Navalia is a framework designed to advance Cowork Agents by incorporating verifiable task synthesis and long-horizon post-training. It uses corporate financial reports to semantically select evidence and construct computation chains for generating analytical tasks with traceable dependencies. Independent LLM review and programmatic checks are employed to verify evidence consistency and answer accuracy. These verified tasks are then used for trajectory collection, supervised fine-tuning, and reinforcement learning to enhance long-horizon execution. The method consistently improves evidence integration, tool use, and task execution across various benchmarks.
A method combining a large language model and reinforcement learning is proposed for autonomous construction and progress updates. An enhanced soft actor-critic reinforcement learning algorithm plans actions for a mobile crane based on construction tasks. Robot kinematics and control algorithms ensure the generated actions are executable by the crane. A customized embedding model extracts relevant information from the construction environment based on user interests. A large language model, such as ChatGPT 5, is then utilized to update progress, provide safety information, and generate reports, with feasibility verified through construction experiments.
The research proposes a method using multi-objective reinforcement learning with Large Language Models for Pareto-optimal Register-Transfer Level (RTL) code generation. Existing methods struggle to generate diverse solutions across multiple trade-offs like Power, Performance, and Area simultaneously. The proposed method trains a generation pipeline to optimize the coverage improvement of the entire Pareto solution set. A reward function based on Hypervolume Contribution (HVC) is used to measure a generated circuit's contribution to the Pareto front. Evaluations show the method significantly improves Hypervolume and circuit diversity compared to baseline methods.
MedPAO-Assist is an edge-deployable, modality-aware agentic AI framework designed for structuring radiology reports. It generates structured reports directly from radiologist dictation and voice commands without requiring external APIs or internet connectivity. The system is built using a Small Language Model and operates with low resource requirements, running on standard laptop CPUs. It supports four modalities, including Chest X-ray, Chest CT, Abdomen CT, and Brain MRI. The workflow involves concept extraction, ontology grounding, and structured report generation, culminating in a human-in-the-loop correction phase.
DataPraxis is a framework designed to scale analytics tasks by evolving workspaces with verified findings and artifacts. It progressively enriches workspaces and inherits these to uncover deeper questions, constructing tasks from accumulated, verified findings and their dependencies. The framework verifies analytical results during evolution and after task assembly. This process leads to the creation of the DataPraxis-5K dataset, which is used to train large models like DataPraxis-27B and DataPraxis-35B-A3B. These models achieve competitive performance on established benchmarks.
Large language models enable agents to perform autonomous planning and execution for smart manufacturing tasks. Challenges in this field stem from dynamic environments, physical constraints, and reliability demands in industrial settings. The paper systematically analyzes industrial agents, proposing a taxonomy and a four-layer system architecture for closed-loop execution. It details core modules and operational mechanisms for these agents. Finally, the work reviews applications across the industrial lifecycle and suggests future research directions.
Automated theorem provers in Lean often discard valuable error information during failed steps when using pure sampling. This work introduces a bounded, online repair module that samples tactics and repairs failed tactics based on the compiler's error messages. A pretrained repair model is used to fix individual rejected steps in place, rather than discarding them. This method improves the solve rate by up to 32% compared to sampling-only baselines. The repair strategy is most effective for shallow tactic-execution failures, yielding significant gains with predictable runtime trade-offs.
MAMath-Bench is a new benchmark designed to evaluate the multimodal mathematical reasoning of multi-agent systems. It assesses systems across three dimensions: problem-solving correctness, reasoning-process reliability, and robustness to adversarial perturbations. The benchmark includes 1,003 problems spanning 11 mathematical categories and modalities like text, images, and audio. Evaluation findings indicate that multi-agent collaboration does not always improve final accuracy but can enhance reasoning coverage. Furthermore, current systems remain vulnerable to errors caused by subtle structural perturbations.
Conversational recommendation systems using vision-language models face challenges in scaling to multiple product domains. LightCRS is a lightweight framework that unifies multi-domain multimodal conversational recommendation. It precomputes multimodal product representations offline and uses a compact, query-conditioned cross-attention module for online inference. The system achieves state-of-the-art performance on a 25-domain benchmark, significantly improving metrics while reducing query latency. This approach enhances performance by jointly training across domains and optimizing the inference process.
Large language models struggle with classical planning, but current benchmarks do not effectively separate structural complexity from semantic interpretability. The zero-training delete-relaxation heuristic $h_{\mathrm{add}}$ significantly outperforms other learned predictors for estimating structural difficulty. Renaming predicates demonstrates that LLM planning success degrades as the semantic complexity increases, while conditioning information predicts difficulty better than trained models. Hybrid planning relies more on calibrating to LLM outcomes than on structural scores. Therefore, improving hybrid planning should focus on modeling LLM behavior sensitive to predicate semantics.
Mechanistic studies of large language models focus on parameters but lack insight into the inference unfolding process. The ordered layer-state trajectory is examined to determine if layer order corresponds to a measurable computation during inference. Empirical results show that layer order aligns with endpoint-preserving shuffles in transformer checkpoints. Direct hidden states offer stronger mechanism coordinates than vocabulary readbacks for this process. Ultimately, layer order reveals an observable and computable aspect of transformer inference without a full theory of LLM dynamics.
