Flow matching models need to satisfy constraints, but existing methods often displace samples from the pretrained distribution. MintFlow introduces a training-free framework that enforces constraints by minimally intervening on the flow trajectory. It calculates the minimal perturbation of an intermediate state required for subsequent evolution to meet the target constraint. An adjoint formulation provides a closed-form expression for this perturbation, avoiding iterative optimization. MintFlow achieves competitive constraint satisfaction while better preserving the pretrained generative distribution across various tasks.
System-1 decision models aim to reduce latency and cost in LLM agent harnesses by performing decision-making in a single forward pass. A paired evaluation assessed an open-weight (Laya) and a hosted (Jev) System-1 model across 11 agent decision points using large sets of data. The hosted Jev model was significantly more accurate on nine of the eleven decision points compared to Laya. The evaluation also uncovered several pipeline issues, such as reporting errors in cost savings and gate accuracy. Overall, the study provides a self-audited comparison of these fast decision models against slower evidence-based methods.
The research investigates whether AI disclosures in annual reports provide a signal about corporate responses to AI risks concerning societal resilience. Using LLMs, the study analyzed 9,821 annual reports from UK listed companies between 2020 and 2025. Empirical patterns show an increase in AI risk and adoption disclosures, with vendor mentions clustering around major providers like Microsoft. Disclosure significantly varies by sector and market, with certain sectors showing lower risk disclosure. Most AI risk disclosures found are not substantive, indicating a gap between reporting and actual risk management.
Training-free safeguards in text-to-image generation often use broad, global safety signals, which is found to have a trade-off between coverage and selectivity. Global unsafe subspaces fail to cover diverse unsafe semantics while broader aggregation distorts benign prompts. CALM is proposed as a training-free safeguard that corrects prompt-local counterfactuals instead of using uniform global removal. CALM routes prompts to specific unsafe categories and minimally modifies representations to the safe side, suppressing unsafe components. This method achieves better unsafe content suppression while preserving the utility of benign prompts.
Large language models performing long-horizon tool use benefit from estimating the value of potential next tool invocations before execution. The proposed Comparative Inference for Tool-use Agents (CITA) trains a model by combining observed behavior, simulator supervision, and LLM comparisons. This model learns to estimate the likelihood that a next tool invocation supports overall task success within the current context. Training with CITA consistently improved Tool F1 and task success across multiple benchmarks and LLMs. The method successfully learns accurate step-level value estimates for comparative tool choices.
World editing is formulated as intervening on existing, executable worlds while preserving necessary properties. Intervention depth is introduced to quantify how strongly edits couple world entities, dynamics, and systems. The capability is demonstrated using industry-grade game modding and a benchmark called IGMBench across Minecraft and Terraria. Frontier coding agents show significant world-editing power, though reliability decreases with intervention depth. This research establishes world editing as a distinct capability separate from world generation and interaction.
The study analyzed 21 topological metrics derived from Natural Visibility Graphs (NVG) for cyber attack detection. Importance analysis methods, including SHAP, Permutation Importance, Boruta, and RFE, were used to select a compact subset of these metrics. Configurations like Top3 achieved the best performance on the CICIDS2018 dataset, showing 97.148% accuracy and a high MCC of 0.9675. This metric reduction significantly decreased the total runtime while maintaining high predictive performance. The findings demonstrate that importance-guided metric reduction yields a compact representation with superior detection capability and efficiency.
DeReAct introduces a modular agent architecture to improve the reliability of AI agents by externalizing gating policies. It incorporates a Critic to validate proposed actions before execution and a Context Manager to reconstruct the environment state and certify task completion. This approach addresses the difficulties in action authorization and completion control present in standard ReAct-based agents. Empirical results show DeReAct increases Pass@1 scores for weaker models, with gains diminishing for stronger models. DeReAct yields more evidence-complete trajectories, suggesting a trade-off between earlier termination and stronger grounding.
Text-to-image generation requires balancing image quality (reward) and visual variety (diversity). The paper formulates this as a satisficing problem where images must meet a reward floor and the batch must meet a diversity cutoff. SatisDive is introduced as a training-free inference-time method to traverse the satisfaction-diversity frontier. This method uses batch-relative reward cutoffs to prioritize reward improvement for lower-scoring candidates and diversity among higher-scoring ones. SatisDive demonstrates Pareto-dominance over existing steering methods across various settings.
A Vision-to-Language Decision Support Framework (THPL) is proposed for managing rainbow trout feeding in Recirculating Aquaculture Systems (RAS). The framework extracts feeding intensity via an Activity Coefficient and uses a Hierarchical Behavior Encoder to generate dual-evidence representations of physical and implicit tokens. These tokens are integrated with environmental data and expert rules to fine-tune a Large Language Model (LLM) using LoRA. Counterfactual multimodal Direct Preference Optimization (mDPO) is then applied to reinforce causal reasoning and improve decision accuracy significantly. This approach offers a novel paradigm for precision aquaculture by linking kinematics with LLM reasoning.
