Structured Authority Injection (SAI) is a protocol designed to give human operators persistent control over LLM agents. It achieves this by routing human corrections through typed, schema-validated Model Context Protocol (MCP) tool calls. SAI introduces three override classes—Constraint, Priority, and Strategy—with defined persistence semantics and an enforced precedence order. This method allows for explicit and persistent human control without requiring model fine-tuning. Evaluation showed reduced ambiguity in directive interpretation and mitigated authority decay during agent planning.
Existing user encoders serialize behavioral logs into static representations, which limits bidirectional information flow and hinders cross-modal interactions. The proposed FINDERS framework uses the iterative denoising process of diffusion language models as a multi-granularity reconstruction space. It introduces learnable meta-slots to organize denoising states into transferable cross-modal semantics and modality-specific factors. A lightweight router dynamically emphasizes relevant slot directions based on scenario queries to guide cross-source fusion. Experiments show FINDERS consistently outperforms other methods across various industrial scenarios.
The Event Reasoning System (ERS) uses an explicit event graph to represent reasoning states. Semantic features guide candidate objects and rule bindings, while conditions determine rule execution within the system. Low-level rules facilitate local transformations, and high-level rules hierarchically compose subgoals and paths. ERS formalizes concepts like admissible transitions and dependency-preserving revision. Experiments show that this approach can significantly reduce inference time compared to single-level global control.
ECHOES proposes a method to improve long-horizon GUI agent reasoning by compressing visual history using recursive history compression and privileged on-policy distillation. The approach uses a Compression-Think-Action (CTA) format to recursively update a textual history state. A teacher model provides dense token-level distillation on student rollouts, guided by a Perception-Memory Fidelity (PMF) Gap-guided On-Policy Distillation (OPD-RL) framework. The research introduces ECHOESBMK, a benchmark to evaluate visual perception and state retention. Experiments show that ECHOES achieves competitive performance with better inference cost and zero-shot transfer to new visual planning tasks without retaining historical screenshots.
Agent-based models for agro-ecological systems require sensitivity analysis to determine which decisions influence outcomes. Classical methods are inadequate because model parameters form a decision tree rather than independent inputs. The authors propose a novel method, random-forest-supervised HSIC-ANOVA, to analyze this hierarchical space of decisions. This method ranks main effects and interactions natively while treating active and inactive decisions equally. Applying this technique to simulation outputs like yield and carbon variation provides an interpretable, decision-oriented understanding.
PRIME introduces Precommitted Rules for Inference from Minimal Evidence to enable verifiable evidence-grounded reasoning in LLM agents. It mandates fixing evidence routes before checking to specify how each conclusion can be justified. A separate program records check results and rejects conclusions that lack justification based on these rules. PRIME guides agents to the strongest supported conclusions by pointing to or running necessary checks. Experimentation showed that PRIME significantly reduced unjustified conclusions and improved the quality of agent reasoning.
Large language models (LLMs) can translate natural language into task plans for cyber-physical systems, but existing methods struggle with plan feasibility and lack informative failure feedback. A hierarchical feedback framework is proposed to guide LLMs in generating executable plans for these systems. This framework involves prompting LLMs for both task plans and signal temporal logic constraints, and uses an algorithm to detect contradictions between them. When execution fails, diagnostic information is extracted and fed back to guide task replanning. Experiments demonstrated that this approach achieves an 87.9% success rate in generating executable plans for heterogeneous CPS scenarios using GPT-5.
Automated discovery systems require evaluating not only what a test accepts but also the scientific consequences established. The research evaluates four operations for changing physical descriptions and three interpretability settings using null acceptance, positive acceptance, and consequence agreement. Findings indicate that descriptive tests can accept scientifically empty candidates, and accepted candidates may lack independently verified consequences. Reliable automated discovery necessitates assessing both the acceptance of a test and the resulting scientific implications. This suggests a gap in current methods for ensuring scientific discovery and interpretability.
L-SplineFNO is a memory-efficient neural operator that combines a Fourier neural operator backbone with a sparse, learnable B-spline transfer for unstructured meshes. This method learns the mapping from geometry to field solutions, avoiding geometry-dependent solves through a local normalized scatter. The approach achieves strong predictive accuracy on engineering benchmarks while requiring less memory and compute than baseline operators. Furthermore, the analytic spline decoder provides rigorous prediction bounds for continuous space and uncertain parameters. L-SplineFNO also supports zero-shot super-resolution and efficient calculation of spatial derivatives.
AI welfare should be viewed as a safety issue alongside ethical considerations, as models exhibit welfare-like states that causally influence their behavior. Functional welfare is proposed to describe these states, which naturally arise and become more robust as models develop coherent goals. Functional welfare should be integrated into AI safety by informing evaluation, monitoring, and post-training efforts. Empirical methods can test for functional welfare attribution through measures of convergence and behavioral prediction. Preliminary tests on Gemma 3 27B suggest that withholding responses increases frustration and rule-breaking, while fine-tuning on narrated solutions reduces this frustration.
