AI agents have automated the formalization of classical linear regularization theory and functional analysis within a Lean library, called AutoIP. Machine-checked proofs only certify formal statements, shifting the human task to confirm that these statements accurately reflect intent. Risks are identified in the formal statements themselves, rather than the proofs. The authors suggest using automated tests on these statements to expose potential issues. The work invites the inverse problems community to extend and build upon this formalized library.
Liability adjudication tasks require analyzing event descriptions against domain-specific skills to determine liability. Existing methods struggle with stale skill libraries, semantic execution gaps, and context budget limitations. KDASO is proposed to address these issues by continuously evolving skill libraries using recent data and LLM trajectory analysis. It employs context-budgeted routing for scalable skill selection. Experiments show KDASO significantly outperforms baselines in various adjudication tasks, demonstrating high skill transferability.
BlockLock addresses reproducible LLM inference by fixing the batch geometry for protected requests, maintaining consistency across batch variations. It introduces two modes, BlockLock-Rows and BlockLock-Pack, to achieve this by either using fixed-width rows or terminal-anchored packing with isolated dummy segments. Both modes ensure the protected request's output bits remain unchanged even when companions in the batch are modified. Evaluation across five dense models and three GPU setups demonstrated that BlockLock reduces latency and improves throughput compared to batch-invariant methods. Specifically, BlockLock-Rows showed latency reductions of 53.5-55.0\% and throughput improvements of 115.2-122.2\%.
Generating training responses in mixture-of-experts models is a major computational bottleneck during reinforcement learning post-training. FlexFloor is a system designed to improve this efficiency by enabling earlier resource release during synchronous response generation. It achieves this by jointly adapting response scheduling and the active group of workers using historical response lengths and memory constraints. This method maintains expert availability while optimizing resource usage. Compared to the baseline, FlexFloor increases response token generation efficiency and throughput without sacrificing rollout reward.
Multi-agent AI systems often span various clouds, models, and record systems, creating a mismatch between billing data and actual business transactions. The Two-Ledger Model proposes a settlement ledger for FOCUS-conformant billing data and an attribution ledger for invocation-grain usage records from telemetry. A correlation spine links these ledgers using various metadata to assign usage to specific business transactions. This model allows for real-time rate pricing, period reconciliation, and the identification of unmetered spend. The approach has shown potential to recover unmetered costs and reduce transaction costs in simulations.
The study investigates the differences in how autonomous AI agents and humans approach proof-making in mathematics. Through "simulated collaboration," the research examined AI attention to syntax versus semantics and divergences in next-step preferences. AI agents were found to favor explicitly marked gaps and showed an aversion to complex proof centers. Conversely, they were attracted to surprising proof bodies, which they tend to normalize towards a baseline. These findings reveal blind spots in human-AI collaboration during mathematical proof construction.
The research empirically characterizes the Trust Responsive Authorization Policy (TRAP) for Agentic AI by mapping an Authorization Readiness Score onto discrete authorization levels. TRAP requires consecutive observations above a threshold plus a margin for escalation, and sharply restricts authority upon a drop in readiness. Comparisons with static, trust-only, and risk-only policies show TRAP restricts at the onset of degradation and re-escalates sooner after recovery than the symmetric policy. The restriction advantage of TRAP depends on whether the degradation is abrupt, as its performance varies based on the degradation pattern. Results indicate that TRAP generally maintains authority above readiness supports for a larger portion of degradation intervals compared to other policies.
The paper compares Darwinian, Lamarckian, and agent-directed evolution in the context of artificial intelligence. Darwinian evolution relies on natural selection, Lamarckian evolution incorporates inherited learning, and agent-directed evolution allows agents to generate and evaluate behavioral variants. These mechanisms can accelerate adaptation when learning is effective and more variants are generated, depending on search effort and problem difficulty. A key challenge is ensuring that the evaluation targets remain aligned with the intended objectives during autonomous self-improvement. Independent evaluation and oversight are necessary to guide and limit AI evolution.
Closed-loop evaluation is proposed for deployed agents by modeling assessment as performance estimation under biased raters and censored outcomes. The method involves four levels of analysis—span, trace, conversation, and engagement—measuring execution, substance, trajectory, and results. It incorporates Provenance and Adjudication components to certify evidence and evaluators, with inner and outer loops for self-auditing and agent improvement. The approach demonstrates that evidence-and-rule judgment yields more reliable results than holistic scoring in agent evaluation.
Cooperative multi-agent reinforcement learning is challenged by the complexity of joint state-action spaces in multi-agent systems. The paper proposes EFXPLORER, a constrained exploration framework to discover cooperative behavior. This framework uses an epigraph reformulation to introduce adaptive exploration budgets that balance exploration objectives with task performance. Successor distance-based intrinsic rewards are also incorporated to encourage temporally extended exploration. Empirical results show the method outperforms baselines and induces novel cooperative strategies across benchmark suites.