A novel framework called Preference-PCA Pluralistic Learning (P3L) is proposed for pluralistic alignment using anonymized crowd feedback. This method learns trajectory embeddings and a shared basis of preference directions via PCA decomposition without needing annotator identities. The approach enables rapid and accurate personalized reward learning and policy optimization at test time. Experiments show that P3L facilitates steerable personal alignment that generalizes to new users and achieves competitive performance. Furthermore, the method reduces computational cost while supporting variable adaptation query budgets.
Large language models deployed in non-verifiable domains require assessment beyond simple binary rewards from reinforcement learning. Subject-matter expert (SME) judgment, which is graded, can be converted into dense reward signals. The method combines this SME judgment with retrieval-based metrics to create a hybrid training signal. This approach addresses the issue where sparse rewards fail to provide useful learning signals in such settings. Experiments demonstrated up to 6% improvement in retrieval metrics compared to sparse-reward baselines.
Capability-Aware Multi-Teacher On-Policy Distillation (CA-MOPD) addresses challenges in multi-teacher on-policy distillation by incorporating student capabilities into supervision allocation. It merges reinforcement-learning specialists into a strong student initialization before guiding distillation. CA-MOPD uses verifier-estimated teacher-student capability gaps to dynamically adjust domain-level distillation strength. Supervision is reduced when capability gaps close and increased when student capability regresses, while retaining dense token-level guidance. Across six benchmarks, CA-MOPD outperformed other methods, showing significant gains over naive MOPD and Open-MOPD.
Continuum-LM introduces a continuous-depth language model by treating pretrained residual transformations as dynamical operators. Hidden states evolve via ordinary differential equations solved by higher-order numerical integration, with solver settings controlling the computation. This approach allows for extending computational depth by reusing parameters beyond simple model scaling. Experiments across eight scientific tasks show Continuum-LM achieves significant accuracy improvements over discrete models. The results suggest that continuous hidden-state evolution is a practical dimension for scientific language models.
Two-layer transformers trained on a sorting task converge to one of two solution classes based on initialization. One class fails to length-generalize, while the other perfectly length-generalizes despite fixed-length training. This difference in generalization stems from how the attention layers interact with input types that become more common at longer sequences. The second class utilizes a mechanism where the second layer leverages contextual information from the first layer to narrow its search space. This research identifies a layer-wise specialization mechanism that allows the second attention layer to handle rare sequences effectively.
Latent world models allow agents to plan in latent space by minimizing the $\ell_2$ distance to a goal latent. The noise in the latent space is structured by a regularizer that promotes variance in signal-less directions. Fitting the state dynamics model $A$ yields the covariance, and its eigenbasis provides a stable basis for planning cost. Restricting the cost to the subspace with least uncertainty improves planning performance significantly. This method achieved gains up to $+54.7$ points on 2D navigation tasks.
Split Federated Learning (SFL) suffers from horizontal and vertical inconsistencies, leading to model divergence and client drift in asynchronous settings. CoASFL proposes a framework to address these issues by incorporating dynamic activation merging and a consistency-aware model aggregation mechanism. It ensures more consistent activation sequences from heterogeneous clients to reduce horizontal inconsistency. Furthermore, it dynamically adjusts aggregation weights based on client participation to mitigate vertical inconsistency. Experiments on an 80-device testbed show CoASFL significantly outperforms other SFL approaches in terms of speed and accuracy.
CorteXforge is a framework designed to generate realistic and reproducible Radio Frequency (RF) datasets using programmable Software Defined Radio (SDR) testbeds. It automatically translates high-level RF scenarios into synchronized experiments on platforms like SLICES-RI/CorteXlab. This process produces fully annotated Signal Metadata Format (SIGMF) datasets that include hardware impairments and propagation effects. Validation involved automatic modulation classification using an over-the-air (OTA) dataset generated by the framework. The work demonstrates that synthetic pretraining is beneficial for adapting to real-world OTA data, offering a foundation for reproducible RF data generation.
ORBIT is a unified deep generative framework designed for autoregressive time series generation by addressing challenges in learning hierarchical representations and coordinating heterogeneous temporal dynamics. It introduces two key components: the Frequency-Hierarchical VQ-VAE for learning hierarchical representations and the Frequency-Band Autoregressive Transformer for autoregressive prediction. The framework uses these components to decompose temporal representations into frequency bands and establish a low-to-high frequency autoregressive direction. This approach disentangles patterns and mitigates interference between different temporal dynamics during generation. Experiments show ORBIT achieves state-of-the-art performance with superior inference efficiency.
The paper determines the sample-complexity exponent for strict replicability when the realizable distribution and its marginal are unknown. It establishes that two independent samples and one shared seed must yield identical functions for strict replicability. For specific parameters, the worst-case sample complexity is shown to be $\Theta((\log M)^2)$ without iterated logarithms. Furthermore, the work provides a matching characterization $\Theta(d^2/(\varepsilon\rho^2)+\varepsilon^{-1}\log(1/\delta))$ for the full Boolean class $P_d$ using a proper Poissonized ball-sampling algorithm. The results involve lower bounds using square-root probabilities and volumetric packing, and an explicit Gibbs upper bound.