Recent deep learning advances in time series prediction necessitate reliable uncertainty quantification. Conformal prediction offers a distribution-free method for prediction intervals, but its guarantees assume data exchangeability, which is often violated in time series. This work systematically investigates using deep sequence models, like RNNs and Transformers, within conformal prediction for time series. The research explores three methods: conditional quantile regression, conditional quantile function estimation, and localized conformal prediction. The study provides theoretical guarantees and demonstrates the effectiveness of deep sequence models in this context through experiments.
Supervised fine-tuning (SFT) exhibits more forgetting than reinforcement fine-tuning (RFT) during classification tasks, despite semantically correct teacher demonstrations. The study investigates this difference by analyzing updates decomposed into semantic and style components across tasks with shared semantic rules. RFT preserves symmetry by maintaining zero semantic error when starting from a neutral policy, whereas SFT develops style drift due to nonuniform teacher preferences during population updates. Explicit conditions show that SFT forgetting has a positive lower bound over a finite interval, while RFT retains zero semantic error. Simulations confirm these theoretical predictions regarding the learning style and forgetting semantics.
ENCORE introduces an exact parallel control method using Replica Exchange for Diffusion Generation. It provides a method for inference-time control, steering a pretrained generative model to a target distribution without retraining. ENCORE avoids simulating intractable time reversal by storing generation trajectories in each replica. The method proves target invariance and establishes that the dynamics are those of non-equilibrium replica exchange with the exact time reversal as a forward proposal. ENCORE achieves competitive accuracy in various tasks, remaining robust to sampler perturbations.
The paper addresses the challenge of selecting external data for private transfer learning, considering issues like relying on aggregated statistics and the potential for negative transfer. It models this selection problem using high-dimensional regression with multiple sources and a weighted ridge estimator. The method provides privacy guarantees using $\rho$-zero-concentrated differential privacy based only on summary statistics. A key contribution is a deterministic equivalent of the test error that incorporates sample size, covariance, model shift, regularization, and privacy noise. This theory enables the optimization of hyperparameters to determine the utility of private external datasets.
Flow-matching models for learning distributions have underdeveloped statistical generalization guarantees, especially concerning the low-dimensional structure of real data. This work derives finite-sample error bounds in the Wasserstein-$p$ distance for score-matching diffusion models learning an unknown distribution. The bounds show that the error scales according to the intrinsic dimension $d_p^\ast(P_{\mathrm{data}})$ rather than the ambient dimension. This demonstrates that flow matching naturally adapts to data geometry, mitigating the curse of dimensionality. These theoretical guarantees provide a basis for the empirical success of flow matching on structured data.
HOST is an efficient algorithm for Gaussian DAG learning that replaces subset search with nodewise hold-out scoring and convex regression. This method avoids the need for a supplied indegree bound. The approach leverages the expectation that hold-out score estimation inflates ordering scores for unselected candidates. HOST then recovers parents by recursively removing indirect effects between nodes. Under suitable conditions, the algorithm recovers a $p$-node DAG with maximum indegree $d$ in polynomial time with a sample complexity of order $d\log p$.
Clustering variables in causal graphs simplifies causal inference but arbitrary clustering can distort causal relations. The paper investigates when clustering operations maintain the identifiability of a causal effect between the original and clustered graphs. It establishes that nonidentifiability in the clustered graph does not necessarily imply nonidentifiability in the original graph without further assumptions. The authors present a class of identification-invariant clustering operations based on the c-components of the original graph. Practical applications for these results are also demonstrated.
Gaussian Processes (GPs) are useful for modeling uncertainty but require design decisions that can lead to misspecified models. Predictively Oriented Gaussian Processes (PrO-GPs) prioritize predictive uncertainty as the main inference target. The authors derived a reduced formulation and sampling scheme to make PrO-GP computations feasible. Experiments demonstrated that PrO-GPs yield better-calibrated predictive distributions than standard GPs when models are misspecified. This approach offers a robust alternative to standard GPs for uncertainty quantification.
SDECast introduces a Neural Stochastic Differential Equation (SDE) framework for continuous-time probabilistic weather forecasting. It extends SDE Matching to learn stochastic dynamics directly in physical space without repeated SDE simulation during training. SDECast successfully recovers drift dynamics and reproduces continuous-time behavior when applied to simulated geophysical flows. The method is scalable, producing skillful probabilistic forecasts for global weather at hourly resolution for up to five-day lead times. This work addresses limitations in traditional autoregressive models by explicitly encoding atmospheric temporal continuity.
DAWIS introduces a unified data assimilation method encompassing filtering, fixed-lag smoothing, and block smoothing. It replaces the state-level prior with a multitask stochastic interpolant over a window of consecutive states, assigning separate flow times to each state. Assimilation cycles invert this window to determine per-state turning points based on observation guidance. This process regenerates the states, where turning points control how each state is fixed, revised, or generated. Experiments show DAWIS improves filtering and smoothing performance over baseline methods in nonlinear systems with noisy observations.