A production decision system is presented for trade-aware customer-to-seller allocation in enterprise go-to-market planning. It integrates a multidimensional account-value layer, an observational continuity-valuation layer, and a multi-objective optimization layer. This framework allows planners to inspect and compare allocation tradeoffs among growth, balance, capacity, and feasibility. Historical backtesting demonstrated that optimizer-generated scenarios improved assignment compliance while maintaining opportunity distribution. Ultimately, the system guides the allocation of scarce seller capacity toward higher-value accounts for increased revenue potential.
Phronix introduces a capability-centric learning paradigm for agents to acquire, validate, and refine reusable skills from their own executions. It operates via a predictor-validator-corrector loop where proposed solutions are validated by tool executions, and outcomes refine stored capabilities. This process results in learned capabilities that solve recurring tasks on the first attempt with high accuracy, reducing solution non-determinism. Phronix efficiently generates durable capabilities from agent executions and significantly reduces the token cost for warm-start runs. Empirical results show Phronix improves task-solving rates and performance in complex performance engineering workflows.
Patronizing and condescending language (PCL) is harmful language that belittles vulnerable groups by assuming pity and reinforcing negative stereotypes. The research proposes automated systems to detect this type of language in both English and Mandarin. Two detection approaches were tested: one used static word embeddings and a convolutional neural network, and the second used fine-tuned pre-trained BERT embeddings. The BERT-based approach proved more successful, achieving an F1 score of 0.528 on the development set. This work addresses the difficulty humans have in detecting PCL despite its harmful nature.
PaperFormalizer is a Lean-based agentic pipeline designed to automatically formalize research papers by authors. It processes LaTeX source material into a dependency graph, audits statements for correctness and ambiguity, and formalizes them using Lean and Mathlib. The process involves interactive steps where the author reviews and approves proposed formalizations. The system generates an annotated paper, a change report, and linked Lean code. This aims to make research based on mathematical proofs more robust and shareable.
Intelligent agents require a method to learn from mistakes to achieve continuous improvement, as current in-context reflection and fine-tuning methods are inefficient. Self-Direction introduces a mechanism where agents reflect on failures to extract lessons and distill them into lightweight low-rank adapters. This process keeps base model weights frozen, training only a small percentage of adapter parameters. Learned lessons are stored in a persistent library, allowing adapters to be automatically applied to solve related future problems. Experiments demonstrated that Self-Direction significantly improved agent success rates and fostered the transfer of learned knowledge across diverse tasks.
TACS-RF presents a framework for tool-driven, agentic circuit synthesis of RF and mmWave modules. A large-language model agent drives industry-standard EDA tools like Cadence and Ansys HFSS to perform the design. The framework was validated by the agent designing a 240-251 GHz sub-THz transmitter on a 90 nm SiGe BiCMOS process. The agent autonomously designed complex sub-blocks including VCOs, power drivers, and antenna structures. The resulting design achieved +5.1 to +6.3 dBm output power with specified phase noise and power consumption.
The study investigates the execution-time reliability of tool-using AI agents, focusing on recovery decisions separate from language-model behavior. It defines distinct obligations for a durable agent, including checkpoint state, permission to dispatch, evidence of external effects, and permission to reconcile. A finite conformance study was conducted across multiple histories on LangGraph and Temporal hosts, checking these obligations. Results indicate that cached retry, revalidation before retry, and stable-key retry successfully pass a portion of the recorded histories. Furthermore, a component study on an enterprise AI platform isolated the roles of authority, provider contracts, and observations in recovery scenarios.
Distilling interactive LLM agents requires choosing between on-policy and off-policy distillation, which corresponds to teacher-generated demonstrations versus student-generated experience. The authors propose a hierarchical agent distillation framework to explicitly control the pairing, mixture proportions, and training schedule of these two levels of on-policiness. Empirical study suggests that student-induced histories paired with teacher action prefixes offer the strongest standalone training signal. Following the student at both levels generally performs poorly, although mixtures can be beneficial. Practical principles are derived for determining where, how much, and when agent distillation should follow the student.
$M^3$ is a multi-LLM framework that formalizes multi-turn jailbreaking as a finite-horizon Markov Decision Process. It uses a state LLM to maintain a compact internal state while an attacker LLM interacts with a target LLM. Empirical results show $M^3$ significantly outperforms state-of-the-art multi-turn red-teaming methods in Attack Success Rates (ASR). The system demonstrates improved effectiveness and more consistent performance across risk categories without requiring predefined taxonomies. This provides a scalable and principled approach for multi-turn safety evaluation.
LLM agents in repeated multi-party negotiations face challenges due to a gap between knowing who to trust and reaching an agreement. The research investigates whether LLM agents can form useful trust estimates from prior interactions to improve future outcomes. A framework is introduced where agents assess pairwise trust after each negotiation game to account for adversaries. Results show a trust-agreement gap, indicating that trust estimates do not always lead to better agreements on their own. Explicit guidance on using trust can increase agreement rates by up to 23 percentage points.