Uncertainty quantification in time series forecasting is challenging due to temporal dependence and distributional shifts. Dynamic Dual-feedback Conformal Inference (DDCI) is proposed to address this by incorporating a dual-feedback mechanism. This mechanism uses both actual feedback from true observations and estimated feedback from conformal inference to adjust prediction intervals. Actual feedback drives the primary interval adjustment, while estimated feedback moderates expansions or contractions using empirical thresholds. DDCI results in more stable and narrower prediction intervals while maintaining practical coverage.
The research focuses on robust estimation of a Gaussian precision matrix that has a sparse component minus a low-rank component, arising from latent variable marginalization. A proposed convex estimator uses a centered Gaussian measure-transformed covariance and penalizes the low-rank part using the nuclear norm. A primal-dual witness theorem guarantees that the raw optimizer is unique and recovers the true structure under certain conditions. The work establishes information-theoretic limits for estimating inertia and signed graphs, providing specific sample complexity bounds. Finally, robustness to Huber contamination is quantified, showing vanishing bias for remote outliers under the MT functional.
The study investigates using Gaussian kernel approximation and regularized logistic classification on smooth, connected, compact, embedded manifolds. The research addresses deriving a fast oracle inequality and constructing Gaussian approximants that leverage intrinsic dimension and regularity. A uniform Sobolev operator expansion is used, and Gaussian kernel integral operators are employed to cancel lower-order bias terms. This method yields an $L_2(X)$ approximation error of order $\gamma^\alpha$ and a squared Gaussian RKHS norm bound of order $\gamma^{-\varrho}$. The results establish bounds on the logistic excess-risk rate, which match the minimax optimal rate under certain conditions.
The study investigates the sample complexity for estimating Ornstein--Uhlenbeck diffusion scores using denoising score matching with shallow ReLU networks. It derives explicit Sobolev bounds by combining estimates from the Gaussian transition kernel and the initial density, which control approximation error and path-norm budgets. The work establishes high-probability risk bounds for empirical score minimizers using truncation and localized Rademacher complexity. Under specific assumptions, the resulting squared $L^2(P_t)$ score risk is $\widetilde O(n^{-(d+3)/(2d+3)})$. Furthermore, aggregating errors leads to a total-variation error bound of $\widetilde O(n^{-(d+3)/(4d+6)})$ for generated versus original data distributions.
The paper studies the communication complexity of representing data that allows for evaluation with a family of functions. It constructs a deterministic encoder that outputs a representation of size $O(d^{*}\log(1/\varepsilon))$ bits for a task $f$ with error $\varepsilon$. The encoder greedily selects functions based on a novel potential function argument comparing entropy gain to pattern count. The lower bounds show the representation length is worst-case optimal for $d^* \ge 1$ and $\varepsilon < 1/4$. The work also provides methods to learn this representation from unlabeled samples with guarantees on size and sample complexity.
Evaluating Large Language Models (LLMs) requires multi-objective assessment across various performance dimensions. The paper introduces Vector-Borda Empirical-Gap Elimination (VB-EGE), an adaptive sampling algorithm for efficiently finding models on the Pareto frontier in arena-style evaluations. VB-EGE progressively removes models once sufficient statistical evidence confirms their Pareto status. The method operates under a multi-objective Bradley-Terry model and provides instance-dependent bounds on sample complexity. Experiments on real-world datasets confirm VB-EGE's empirical effectiveness.
Alarms in industrial systems require explanations tracing changes in the data-generating process to the alarm score, despite scarce incident data. The paper formulates the alarm-score change as a Shapley game over hybrid structural causal model (SCM) states to attribute propagation. It distinguishes between changes in structural functions and exogenous laws for a finer-grained report of influence. Exact Shapley vectors are computed in polynomial time for linear SCMs and quadratic scores using first- and second-moment recursion. The method adapts a reference SCM to estimate these changes from limited data, providing attribution-error bounds.
Bayesian neural networks are introduced to quantify epistemic uncertainty, addressing limitations of standard weight-space priors in deep reinforcement learning. The method places Bayesian nonparametric Dirichlet Process priors directly on the data-generating distribution instead of network parameters. This leads to the development of DP-DQN, a reinforcement learning algorithm that uses these DP-BNNs for value function uncertainty. DP-DQN combines data-driven reweighting with generative prior hypotheses for deep exploration. This approach avoids the computational cost of multi-network ensembles.
Offline reinforcement learning is limited by data coverage, prompting the study of methods to acquire information without direct environment interaction. Query-based collaborative reinforcement learning introduces a method where a learner queries a collaborator who samples actions based on a fixed policy. This approach is used to efficiently obtain complementary information from the offline dataset for tasks like collaborator-mediated simulation. For finite-horizon problems, the work derives information-theoretic lower bounds and develops an adaptive query-allocation algorithm. Experiments confirm the empirical efficiency of this method across various synthetic and semi-synthetic environments.