An explainable, header-centric framework is presented for metadata-only Semantic Table Interpretation and Data Quality Assessment. This framework maps column headers to 39 interpretable FinalFormat types using lexical resources and token-level traceability. It activates validation rules based on a Data Quality Issues taxonomy to detect problems like missing data or wrong data types. These detections are aggregated into HeadersIQ, a lightweight quality metric. The method was evaluated on heterogeneous benchmarks, showing broad practical coverage across real-world metadata.
Synthesis Through Simulation (STS) is a schema-free data synthesis paradigm where an LLM agent generates data by interacting with policy-enforcing APIs in simulated enterprise environments. STS guarantees structural validity by construction because data generation occurs within the validating environment, decoupling validity from distribution modeling. The Generalist Populator (GP) agent addresses distributional fidelity and scalability, achieving 0.88 average marginal fidelity across ten environments without needing database schemas. The GP agent successfully satisfies constraints in all ten environments, unlike schema-privileged agents which fail on complex workflows. The full framework, environments, and generated datasets are open-sourced.
Agent-controlled forgetting involves having the acting model select previously observed tool results and replace them with short notes while retaining the original results in a recoverable archive. A Python harness allows for batch archival and explicit recovery without task-specific model training, protecting user instructions. Experimental results showed significant resource savings in noisy tool-use trajectories, with reduced input tokens and lower estimated API costs. However, some specific evaluations did not fully satisfy the follow-up assessments. The findings indicate that workload dependence is a central factor when managing reversible context for these agents.
Agentic AI can improve through methods like longer searches or modifications to output verification. The work compares these mechanisms using bounded verification with hidden terminal randomness to define mathematical languages. The study shows that independent majority amplification preserves certain languages, while existential acceptance over random tapes can allow incorrect outputs. Exact verification is achieved in the zero-randomness case, and the framework links self-improvement claims to requirements on correctness and evidence. Ultimately, the work establishes a framework for auditing self-improvement in agentic systems.
Self-improving LLM agents in credit pipelines must confine their evolution to the runtime harness, such as instruction text and tool-call logic, while keeping model weights fixed. A dual-loop engine with a measured admission gate was developed to ensure self-evolution remains reviewable with attached change logs and tests. Simulation showed the gated loop successfully admitted changes that did not worsen errors or raised missed flags in supervisory evaluations. Replacing the gate with an unbounded system allowed for more harmful changes, demonstrating the gate's role in error control. The findings relate mechanisms to the EU AI Act's provisions for high-risk credit scoring.
Planning for long-horizon agent tasks requires the ability to revise plans effectively when outcomes are unexpected. Plan-and-Patch introduces a framework using diffusion language models (dLLMs) to generate structured plans and repair them by filling in modified regions while preserving the surrounding structure. The method compares diffusion models against autoregressive planners in plan generation and repair tasks. Diffusion models show superior plan repair success rates over autoregressive models without task-specific training. Overall, Plan-and-Patch offers a framework for faster plan generation and more effective plan repair for agents.
Understanding human interpretations is crucial for developing human-centered AI, as conventional methods often assume a single definitive ground truth. Human judgments frequently involve inherent ambiguity, where multiple interpretations can be valid simultaneously. The paper advocates for shifting focus to modeling the space of plausible human judgments rather than reducing ambiguity to a single target. This approach requires distinguishing meaningful ambiguity from mere annotation noise. Ultimately, this perspective should guide the representation, learning, evaluation, and governance of AI systems to reflect human diversity.
The study investigates whether small LLM agents can operate effectively while adhering to explicit time budgets and using the allocated time productively. Agents often fail to control time use due to a lack of time awareness and inability to anticipate action durations. Interventions involving timing information and reinforcement learning with budget-aware rewards improve budget adherence but do not necessarily increase task performance. A central challenge identified is the gap between respecting time limits and actually utilizing the time to improve task performance. Results indicate that even when adhering to budgets, agents still struggle to allocate extra time productively.
SVD-based KV-cache compression involves using an encoder and decoder to factorize key/value weights, followed by a fine-tuning step to recover accuracy. A shared learning rate often creates misleading advantages when comparing fine-tuning recipes with different parameter counts. Freezing the decoder and only healing the encoder demonstrated a superior result compared to other methods under a shared learning rate. Tuning learning rates per arm eliminates this apparent advantage, showing that encoder-only healing achieves parity with other approaches. This method offers a memory-efficient way to retrofit low-rank KV-cache compression.
