Long-horizon language agents use external memory where memory use exhibits heavy-tailed traces, concentrating on a core while rare states accumulate prediction errors. This effect is reproducible and depends on the agent's policy, with semantic LLM policies yielding the strongest truncated-power-law-compatible traces. The authors propose the Core--Tail World Model (CTWM), a rank-based memory controller to manage prompt budget effectively. Applying CTWM reduces prompt tokens and lowers prediction errors relative to baseline memory methods on synthetic graph worlds. These findings suggest heavy-tailed memory traces serve as a practical control signal for token-efficient agent world models.
Causal world models are studied for modular LLM agents where actions in one module affect others. The research introduces FedCausalCompose, a framework using causal world-models to provide intervention-response evidence for cross-module interfaces. Causal interfaces prove most beneficial in structured tool environments where API signatures reveal preconditions and effects. However, this benefit is less pronounced in dialogue and narrative settings unless causal information is explicitly made relevant at action time. Therefore, causal world models are most helpful when cross-module interfaces are statistically identifiable and presented in an actionable format.
Praxa is an agent harness designed to explicitly represent the states of proposal, authority, dispatch, and verified external effect in AI agent execution. It introduces mechanisms for deterministic admission, brokered execution, external read-back, reconciliation, and reviewed promotion. The research presents four evidence lanes, including repository audits, pilot testing, and comparative development analyses. While demonstrating an evidence-bound architecture for making authority-to-effect transitions testable, the current evidence does not establish superiority in security, safety, or performance metrics. Ultimately, Praxa's contribution is an explicit and testable framework rather than a proven lift in autonomous capability.
Role-specialized QA pipelines pass rationales between reasoners and verifiers, but the effect is unclear regarding answer accuracy or support assessment. A message-intervention diagnostic showed that adding faithful rationales adds little to answer accuracy, but corrupted rationales significantly alter support judgments. Corrupted rationales strongly shift support judgments, and human audits reveal that models often accept corrupted rationales. The study concludes that rationale sharing should be evaluated as a verification-message mechanism rather than solely as a route to better answers.
Agent harnesses often require small language models (SLMs) for microtasks like command approval and tool selection. A benchmark was created to measure if off-the-shelf SLMs meet practitioner-defined thresholds, which are anchored to a non-LLM baseline. The study found an "eligibility gap" where most configurations failed to meet these thresholds across various models. Quantization to 4-bit did not substantially change this gap, suggesting the issue tracks model size rather than mere precision. The practical implication is to use SLMs only when a performance deficit compared to a baseline is observed.
The study investigates the effectiveness of inference-time PRM-Pruned Fragment Grafting (PPFG) for transferring knowledge in reasoning Language Models. The research tests PPFG on Qwen2.5-7B-Instruct with Math-Shepherd across multiple benchmarks and models. Results indicate that PPFG is statistically indistinguishable from an independent parallel Chain-of-Thought baseline in stagnation and random-targeting scenarios. The analysis reveals that injected grafts often target chains that were already successful or near completion, suggesting the mechanism is not heuristic-specific. The work contributes an equivalence-testing template to establish nulls for inference-time mechanism effects.
Personalizing large language models requires selecting preference pairs based on user preferences rather than aggregate quality. Existing methods use heuristics for pair selection, which can degrade personalization. The research frames personalized preference learning as a geometry-aligned optimization problem by examining gradient interactions. This insight suggests that pair selection is a geometric decision influencing personalization advancement. The proposed GAP-DPO algorithm uses utility-aware, geometry-aligned selection to improve stylistic fidelity and generation quality.
K-Dense BYOK is a free, open-source AI research assistant designed for scientists that runs directly on the researcher's local computer. It allows users to leverage any chosen model while providing scientific scaffolding, workflow templates, and a complete, hash-chained lab notebook. The system records agent actions rather than relying solely on its output to combat model overclaiming. It guides research through written procedures and roles, tracking all work in a verifiable, immutable record. This tool demonstrated strong performance in research execution and quality across various prompts.
Profession-specific system prompts increase token usage and cost without providing a consistent accuracy improvement for scientific tasks. Evaluation of Scientific Agents showed no statistically clear accuracy gain from using these detailed profiles over baseline settings. Matched profiles resulted in significantly higher token usage and cost per successful call to the model. While longer prompts offered a minor advantage in specific error scenarios, this was attributed more to prompt length than domain expertise. Loading full profession profiles by default did not improve accuracy or cost considerably.
The research investigates reward exploits and countermeasures in training language models for conversational humor. Two methods are explored to capture understandable surprise and predicted audience amusement, testing how models react to word-shuffled replies. A fluency filter attempts to detect these shuffles, but the combined reward system still allows some exploits. Normalizing laughter cues across speakers blocks some attacks, though unmatched expressions remain exploitable. The study concludes that countermeasures must block exploitable shortcuts while maintaining the intended behavior of the reward.
