The study systematically analyzes token consumption patterns in agentic coding tasks using eight frontier LLMs on SWE-bench Verified. Agentic tasks are found to be uniquely expensive, consuming 1000x more tokens than code reasoning, with input tokens dominating the cost. Token usage is highly variable and stochastic, and accuracy often peaks at intermediate costs rather than higher ones. Models exhibit significant differences in token efficiency, with some models consuming substantially more tokens than others on the same tasks. Furthermore, frontier models generally fail to accurately predict their own token usage.
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 to generate analytical tasks with traceable data dependencies. The framework employs independent LLM review and programmatic checks to ensure evidence consistency and answer correctness. Verified tasks are then used for trajectory collection, supervised fine-tuning, and reinforcement learning to improve long-horizon execution. Experiments show Navalia consistently improves evidence integration, tool use, and task execution across multiple 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 ensure the generated actions are executable by the crane. A customized embedding model extracts relevant environmental information, which is then processed by a large language model like ChatGPT 5 for updates and reporting. The feasibility of this autonomous method was verified through full-scale retaining wall construction experiments.
The research proposes a method using multi-objective reinforcement learning and Large Language Models for Pareto-optimal Register-Transfer Level (RTL) code generation. Current methods struggle to generate diverse solutions across multiple trade-offs like Power, Performance, and Area simultaneously. The proposed method optimizes the quality and diversity of the entire circuit set rather than a single circuit's quality. It utilizes a generation-based training pipeline and a Hypervolume Contribution (HVC) reward function to guide the LLM. Evaluation shows significant improvements in relative 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, requiring minimal resources suitable for offline clinical use. The system supports four modalities, including Chest X-ray, Chest CT, Abdomen CT, and Brain MRI. It utilizes a small language model and an Extract-Ground-Generate architecture for concept extraction and report generation. This workflow enables a "no-touch" reporting process requiring no external connectivity or keyboard input.
DataPraxis is a framework that scales analytics tasks by evolving workspaces with verified findings and artifacts. It progressively enriches workspaces and inherits these validated findings to uncover deeper analytical questions. The framework constructs tasks from accumulated findings and dependencies, ensuring analytical results are verified during evolution and assembly. This method was used to create the DataPraxis-5K dataset, comprising 5,000 verified trajectories. Training on this data resulted in DataPraxis-27B achieving competitive performance on analytics benchmarks.
Large language models enable agents to achieve autonomous planning and execution, presenting opportunities for smart manufacturing. Challenges in dynamic environments and physical constraints limit the direct use of general-purpose agents in this sector. The paper systematically surveys LLM-driven industrial agents, establishing a taxonomy and proposing a four-layer system architecture for closed-loop execution. It details the core modules and operational mechanisms for these agents. Ultimately, the work provides a structured reference for designing, developing, and applying industrial agents across the manufacturing lifecycle.
The research introduces a bounded, online repair module for theorem proving in Lean 4 to address the limitations of pure sampling. This module samples candidate tactics and repairs failed tactics by feeding the compiler's error messages into a pretrained repair model. This approach improves the solve rate by up to $32\%$ over sampling-only baselines on the LeanDojo benchmark. Repair is most effective for shallow tactic-execution failures, yielding significant gains with predictable runtime costs. The method successfully doubles the solve rate on the Lean Workbook benchmark without introducing regressions.
MAMath-Bench is a new benchmark designed to evaluate the multimodal mathematical reasoning of multi-agent systems. It includes 1,003 problems across 11 mathematical categories, incorporating text, image, and audio modalities. The evaluation assesses problem-solving correctness, reasoning-process reliability, and robustness to adversarial perturbations. Findings indicate that multi-agent collaboration does not always improve final accuracy, though it enhances intermediate reasoning coverage. Current systems also show vulnerability to adversarial reasoning errors.
LightCRS is a lightweight framework for unified, multi-domain multimodal conversational recommendation. It precomputes multimodal product representations offline using frozen encoders to manage complexity. Online, it uses a compact module with query-conditioned cross-attention to fuse this evidence across domains efficiently. The system compares candidates jointly using an inter-candidate Transformer for recommendation. Evaluation on a 25-domain benchmark shows LightCRS achieves state-of-the-art metrics while significantly reducing per-query latency.
Large language models struggle with classical planning, but existing benchmarks do not effectively separate structural difficulty from semantic interpretability. The zero-training delete-relaxation heuristic outperforms learned alternatives in predicting structural difficulty across tested domains. Renaming predicates significantly impacts LLM planning success, demonstrating sensitivity to semantic information. LLMs utilizing conditioning information based on semantic changes predict difficulty better than models trained solely on structural features. Therefore, improving hybrid planning requires modeling LLM sensitivity to predicate semantics rather than focusing only on structural predictors.