SPERA is an EEG foundation model designed to overcome challenges in building general-purpose EEG models due to data heterogeneity. It employs a joint-embedding predictive architecture to predict information in a latent space. The model incorporates a Legendre-polynomial spatial prior to encode scalp electrode geometries and uses factorized temporal and spatial attention. Pretrained on extensive EEG data, SPERA shows high balanced accuracy across various downstream tasks and demonstrates strong parameter efficiency.
Activation probes monitoring language models require synthetic data whose amount is determined by coverage rather than difficulty. The required sample size is set by the need to saturate the distribution for each type of monitored concept. Median "half-gain" size for each concept is found to be between 7 and 11 own-kind synthetic samples. The required breadth of data differs by concept, with instruction probes needing more kinds of data, while high-stakes probes show redundancy. Coverage, not difficulty, dictates the necessary quantity of synthetic data for effective monitoring.
A privacy-preserving framework for anomaly detection in in-vehicle networks based on the Temporal Transformer CAN Encoder with Federated Lightweight Heads is presented. This method aims to identify subtle temporal and contextual anomalies in CAN bus communications, which are often indicative of malfunctions or interference. A lightweight Transformer encoder learns the evolution of signals over time to capture these irregularities effectively. Federated learning allows multiple vehicles or ECUs to collaboratively improve a shared model without sharing sensitive raw CAN data. This combination achieves robust anomaly detection performance while ensuring efficiency and privacy.
Mixture-of-Experts (MoE) models generate routing telemetry during inference, which can reveal information about the model's internal computation. This telemetry introduces a privacy concern regarding membership inference, specifically whether an example was used for fine-tuning. A router-augmented membership inference attack, combining output signals with routing features, consistently improves membership inference across various MoE architectures and data domains. Leakage persists even when fine-tuning is frozen, suggesting the router exposes a projection of fine-tuning signals. Consequently, router telemetry acts as an additional privacy surface for fine-tuned MoE models.
Spontaneous self-organization in closed systems is investigated through constraint-induced self-organization via geometric radiation in coupled metric evolution systems. Simulations reveal a four-stage cycle: stress accumulation, super-exponential radiation, chaotic collapse, and convergence to a fractal limit cycle, termed the wedge-shaped attractor. Four jointly sufficient conditions, including an irreversible geometric horizon and persistent stress injection, are identified as triggers for this process. The research proves theorems related to geometric horizons, energy dissipation, and radiation as a phase transition channel. This work establishes a new paradigm for closed-system self-organization by linking nonlinear constraints and gravitational horizons.
Looped Language Models (LoopLMs) present a parameter-efficient method for scaling reasoning by reusing parameters across recurrent steps, but post-training is difficult. LoopOPD introduces a cross-loop on-policy distillation framework using recurrent computation within the LoopLM as its own supervision source. Dynamic LoopOPD further refines this by continually updating the terminal loop teacher during parameter updates for recurrent self-improvement. Experiments show that this method improves mathematical reasoning and extends to general reasoning and code generation benchmarks. Recurrent computation proves to be an effective source of supervision for LoopLMs.
Neural surrogates for vector-valued PDEs fail when physical states are rotated in a coordinate frame. The Invariant-Conditioned Isotropic Kernel Neural Operator (IKNO) is introduced to address this by building interactions from rotation-invariant scalar quantities. This method ensures that rotating inputs result in the same rotation of the predicted velocity change. IKNO achieves this rotation-robustness while matching forecasting accuracy with significantly fewer parameters than other rotation-aware models. Consequently, a compact, PDE-specialized model can eliminate coordinate dependence without sacrificing predictive quality.
Kolmogorov-Arnold Networks offer higher parameter efficiency and interpretability compared to Multi-Layer-Perceptron architectures for learning physical relationships. Computational experiments on the Feynman dataset and Gymnasium RL benchmark showed significant sample-efficiency gains. The Kolmogorov-Arnold architecture achieved similar performance with 40% fewer samples during training. Relative performance improvements of up to 50% were observed during the training process. These sample-efficiency gains were robust despite varying levels of reward noise.
Phase-HDC proposes replacing optimizer history storage with gradient thresholds in discrete phase learning. It achieves this by updating each angle by at most one step based on the sign of the current gradient when the gradient is sufficiently large. This method is shown to be the exact solution for a first-order loss model where parameter changes have a fixed cost. Phase-HDC achieves accuracy comparable to Adam with 6-bit moments while storing significantly less memory across various datasets. The method demonstrates substantial storage savings compared to standard float32 and 8-bit Adam implementations.