Research uses machine learning and deep learning to accurately predict travel times within supply chain transportation and logistics. Accurate travel time estimation is vital for improving logistics consistency, demand forecasting, lead time management, and assembly planning. The study leverages large historical data and novel techniques to build an accurate model for predicting inventory travel times. This prediction is crucial for supply chain members to enhance delivery performance and ultimately improve customer satisfaction.
EHR2Trace is a system designed to convert Electronic Health Record (EHR) data from various sources into traceable patient events for model training and evaluation. It links these events to their source records and separates event time from information availability, distinguishing between medication orders, dispensing, and administration. The system supports shared event representations compatible with OMOP and MEDS exports, incorporating automated validation. Experiments demonstrated that data structure significantly impacts model performance, with assigning later diagnoses affecting measured accuracy. Ultimately, EHR2Trace provides a reusable data foundation for developing patient world models and clinical agents.
The method proposes a moving-horizon approximate branch-and-reduce approach for training deep classification trees on large datasets with continuous features. It leverages a hierarchical root-subtree optimization framework, using branch-and-reduce for the root and greedy heuristics for subtree approximation. This approximation acts as a lookahead rollout, enhancing efficiency for deeper structures, and a low-cost strategy refines model accuracy iteratively. Numerical results show the method surpasses heuristic baselines in accuracy while offering superior scalability concerning dataset size and tree depth.
A data-free physics-informed neural operator is developed for level-set interface advection. This operator maps an initial interface to its full spatiotemporal trajectory under a prescribed flow. Training relies solely on the transport residual and a geometric constraint, without using any reference solutions. The method demonstrates significant performance gains over supervised baselines, achieving error reductions of a factor of 4.4 for a single vortex. The physics-trained operator conserves enclosed area better than the supervised baseline in certain conditions.
The Conformal Adversarial Generative Ensemble (CAGE) is a novel framework designed to improve time series forecasting by mitigating the influence of extreme forecasts. It integrates generative modeling, adversarial discrimination, and conformal prediction to enhance forecast reliability. CAGE uses multiple generative models to create initial forecasts, which are then assessed by a discriminative component employing conformal prediction. P-values from nonconformity scores dynamically adjust model weights, ensuring only credible predictions influence the final ensemble output. Empirical results on milk collection and public health data show CAGE outperforms traditional ensembles in handling noise and outliers.
DualCast is a dual-path language model framework designed for bimodal financial time-series forecasting. It utilizes a frozen language model augmented with a discrete financial vocabulary, representing log-return patches with summary tokens. A fast path trains embeddings on a frozen backbone, while a slow path uses a toggleable LoRA adapter to condition on external news for revised predictions. The slow path, optimized with a return-space group policy, achieved the lowest mean absolute percentage error in several evaluations. This approach effectively combines a fast numerical forecaster with an optional text-conditioned revision mechanism.
Diffusion maps and kernel methods offer an interpretable spectral representation for geometric learning. Dense Gaussian kernel methods require $O(N^2)$ memory, which is prohibitive for high-rank matrices. FlashDiffusion is a matrix-free method that evaluates dense Gaussian kernel blocks using fused GPU tiles. It couples the eigensolver with an empirical $\beta$-flow to select the finite-sample resolution scale. This approach warm-starts expensive spectral solves by continuing over coarser resolutions.
Alcohol intoxication is a risk factor for e-scooter crashes, necessitating real-time impairment detection. Researchers used sensor data from an instrumented e-scooter and a six-axis IMU to study kinematic signatures of impairment. Analysis of normalized permutation entropy and standard deviation revealed that intoxicated riders exhibit reduced temporal complexity and increased signal amplitude compared to sober riders. A machine learning classifier achieved 85% accuracy in distinguishing between sober and intoxicated states based on these kinematic features. This work establishes a method using onboard sensing and entropy analysis for automatic intoxication detection in e-scooter riders.
Test-time scaling requires effectively managing context allocation and reuse across multiple context windows during inference. The authors introduce Hermes, a family of harnesses, and Hermes-Learn, a two-stage framework, to enable models to make these allocation decisions. Capable models can utilize this flexibility to scale with extra compute, whereas smaller models struggle without training with Hermes-Learn. This training induces adaptive contextual reasoning strategies that generalize across benchmarks and transfer to other scaling methods. Ultimately, learning contextual reasoning unlocks the ability to scale model performance during inference.