A surrogate-accelerated Bayesian history-matching framework is presented for joint 4-D seismic and production data assimilation. It employs an autoregressive Physics-Informed Neural Operator (PINO) to model the forward map, incorporating black-oil residuals as a penalty. Adaptive regularised ensemble Kalman inversion (aREKI) is used to assimilate data, followed by an ensemble smoother for joint estimation. The method ensures stability and tractability by carefully constraining the inversion space, leading to significant computational speed-ups. The approach demonstrates high accuracy and efficiency in history matching.
The study examines the early training dynamics of nonlinear residual networks modeled with zero hidden layers to understand general residual networks. The research uses a sequential limit where network width, depth, training data size, and input dimension all diverge. A key finding is that in the limit of large input dimension, the network output can only learn a linear function of inputs early in training. This suggests that mean-field limits of ResNets are rigid and may be insufficient for feature learning in high-dimensional regimes. The work also investigates methods for transferring learning rate hyperparameter based on dataset size and input dimension.
The paper investigates the PAC-learnability of function classes mapping inputs to conditional probability distributions over a finite space. It introduces the fat$_\gamma$Natarajan dimension, which is shown to be a necessary and sufficient condition for learnability when finite at every positive scale. A fat$_\gamma$Graph dimension is also introduced, along with inequalities comparing it to the former, establishing an equivalent characterization of learnability. The authors further demonstrate that a distribution-valued function class is learnable if and only if its coordinate classes are learnable in the probabilistic concept model.
Offline maximum entropy reinforcement learning estimates the optimal soft action-value function $Q^*$ from fixed datasets without further interaction. The proposed framework uses conditional diffusion models based on probability flow ODEs to estimate the conditional reward distribution and transition kernel. These distributional estimators approximate the soft optimal Bellman operator, which is then minimized over deep neural networks to estimate $Q^*$. The method establishes nonasymptotic guarantees for operator learning and value learning, providing an excess Bellman residual risk bound. Numerical experiments confirm the theoretical effectiveness of this approach.
A neural exploratory framework is proposed for variational problems over graphons, which are limits of dense graph sequences. Graphons are represented by implicit neural networks, and objective optimization is performed using gradient descent. The framework incorporates a multi-scale residual architecture, an embedded solver enforcing density constraints, and symmetry-aware Monte Carlo estimators. This method successfully finds optimal graphons for generalized Turán problems and variational problems from large deviation theory. The work yields new candidate optima and theorems regarding optimal graphons for these problems.
The study investigates federated nonparametric regression for clients with heterogeneous data distributions and estimation uncertainties. It addresses whether a federated predictor benefits an individual client based on population heterogeneity and uncertainty. The method allows clients to transmit local kernel summaries so the server can construct a covariate-density-weighted estimator without centralizing individual data. Client gains are characterized, and the framework provides pointwise inference for the federated estimator. Robustness is analyzed under perturbation models, using Kullback-Leibler and $L_2$ neighborhoods to assess variance-sensitive and exact fitting properties.
The work establishes empirical concentration inequalities for stationary chains with infinite memory, extending previous results from Markov chains. It develops a bounded-differences inequality on bounded Polish metric spaces using a summable Wasserstein contraction condition. Dependence coefficients are estimated from a single trajectory using transition counts or kernel conditional-density estimates. The procedure yields empirical concentration bounds that are asymptotically as tight as their theoretical counterparts. These bounds are fully adaptive, requiring no prior knowledge of the inequality's constants.
The study investigates squared-error estimation when a nuisance parameter is unidentified in a submodel affecting a target that vanishes there. Sufficient conditions are provided for the largest critical-path minimax risk to equal the exact asymptotic minimax risk over the full parameter space when using higher-order profile expansions. An estimator achieving the leading constant uniformly is possible under conditions involving efficient observable features and stable recovery. The method employs calibration estimates, a local feature mean, and an empirical recovery feature. The results are verified in models involving logistic errors-in-variables regression and replicated-network models.
The research investigates online learning of state-dependent model-error corrections within the Integrated Forecasting System using 4D-Var. It proposes a batched extension for this approach, processing assimilation windows in random order and averaging parameter increments like in mini-batch optimization. Experiments show that the batched formulation consistently improves forecast skill in out-of-sample tests, unlike sequential online learning which fails to generalize. Larger batch sizes enhance convergence and robustness by mitigating parameter update noise. The findings suggest that the formulation of the online learning problem is crucial for generalization in hybrid machine-learning/data-assimilation systems.
A method using a linear surrogate approximates Gaussian Process performance in high-dimensional test functions. While this approach yields exact acquisition for spherical decision regions, the unit cube is more common in Bayesian Optimization. The paper investigates the complexity of globally solving acquisition subproblems over the unit cube using standard multi-start quasi-Newton methods. A new acquisition function is proposed that is globally optimizable faster than existing methods while maintaining or improving performance. This resulting method is competitive with GP-based BO in accuracy per evaluation and superior in accuracy per walltime.