The research investigates the optimal information complexity for VC learning using the Conditional Mutual Information (CMI) framework. The authors explore whether the optimal PAC guarantee for VC classes can be recovered from algorithm-dependent information-theoretic quantities. They demonstrate that this guarantee can be recovered by constructing a learning algorithm whose eCMI is of order $O(d)$, where $d$ is the VC-dimension. Specifically, this is achieved with a randomized Majority-of-5 base learner that provides optimal in-expectation generalization. This work connects information theory to the generalization guarantees of machine learning algorithms.
The paper explores constructing an explanatory space for a weighted Naive Bayes classifier using a supervised representation. It reformulates the classical supervised distance based on conditional log-likelihoods into a discriminative representation using log-odds. This representation induces a distance that exactly matches the $\ell_1$ distance between vectors of analytical Shapley values. The work formally interprets the geometry induced by the model through this link to supervised distance and local explanation. Finally, empirical comparisons are made across several supervised distances using a $k$-nearest neighbors classifier.
Linear attention is a tractable model for in-context learning mechanisms in transformers. The research extends this theory to nonlinear single-index targets by showing a nonlinearity-noise equivalence. Linear attention only captures the linear component of the underlying function, with the nonlinear structure contributing to generalization error as effective noise. This reduction allows linear theory results to be applied to nonlinear tasks. The findings suggest a limitation in the reduced linear-attention model and offer a starting point for studying nonlinear in-context learning.
Distributionally robust optimization (DRO) optimizes decisions over an ambiguity set, but poorly aligned sets can lead to overly conservative results. Diagnostic-transport DRO (DT-DRO) addresses this by using calibration data to adapt the ambiguity set geometry based on observed predictive errors. DT-DRO employs the conditional probability integral transform cumulative distribution function to diagnose misallocation and translate this into an outcome-level transport. This method results in a computationally tractable dual reformulation and provides tighter robustness guarantees. Experiments show DT-DRO improves decision quality, especially under structural and tail misspecification.
Conformal prediction offers prediction sets with finite-sample guarantees, but label verification for calibration is costly. A partial verification method is developed to return identical prediction sets to complete verification. This method characterizes calibration certificates and coordinates verification across calibration examples. Verification cost is reduced by 15-82% compared to checking calibration examples individually. This efficiency gain is achieved while maintaining identical prediction sets.
MCMC methods often use control variates to reduce estimation variance, but finding these requires solving a complex Poisson equation. The work extends solutions for samplers with a known spectral decomposition to general targets using a learned change of variables, like a normalizing flow. This transformation allows Markov kernels and their Poisson solutions to be mapped, yielding explicit control variates in the latent space. Running samplers in this latent space produces estimators that are consistent under mild conditions. Experiments confirm that this method performs well compared to existing state-of-the-art techniques.
The paper establishes a general $\Omega(\sqrt{T\gamma_T/\log T})$ minimax regret lower bound for kernel bandits with non-constant continuous kernels. This result addresses the lack of lower bounds for general kernels, contrasting with existing upper bounds that scale as $\sqrt{T\gamma_T}$ up to log factors. The authors show that the log factor is generally unavoidable but can be removed under specific conditions. The findings imply that the minimax-optimal scaling is exactly $\Theta(\sqrt{T\gamma_T})$ for certain specific kernels like Matérn and $\gamma$-exponential.
Anchored Langevin dynamics (ALD) and reflected anchored Langevin dynamics (RALD) are useful for sampling non-smooth, heavy-tailed distributions. The paper proposes non-reversible anchored Langevin dynamics (NALD) and non-reversible reflected anchored Langevin dynamics (NRALD) to handle non-differentiable densities. These methods introduce a circulation drift to accelerate sampling by breaking reversibility. Finite-time non-asymptotic analysis demonstrates that NALD and NRALD can converge faster than reversible methods. Numerical experiments confirm the efficiency of these proposed algorithms for sampling.
Synthetic tabular data generators often focus on distributional fidelity, which does not guarantee the preservation of causal effects. The authors address this by proposing a framework to improve causal fidelity directly within a generative model. This involves adding a causal discrepancy penalty to the standard generative objective. They implement this using a causal-penalized TabDDPM optimized with an on-policy score-function estimator. Experiments show that this causal regularization can improve expected causal fidelity while maintaining competitive statistical fidelity.