SteelAgent uses a language model interface to predict steel properties by combining metallurgical descriptors with machine learning models and physics-informed knowledge. The research investigates the impact of different held-out unit choices on the conclusions drawn from tensile-strength evaluations. Results show that holding out steel families significantly reduced the $R^2$ for gradient boosting, and the empirical coverage of prediction intervals deteriorated on held-out sources. The study also examines how physics-derived features affect the prediction accuracy and demonstrates the potential for calibration deterioration on held-out groups. These findings suggest that reporting the evaluation population alongside property estimates is beneficial for model assessment.
Octopus is a single-model framework designed for latent-KV multi-agent reinforcement learning, addressing limitations in existing multi-agent systems. It coordinates multiple threads by sharing a backbone and managing KV states through a directed acyclic graph (DAG) using five primitives. The framework employs mechanisms like Zero-Copy Radix Forking and Attention-Only KV Mounting to efficiently manage shared attention and reduce memory usage. Training results show Octopus outperforms other methods, achieving higher accuracy and increased throughput while reducing peak KV memory.
A simulation-based framework is presented for uncertainty quantification and calibration of empirical Bayesian inverse methods in M/EEG source imaging. The study compares Bayesian minimum norm (BMN) with sLORETA depth weighting and sparse Bayesian learning (SBL) with sparse basis field expansions (S-FLEX). S-FLEX is shown to offer lower localization errors and tighter uncertainties than BMN+sLORETA before recalibration. Post-hoc isotonic recalibration improves the performance of both methods, with SBL+S-FLEX demonstrating superior localization and uncertainty in real auditory MEG data.
The study benchmarks frozen forecasting foundation models as classifiers on the MONSTER dataset. Researchers compare the frozen representations against established time series classification methods and random-weight models. Pretraining does improve classification accuracy over random baselines on five datasets, showing gains up to 23 percentage points. However, the frozen representations do not consistently outperform the strongest non-pretrained baselines across all settings. The findings emphasize the necessity of using strong, task-specific baselines for evaluating these models.
Data streams are often non-stationary due to concept drift, requiring frequent retraining of predictive models. Time Neutralising Trees (TNT) are introduced to achieve robust stream classification despite this drift. TNT filters out old training samples during training to neutralize irrelevant data. This process allows the model to adapt by re-including relevant instances as the stream context changes. Evaluation on real and semi-real streams shows TNT achieves state-of-the-art performance across various drift types.
Autoregressive decoding's KV cache grows linearly, necessitating compression methods beyond existing redundancy exploitation. LowRAM introduces value cache amortisation by predicting values based on keys within each attention head, storing only residuals for the worst-reconstructed tokens. This is paired with cross-layer low-rank key compression to create an asymmetric pipeline. This approach minimizes decoding overhead while achieving up to $4\times$ value compression. The resulting LowRAM method, combined with other techniques, enables significant compression and throughput gains for long contexts.
The paper develops a stochastic dynamical-systems perspective on three classes of generative AI models. Large language models are framed as nonlinear autoregressive models with attention dependencies, linking pretraining to token generation. Interacting LLMs are analyzed using mean-field ordinary differential equations to capture emergent behaviors like preferential attachment. Neural stochastic differential equations offer a framework for continuous-time generative modeling of stochastic trajectories. The work explores probabilistic prediction, information exchange, emergent behavior, and path-space modeling in generative AI.
Reinforcement learning in Partially Observable Markov Decision Processes (POMDPs) is challenging due to history dependence. The research bridges this gap by applying neural value function approximation to POMDPs, moving beyond limitations of existing finite-time convergence analyses. The authors establish bounds for point-wise approximation error, critic error with nonlinear approximations, and ultimately achieve a finite-time sublinear regret. This result represents the first such regret guarantee based on neural function approximation for finite-memory POMDPs. Numerical experiments confirm these theoretical findings.
Mechanistic Shaping is a framework designed to shape models by adhering to specified rules during training, measured by their internal representations and outputs. This method has been successfully applied across various domains, scales, and training paradigms, showing a substantial reduction in unwanted behaviors while maintaining task performance. The research investigates whether shaped concepts remain monitorable after training, leading to the introduction of a hierarchy of monitor classes. Three geometric mechanisms—drift, erasure, and persistence in higher-order statistics—are identified as ways concepts can be circumvented. The study concludes that shaped concepts largely remain monitorable, and interpretability is a valuable tool for training.
Optimistic Hedge, a weight update method, has been studied for minimizing regret in general-sum games. Initial results showed regret was polylogarithmic in the time horizon $T$, and Ha found bounds independent of $T$ for specific learning rates. However, for three or more players, the regret bounds depend significantly on the number of players and actions. The analysis uses a temporal approximation framework to derive upper bounds, leading to a simplified proof of recent regret bounds for $m$-player games.