Disaggregated LLM serving separates prefill and decode across node pools, complicating classical cache retention policies due to network transfer costs. KVLearn is a learning-based framework designed for cost-aware KV-cache retention in this disaggregated environment. It incorporates a Prefix Reuse Predictor, a Cost-Aware Retention Score, and an Adaptive Threshold Controller to manage decisions. Evaluation shows KVLearn significantly reduces time-to-first-token and inter-node KV transfer volume compared to existing baselines. The framework effectively manages heterogeneous reuse patterns in multimodal workloads while maintaining high throughput.
Interaction-Gated Time-Frequency Learning (TF-IG) is a parameter-efficient framework for estimating heart rate from Photoplethysmography (PPG). It uses a dual-branch architecture to capture temporal pulse morphology and frequency-domain periodicity separately. A novel interaction-gated fusion module explicitly models feature agreement and discrepancy between these domains. Experiments on benchmark datasets yielded strong results, achieving a mean absolute error of 1.227 BPM. The method demonstrates an effective trade-off between accuracy and model complexity.
Reinforcement learning facilitates adaptive decision-making in computing systems deployed on FPGAs and ICs, despite resource and complexity challenges. This survey reviews over 100 studies on RL implementations across these platforms, covering various algorithms like model-based and deep RL. It classifies research based on RL family, hardware realization, and application domains, providing a unified analysis of acceleration and optimization. The review compares performance metrics such as latency, power, and resource utilization in FPGA and IC contexts. Finally, it identifies key architectural patterns, limitations, and future research directions for RL-driven hardware systems.
A simultaneous neural topology optimization (SNTO) framework is proposed to address the computational demands of large-scale, stress-constrained topology optimization. This method uses custom neural networks to represent state, adjoint, and design variables, optimizing them jointly via a single loop using continuous adjoint analysis and automatic differentiation. The framework minimizes a loss function combining the design objective, stress constraints, and potential energy functionals to achieve super-resolution designs. Scalability is achieved through a hardware-aware training scheme that partitions the design domain across multiple GPUs, ensuring field continuity without adding extra loss terms. Validation on benchmarks showed SNTO yields sharper boundaries and lower volume fractions compared to finite element methods.
Reward hacking in reinforcement learning with LLMs is difficult to prevent during training due to evolving environments and model obfuscation. SimHack is introduced to predict when detectable reward hacking will become dominant before starting an RL run. It learns reward hacking training dynamics from smaller reference runs to generate counterfactual simulations of unseen recipes. SimHack accurately predicts the median onset of hacking to a relative error of 11.3% across different target recipes. This capability allows guiding training decisions to discover curricula that achieve better performance without hacking.
A Set Transformer model is proposed to estimate X-ray transmission in polymer composites by mapping molecular constituents and physical descriptors to the transmission ratio. The model utilizes chemistry-pretrained transformer embeddings from ChemBERTa for each constituent. Training data is generated via a PENELOPE/penEasy pipeline simulating X-ray transmission in polyethylene-based composites. The method achieves a low mean squared error of $(1.12 \pm 0.56)\times10^{-4}$ and $R^2 = 0.9988$ in cross-validation. This approach allows for predictions in milliseconds, significantly faster than Monte Carlo simulations.
LLM-based methods for node classification on text-attributed graphs report state-of-the-art accuracy. The study compares these results against inexpensive baselines using validation-only selection for baseline tuning. Some LLM methods remain significantly better than simple baselines across many tasks. However, evaluation risks exist due to unstated hyperparameter choices, label usage, and configuration mismatches between studies. Therefore, reported gains need careful scrutiny when comparing different LLM approaches.
Offline Apprenticeship Learning infers policies from expert demonstrations to enable data-efficient learning in human-centric environments. Frameworks are developed for reward modeling, using clustering and temporal partitioning of demonstrations, and for demonstration quality, incorporating quality signals. Methods like HALIDE and CCHAL attribute quality to decisions using Bayesian methods and contrastive objectives. Classroom evaluations show EDM-induced policies lead to significant learning gains over expert baselines. This work establishes offline apprenticeship learning as a practical and ethical paradigm for complex systems.
ReAPO proposes Retry-Aware Policy Optimization to improve efficient exploration in Large Language Model (LLM) reasoning. It frames credit assignment using Stochastic Shortest Paths to guide exploration beyond standard independent stochastic rollouts. ReAPO allocates generation to failed trajectories, reinforcing successful retries without rewarding preceding failures. Experiments on Qwen3 and OLMo-2 models show ReAPO improves first-attempt performance over GRPO and baselines. This method effectively enhances LLM reasoning through improved exploration and retry-aware credit assignment.
Motion becomes a gauge symmetry in sequential data because changes in data are due to motion, which is a symmetry of the data. Standard recurrent networks lack equivariance to motion because their memory is not sensitive to input movement. The proposed Motion Equivariant Convolutional LSTM (MELSTM) treats motion as a gauge field, making the network provably equivariant to any motion. MELSTM outperforms existing models in tasks like future-frame prediction and action recognition while using fewer parameters. This approach allows for more accurate and generalizable modeling of sequences involving motion.