Activation-Conditioned Self-Distillation (ACSD) extracts a steering vector by contrasting activations of self-generated trajectories leading to correct answers against other trajectories. A frozen base model applies this vector to predict the next token based on student-generated prefixes. The student learns from the next-token distributions of these prefixes, utilizing outcome verification for direction construction and calibration. Distillation is performed without needing problem-specific reference text or teacher parameter updates. ACSD achieves the highest mean accuracy across four mathematical benchmarks among evaluated methods.
The paper investigates the risk retained by a small reweighted training set in finite weighted least squares with a minimum-norm learner. It proves the exact law $\Gamma_d(n)=3-n/d$ for the specific range $\lceil3d/2\rceil\leq n\leq2d-1$. The guarantee is achieved by using selections that preserve the full feature span, employing balanced simplex anchors, and methods like positive-weight lifting. For a specific small budget $(d,n)=(5,6)$, the bound $\Gamma_5(6)=11/5$ is matched, with a sharpness of $6/5$ verified using circuit-based and geometric methods.
Weight space networks predict properties of neural networks by taking another network's weights as input. Prior work struggled with generalizing performance to networks of different sizes. Transferable Graph Metanetworks modify the model to improve performance transfer across networks of varying widths. The modifications are based on invariance and continuity principles for function representation. Empirical results show significant size generalization, especially when training under maximal-update parameterization, supported by infinite-width limit theory.
The research quantifies the gap between how much a causal predictor can improve and the gain actually learned during a causal repair in a Gaussian model. It uses auxiliary data to estimate effect magnitudes up to bounded contamination and diagnostics to identify direction in a causal experiment. Joint optimization of the learner and assessor under a uniform learning MSE helps avoid trivial solutions, showing a $k^{-2}$ assessment floor at the usual learning scale. A leading-log frontier is characterized by an assessment exponent of $\min{\ell_k,2k\eta_k/U}$ in the magnitude-rich regime. This work specifies how the assessment target influences information requirements in causal experiments.
ChainLoRA is a replay-free continual learning framework designed for large language models by merging task vectors in a geometry-preserving manner. It formulates forgetting through a geometric view of task updates, separating directional overlap from coefficient coupling. The method combines chain-updated training with post-stream adaptive SVD merging for parameter integration. Procrustes adaptation aligns a shared carrier to the latest task, facilitating the geometric separation of task-specific components. Experiments demonstrate that ChainLoRA achieves state-of-the-art performance on benchmarks while managing parameter budgets effectively.
Fractional Laplace neural operators (fLNOs) are introduced to embed the memory structure found in hereditary network dynamics, which are described by Volterra resolvents. For commuting excitation-Laplacian pairs, a block graph-spectral layer exactly represents the linear Volterra solution operator. The expressivity frontier shows that finite rational realizations approximate fractional memory geometrically but cannot capture non-integer critical asymptotics. The framework allows for trainable parametrizations that enforce stability margins and separate genuine operator consistency from parameter sharing.
Uncertainty quantification is necessary for black-box image regression models in high-stakes applications. Adaptive Conformal Prediction using Nearest Neighbors (ACPNN) is proposed as an input-adaptive conformal framework for image regression. ACPNN generates locally adaptive uncertainty estimates by leveraging neighboring samples defined by a Gaussian Process with an ARD kernel. The method is computationally efficient while maintaining reliable uncertainty quantification. This approach was demonstrated effectively on a diffusion model emulating inertial confinement fusion simulations.
A neural-network-based algorithm is proposed for contextual energy pricing to optimize demand response. This method models pricing as a Stackelberg game and uses a mean-field solution representation. The algorithm learns constrained mappings from contextual features, such as weather and time patterns, to feasible price signals. Simulations across US cities demonstrate that incorporating this contextual information significantly increases the value of demand response programs.
The research focuses on inferring parameters of stochastic differential equations driven by weighted sub-fractional Brownian motion using neural networks and the Euler approximation. The framework considers an SDE driven by a Gaussian process with a specific covariance structure related to a temporal weight function. The method uses the Euler approximation to reconstruct Gaussian driving increments and estimate a trajectory likelihood. Neural and radial-basis representations are employed to model the drift, diffusion, and temporal weight. The approach compares its performance against other neural methods for parameter estimation.
The study addresses a nonstationary stochastic bandit problem featuring latent linear dynamics where actions influence future state evolution and rewards are bilinear in the action and state. Previous methods achieve $\tilde{O}(T^{2/3})$ regret, but the authors propose an adaptive block-level optimism strategy. This involves approximating the infinite-memory reward process with a finite-memory proxy to facilitate planning. A UCB-based block algorithm is introduced to select blocks optimistically based on confidence sets for the truncated dynamics parameters. This approach yields a superior regret bound of $\tilde{O}(\sqrt{T})$ for this class of problems.