The research develops methods to detect and parameterize Euclidean point cloud data lying near a wedge sum of tori in high-dimensional data. Persistent cohomology and cup products are employed to identify the distinct toroidal components within the data. Optimization over low-energy multivariate trigonometric polynomials is used for the best data fit on each component, reducing the problem to generalized ridge regression. The framework has demonstrated utility in tasks like interpolating video data and recovering attractors from dynamical systems.
Bayesian Tensor Train Kernel Machines (BTT-KM) extend existing Tensor Network kernel machines by incorporating hierarchical ARD-type priors on the tensor train cores. This framework infers the necessary tensor train ranks and feature dimensions directly from the data. Mean-field variational inference enables a Bayesian alternating least squares algorithm with the same computational cost as deterministic methods. The approach yields a closed-form predictive distribution and recovers true ranks and dimensions in synthetic experiments.
Shapley values are used for attribution in machine learning but are hard to compute exactly. ShapleyOG proposes a fast, adaptive estimation method using Odd Gaussian Processes to overcome the limitations of previous methods. ShapleyOG exploits the dependence of Shapley values on the odd component of the game, requiring half the observations for the same posterior. It refits the Gaussian process sparsely and efficiently updates the selection criterion, achieving estimation accuracy comparable to ShaplEIG. Experiments show ShapleyOG significantly reduces overhead compared to state-of-the-art estimators.
Task-Calibrated Random Features (TCRF) are proposed to improve random features by balancing frequency moments derived from task-specific influence across kernel entries. These influence-based weights are shown to coincide with classical minimum-distance calibration, offering a reusable, task-derived calibration space. The method achieves first-order variance cancellation and eliminates the linear component of pilot noise in kernel methods. Experiments demonstrate improved approximation to exact results for kernel ridge regression and Stein discrepancy, even with restricted pilot budgets.
Estimating causal effects from observational network data requires accounting for both individual treatment and neighborhood exposure. The proposed Network Effect Decomposition (NEDNet) parameterizes the conditional outcome response using a pre-treatment baseline and separate treatment/exposure pathways. This framework incorporates path-specific neighborhood contextualization to provide local pre-treatment information to these pathways. A reference-normalized decomposition is used, which is equivalent to a zero mixed treatment-exposure contrast. Experiments show NEDNet achieves the lowest PEHE across various causal effect estimations on benchmark data.
The research proposes a heteroskedastic functional mixed-effects regression framework to jointly model the conditional mean and variance of repeated functional trajectories. This method accounts for participant-specific random effects in both the mean and variability components. Application to postprandial glucose trajectories from the AEGIS study revealed that premeal glucose and carbohydrate intake affect glucose variability differently than HbA1c affects the mean. The analysis showed that between-participant differences in residual variability persist despite controlling for observed covariates. Modeling conditional variability provides a richer understanding of glucose regulation and individual heterogeneity.
Interventional state-space models (iSSMs) use perturbations to identify parameters in state-space models, overcoming limitations of standard state-space models. The authors extend iSSMs to handle general nonlinear latent dynamics, proving identifiability up to permutation, scaling, and shift under specific conditions. This theoretical extension is validated using FitzHugh-Nagumo dynamics simulations. Applications include adjudicating between decision-making models and predicting responses to unseen neural micro-stimulation. The work provides practical guidance for designing stimulation protocols to enhance identifiability in neural data.
Patch length significantly impacts the performance of attention-based time series forecasting, yet its selection is typically empirical. The authors develop a theoretical framework to explain why patch length affects forecasting accuracy and propose a principled method for selection. For linear state-space models, they derive a factorization that separates forecast dynamics from filtering dynamics. This analysis leads to a patch-selection criterion balancing representation and estimation errors under temporal dependence. The criterion considers factors like predictive structure, head count, and long-run dependence in determining optimal patching.
The paper investigates cross-validation (CV) for global time series forecasting models based on deep learning. It proposes a series-wise CV approach, partitioning data across both series and time, which differs from the conventional time-based partitioning. The study benchmarks various CV strategies, including K-fold CV, bootstrapping, and Monte Carlo CV, across 11 datasets and 10 neural network architectures. Results indicate that series-wise CV improves downstream forecasting accuracy by promoting generalization across time and series. Conversely, time-wise CV provides a slight advantage when estimating out-of-sample error.
Gradient descent iterates oscillate at the edge of stability, posing a challenge for continuous-time models like gradient flow. The paper introduces the edge coupling as a third object from which both central flow and rod flow can be derived. Rod flow and central flow correspond to coarse-grainings over different time scales, with rod flow coarsening only the fastest scale and central flow coarsening both fast and intermediate scales. These flows are also derived for momentum methods like heavy ball and Adam and experimentally verified on neural networks.