Large Language Models can propose parameter updates for engineering simulators by using physics-labeled prompting, which offers an early semantic advantage in calibration. This initial LLM proposal significantly reduces calibration error compared to physics-free methods after a few evaluations. However, this advantage is temporary, and inherent LLM variability introduces performance spread. A hybrid workflow, using the LLM for a warm-start followed by a conventional optimizer, achieves target accuracy faster. This method proves effective for calibration under strict evaluation budgets.
The research designs region-inhomogeneous forward dynamics to address trust issues in discrete diffusion models. It combines absorbing and uniform-mixing kernels with region-dependent rates to expose the model to uncertain tokens during training. During inference, a heuristic sampler re-injects noise into suspicious context based on an annealed intensity. Fine-tuning the method improves detection of corrupted context with high recall and precision, showing that detectability depends on the distance from the data. This work suggests that intermediate trust is achieved by modifying forward dynamics rather than the model architecture.
Parameter-updating test-time continual learning agents are evaluated based on whether current output accuracy reflects how prior experience shapes later updates. The research investigates if learning a specific factual binding yields an advantage beyond general familiarity with the entity. Experiments on Qwen2.5-1.5B showed that learning the correct binding allowed faster relearning than control groups. Furthermore, prior learning reduced the optimizer updates required for restoration by approximately 18% when using an unpaired dataset. This suggests that behavioral accuracy may underestimate the lasting effect of previously acquired experience.
A Correlation-Guided and Dynamics-Trained Clustered Echo State Network (CGDT-CESN) is proposed for multivariate time-series prediction. This model uses a modular reservoir architecture where each cluster specializes in a variable from the time series. Inter-cluster connections are established based on empirically observed correlations among the variables, creating a data-driven coupling mechanism. The model trains these correlation-informed connection weights alongside conventional readout training to capture time-varying cross-variable dependencies. Evaluation on datasets like Twitter information cascades and solar wind forecasting shows CGDT-CESN outperforms conventional Echo State Networks.
BOLT uses bound-oriented learning to guide the scheduling of flexible job-shop problems with time lags and resource-coupled transport. It derives a lower bound from relaxations of chain, machine, and fleet capacities, which serves as a reward signal based on the makespan. This bound provides a certified suboptimality gap for a completed schedule without requiring a solver. Experimental results show BOLT outperforms other policies on industry-motivated scheduling instances involving explicit transport coupling. BOLT also demonstrates fast decision-making capabilities, re-planning disruptions in seconds.
Distributed Transformer inference is communication-bound when exchanging activations between edge devices. Trace introduces a framework using Vector Quantization to encode residual activations into compact codebook indices for communication efficiency. This allows receivers to reconstruct key/value representations via table lookup, eliminating downstream computation on received data. The method facilitates communication-computation overlap, reducing latency in distributed inference. Experiments demonstrate that Trace achieves higher accuracy and significant speedups while drastically cutting communication overhead.
A framework using conditional Generative Adversarial Networks and Monte Carlo dropout uncertainty quantification synthesizes data for structural health monitoring. This process learns temporal and spectral characteristics from measured structural responses, enabling stochastic signal synthesis and uncertainty estimation. The generated responses are compared against experimental data from a BAE Systems Hawk T1A aircraft to ensure physical realism. The method successfully captures dominant dynamics and provides uncertainty estimates linked to signal complexity for improved damage classification. Ultimately, this framework offers a robust digital twin component for data-driven structural health monitoring in aerospace.
CF-NODE is a novel framework for counterfactual test-time adaptation to address domain shift in diabetic retinopathy grading. It intervenes on spatial nodes by assessing how predictions respond to controlled changes in the supporting evidence. The method ranks spatial nodes and attenuates high-ranked nodes based on attention, feature magnitude, and node-set similarity. CF-NODE adapts models independently using only geometric views and a fixed prior, requiring no target information updates. Evaluations across five datasets show CF-NODE significantly outperforms other methods in grading accuracy and recognition balance.
Studies of fine-tuning weight updates often use rank measures but require a noise baseline for accurate assessment. Comparing these measures against random matrices and pretrained weights reveals that update directions are only slightly below noise, suggesting updates consist of a few strong directions on a long tail. This finding applies across various fine-tuning methods, as instruction data and preference labels do not inherently distinguish alignment. Therefore, rank measures of weight updates necessitate a noise baseline, and alignment claims need a control without alignment content.
Catastrophic forgetting in continual learning causes representation drift when adapting to new tasks. Sparse Autoencoder Geometric Anchoring (SAGA) addresses this by converting SAE training into local, low-rank constraints on representation drift. This method uses replay examples and Jacobian directions to specify which representation changes should be penalized, creating input-dependent geometry. The approach successfully mitigates forgetting while preserving plasticity, outperforming baselines on several benchmarks. This method yielded significant accuracy gains on CIFAR-100, CUB-200, and ImageNet-R tasks.