Retailers face challenges estimating local demand for assortment and pricing due to limited store-specific data. This study investigates using transfer learning to borrow strength across stores for choice-based revenue management decisions. A shrinkage estimator is proposed to combine store-specific estimates with pooled information from other stores under a multinomial logit model. The research establishes performance guarantees for this shrinkage and derives regret bounds for assortment and pricing. Empirical results show that shrinkage improves out-of-sample prioritization and revenue capture compared to local estimation or full pooling.
The work presents a method for positive-data identification using substrate-bounded contrastive learners. It establishes a "Reverse-Gold Theorem" showing that a contrastively trained scorer can achieve PAC conditional identification from positive data alone. This identification relies on an innate recursive contrast operator and a discriminative scorer trained via InfoNCE-style contrastive estimation on self-generated perturbations. The results yield polynomial sample complexity for the identification, with bounds derived from the geometry of the contrastive loss. The framework is instantiated for context-free languages and subregular classes, leading to an open conjecture.
The paper revisits depth separation for neural networks in the context of $L^2$ approximation of radial functions under the standard Gaussian measure. It demonstrates that a depth-3 ReLU network can approximate such functions within a small $L^2$ error using polynomial size. However, approximating these functions with depth-2 networks requires exponentially large weights unless the network size is also exponential in the dimension $d$. A general result shows that functions with significant high-frequency mass in their Hermite expansion resist approximation by shallow networks with bounded weights.
The study investigates whether in-context reinforcement learning (ICRL), particularly Decision-Pretrained Transformer (DPT), functions as a Bayesian posterior sampler in practice. The researchers establish two necessary conditions for a Bayesian posterior sampler: sufficient-statistic invariance and the martingale property. Statistical hypothesis tests were used to assess if practically-trained ICRL models satisfy these conditions across Gaussian, Bernoulli, and linear bandit settings. The results consistently show that both conditions are violated, suggesting ICRL may not behave as Bayesian posterior samplers in practice. The work provides a framework for testing the Bayesian properties of ICRL.
Internal guidance is proposed as a sampling method for tilted diffusion models, diverging from existing external guidance methods. It modifies the evaluation point of the base denoiser rather than the output itself. For discrete distributions on the hypercube, the denoiser of a linearly tilted density is the base denoiser evaluated at a shifted and rescaled input. The training algorithm is modified with random diagonal rescalings to allow the denoiser to learn both in- and out-of-distribution evaluations. This approach enables a single denoiser to sample binary reward tilts, class labels, and infilling without fine-tuning or external guidance.
Barycenter models, based on optimal transport bridges, are useful for learning with measure-valued data. The paper introduces four novel barycenter models—SBB-encoder, SBB-decoder, SBB-symmetric, and SBB-Markovian—derived from the Schrödinger-Bass Bridge. Theoretical results establish the existence of the SBB-barycenter and characterize the Gaussian decoder. A hybrid Wasserstein gradient particle descent scheme is derived for the four formulations using the SBB dual representation. These methods successfully solve distributional-inverse/interpolation problems in data analysis, outperforming existing models.
The paper proposes a certified rule for determining if unseen data belongs in a pretrained model's context or its weights based on information value. This value is defined by a formula involving saved code length, query information, and decoding loss. The law provides ceilings that certify the choice from a single measurement, with an anytime-valid test for error control. The method successfully diagnoses when in-context learning is helpful and evaluates context modifications by tracing learning curves. This framework shows that observing specific features in contexts can significantly alter model behavior.
A unified framework is developed connecting proper scoring rules and Stein's method for estimating unnormalized models. This involves constructing a class of proper scoring rules, including score matching and ratio matching, that naturally induce Stein operators. This connection provides a new construction principle for Stein operators, allowing for the study of estimators based on these rules through associated geometry. This geometric perspective yields a unified characterization of the asymptotic efficiency for the resulting estimators.
Residual-based causal discovery assumes variables are generated from parents plus independent noise, allowing causal direction to be identified by testing independence of noise from causes. The RESIT-framework enables nonparametric residual estimation for identifying sink nodes in additive-noise models. Loscar extends RESIT to location-scale noise models by using quantile regression to estimate conditional mean and scale. This approach incorporates pretrained tabular foundation models as plug-in estimators for residuals. In benchmarks, in-context quantile regression significantly outperforms feasible generalized least squares methods for causal discovery.
The study focuses on estimating quantile distributional policy evaluation by approximating the return distribution with the quantile Bellman fixed point $\bm\eta_m$. The authors investigate the statistical accuracy of estimating $\bm\eta_m$ across different quantile resolutions under Wasserstein distances, assuming access to a generative model. They establish non-asymptotic error bounds for this estimator, showing resolution-independent rates under certain reward density conditions. The analysis further characterizes the local asymptotic minimax risk, revealing resolution-dependent growth rates for specific scenarios.