconformity scores or conditional densities into these hyperrectangular sets with finite-sample coverage. Key geometric properties are established for this framework, and a representation theorem links hyperrectangular methods to rectangularizations of conditional densities. Experiments quantify the efficiency cost of this rectangularization, highlighting its utility in uncertainty quantification.
The paper introduces a unified framework using Gibbs probability measures to handle training-dataset aggregation and exact machine unlearning. It studies operations on these measures, such as renormalization, normalized log-linear combination, and nesting, showing they yield measures with explicit objective functions. Nesting is shown to be equivalent to normalized log-linear combinations, establishing a geometric interpretation involving Pythagorean identities. These operations are applied to Gibbs algorithms to aggregate local datasets, enabling exact machine unlearning and centralized-equivalent performance in distributed learning systems.
The study addresses sample complexity for multi-objective reinforcement learning where preferences are specified post-data collection. It focuses on finite discounted Markov decision processes with unknown transitions and $m$ known reward functions. The goal is for the learner to provide an $\varepsilon$-optimal policy for any convex combination of rewards without additional sampling. A minimax sample complexity bound of $\widetilde\Theta\left(\frac{SAH^3}{\varepsilon^2}\min\{S,\log(em)\} \right)$ is established. The analysis demonstrates that the worst-case statistical cost grows logarithmically with the number of objectives and saturates at the number of states.
Extreme events often occur sequentially, motivating the use of generative modeling for cascading phenomena like earthquake or financial market losses. The method introduces an extremal token representation that encodes multivariate observations as a categorical sequence, focusing on dominant exceedances. A decoder-only transformer estimates the history-dependent next-token kernel, enabling recursive sampling to generate future sequences. The model is pretrained across simulated process families to learn which dimensions dominate multivariate extremes. This approach supports task-specific tokenization and simulation-based pretraining for modeling multivariate extremes.
The paper studies non-parametric bandits where arm evaluations use Cumulative Prospect Theory (CPT) instead of expected reward. It proposes Non-Parametric CPT Bandits ($\mathtt{NPCB}$), an optimistic algorithm that separates mean estimation from CPT-based decisions in continuous action spaces with arm-dependent noise. The CPT regret is shown to be of order $\widetilde{\mathcal O}\bigl(\sqrt{T\,\dim_E(\mathcal F,T^{-2})\,\log N(\mathcal F,T^{-2},\|\cdot\|_\infty)}\bigr)$, matching classical non-parametric bandit regret. The work extends $\mathtt{NPCB}$ to unknown noise CDFs and analyzes Kernel-CPT bandits, yielding information-gain-based regret bounds.
TestBED is an open-source toolkit designed for Bayesian experimental design, aiming to standardize this framework. It decomposes the BED process into modular components covering tasks, objectives, training, and evaluation procedures. The toolkit provides scalable implementations of core primitives like expected information gain estimators. It allows users to compose existing methods and easily implement and evaluate new methodologies. TestBED also facilitates large-scale benchmarking of contemporary BED methods.
Current generative AI alignment relies on a myopic paradigm prioritizing immediate conversational satisfaction. This approach ignores the long-term consequences of human-AI interaction, such as cognitive atrophy and diminished human agency. The authors propose shifting to a long-horizon alignment strategy that considers how AI alters human experience over time. This involves mapping the trade-offs between near-term utility and long-term costs across four facets of human well-being. The goal is to move beyond simple preference optimization to foster long-term human well-being.
The paper establishes tight lower bounds for minimizing high-order Hölder-smooth and uniformly convex functions when their smoothness orders and convexity degrees are matched. Specifically, for functions where the $p^{th}$-order derivatives are Hölder continuous with degree $\nu$ and the functions are uniformly convex with degree $q$, the lower bound involves $\Omega\left(\left(\frac{H}{\sigma}\right)^{\frac{2}{3(p+\nu)-2}}\log\frac{\Delta_0}{\epsilon}\right)$ oracle calls. This result aligns with the upper bound, providing a higher-order counterpart to Nesterov's classical first-order lower bound. The authors develop a multiscale phase-wise construction to achieve this, preserving previous oracle responses while maintaining the required smoothness and convexity. This work resolves an open case regarding the lower bounds for the two asymmetric regimes of such functions.
Whole-page ranking optimizes ads and content by balancing business goals and user engagement under complex constraints. Existing methods struggle to capture order-dependent interactions efficiently due to either independent scoring or high autoregressive inference overhead. The proposed Multi-Item Scoring (MIS) framework uses a prefix-aware approach to integrate sequential context into efficient multi-item scoring. MIS computes reusable representations once and performs lightweight scoring during beam search to capture dependencies without repeated full-model inference. Experiments show MIS matches autoregressive quality while achieving significantly faster CPU decoding and reduced service consumption.
Support Vector Machines are shown to be equivalent to minimizing the Kullback-Leibler (KL) divergence between two Gaussian measures. This research demonstrates that the dual objectives of SVM classification and regression are exact KL divergence minimizations between Gaussian measures. The KL divergence formulation also recovers the objective function for kernel ridge regression when domain constraints are lifted and $L^2$ regularization is applied. Consequently, the work establishes a unified information-theoretic model for standard kernel machines. This provides a foundation for unifying broader kernel-based algorithms using probability theory.