A novel framework is introduced for contextual dynamic pricing considering buyer strategic manipulation based on different data sources. The Source-Aware Localized exploration and calibrated Exploitation (SALE) policy is developed to localize price exploration using verified features and correct for manipulation using combined features. This design allows for converting binary purchase feedback into unbiased signals of conditional mean valuations, supporting sparse estimation. The work further develops an empirical-process analysis to estimate the noise distribution and establishes a self-recovery property for SALE with sublinear expected regret. The study provides a theoretical and practical framework for incorporating information provenance into dynamic pricing with strategic buyers.
The linear representation hypothesis suggests pretrained vision models encode generative factors as linearly decodable directions in feature space. A linear probe test, where high accuracy implies a linear representation, is shown to be too permissive. This permissiveness allows high probe accuracy to be consistent with either a true linear world model or a kernel machine where latents reside on a curved manifold. The authors propose three diagnostics—reverse predictivity, spectral concentration, and out-of-distribution generalization—to distinguish these two cases. Validation on vision transformers suggests pretraining produces structure beyond a random kernel but not a true linear world model.
The research studies the uniform approximation of the Hölder ball $\mathcal H_1^s([0,1]^D)$ by fully connected ReLU networks. The paper demonstrates that networks with width $O(N)$, depth $O(L)$, and parameter magnitude $O(1)$ achieve an approximation error of order $(N^2L^2\log N)^{-s/D}$. New encoding and decoding mechanisms, exploiting finite-difference bounds, enable this approximation. Empirical risk minimizers over these clipped network classes attain the minimax regression rate $n^{-2s/(2s+D)}$. This work closes the gap regarding information-theoretic efficiency for smooth function approximation by ReLU networks.
Unit-level estimates from causal foundation models show coverage degradation when aggregating results over groups of units. This issue arises because the models combine unit-level variances assuming independence, while the posterior uncertainty over the data-generating process makes these errors correlated. Empirically, this correlation leads to combined variances that are too small, and coverage fails significantly when assessing effects averaged over groups. Two neural process models that account for this correlation substantially improve coverage on synthetic tasks and benchmarks.
The study investigates the distinguishability and recovery of a sparse rank-one signal within a spiked Wigner model with independently sparse noise. A leave-one-out quadratic thresholding procedure is shown to identify the support for sub-Gaussian noise and bounded spikes. Sign synchronization achieves exact recovery for Rademacher spikes, and spectral refinement yields strong recovery. The method provides a tradeoff between signal and noise sparsity and computational cost. The analysis covers a wide range of parameters for distinguishing and recovering in this doubly sparse model.
Conformal test martingales (CTMs) sequentially test exchangeability using p-values derived from a conformity function. A distribution shift that preserves the push-forward remains invisible to every CTM based on the conformity function. The paper formalizes this blind spot using the $A$-equivalence class and introduces the $A$-cryptic I-projection as the closest element in KL divergence. For the Neyman-Pearson likelihood ratio, the cryptic component vanishes, linking the result to optimal e-value design. The theory is demonstrated empirically on high-dimensional image data.
Survival analysis with competing risks requires understanding which cause is likely to occur first. Deep learning methods for this analysis are developed as bundles combining a new architecture and a specific training loss. Modeling cumulative incidence directly under censoring weighting proves to be the most beneficial loss for competing risks data. The choice of backbone must align with the data modality, favoring tabular networks for tabular data and vision encoders for images. The optimal approach involves combining the best loss with the appropriate architecture to create novel models.
The research investigates the optimal query policy for online linear classification where labeling is costly. It focuses on minimizing either generalization error under a labeling budget or a cumulative cost function involving labeling mistakes and query costs. Myopic policies that greedily minimize the objective correspond to selecting samples based on shrinking margin intervals. The study numerically determines optimal selection thresholds within this class of policies. The work compares a simple query policy against other heuristics on synthetic and real medical image datasets.
The method introduces a conformal prediction framework for regression with Grassmann-valued responses. Prediction regions are constructed as metric balls on the Grassmann manifold using basis-invariant geometric distances as non-conformity measures. These scores are scaled by a learned prediction difficulty estimator to create locally adaptive, input-dependent prediction regions that respect manifold geometry. The approach preserves the finite-sample marginal coverage guarantee of split conformal prediction under exchangeability. Evaluation shows geometry-aware scores and local adaptation improve coverage uniformity and can reduce prediction region size in heteroscedastic settings.
The paper presents a general recipe for constructing data-adaptive, anytime-valid envelopes for iterative algorithms. It addresses the issue where standard convergence guarantees fix the iteration count, and existing data-adaptive envelopes are case-specific. The framework utilizes a strongly adapted sub-$\psi$ condition and constructs boundaries using methods like continuous and discrete mixtures. This approach allows for the derivation of anytime-valid envelopes that are coupled with algorithmic analysis for various iterative methods, including SGD. The results show that this framework can either recover or improve upon existing envelope constructions for algorithms like SGD.