Query clarification is necessary in Retrieval-Augmented Generation (RAG) systems when user queries are ambiguous or contextually unclear. The study addresses scenarios where queries are ill-formed or have multiple relevant interpretations based on retrieved documents. A new dataset, ClariEdge, containing ambiguous and non-ambiguous queries is introduced for this task. A lightweight fine-tuned LoRA model achieves high classification accuracy, outperforming larger models in clarifying query quality. Hard negative examples are shown to be crucial for distinguishing truly ambiguous queries from those that are merely contextually complex.
Continued pre-training across data volume, diversity, and model size was investigated for embodied agents in diverse virtual worlds. Results indicated that more data improves in-corpus performance, diversity helps transfer to related worlds, and model size affects data efficiency and generalization to distant worlds. An early interface phase followed by a later world phase was observed, aligning with an empirical scaling law for validation loss. Models proved competitive in 3D settings, yet limitations in imitation suggest further training stages are necessary for generalist agents.
SpecEvo is a multi-LLM framework designed for cost-efficient scientific discovery by coordinating lightweight and frontier models. It uses lightweight models for parallel speculation and evidence consolidation, while frontier models intervene strategically at sparse checkpoints. This workflow delegates routine exploration to inexpensive models, reserving high-level reasoning for crucial interventions. Results show SpecEvo outperforms baselines while reducing API costs by up to 80% across various discovery tasks. The method demonstrates effective trajectory-conditioned model coordination for efficient scientific exploration.
Frontier coding agents are tested on the demanding task of improving published scientific models, which requires domain insight beyond standard ML engineering. Agent4Sci is introduced as a long-horizon testbed to measure the extent to which these agents can advance scientific ML benchmarks in real-world conditions. Empirically, agents achieved relative improvements by using engineering tricks rather than deep research insight. Trajectory analysis revealed limitations such as novelty miscalibration, resource under-utilization, and rapid regression to local engineering sweeps. The work proposes lightweight plug-in skills to bridge the gap between ML engineering and domain insight for these agents.
Agentic harness evolution uses task execution experiences to optimize an agent's instructions and tools. A failure mode emerges where expanding the task pool causes existing methods to degrade due to suboptimal evidence selection. Successive Scale Racing (SSR) is proposed to address this by adaptively searching for optimal evidence combinations in each evolution round. SSR races candidate harnesses based on task quality and coverage within a limited compute budget. Empirically, SSR achieved state-of-the-art results across benchmarks while improving performance as the task pool increased.
Decision language models (like Jev) offer an affordable and rapid alternative to using large language model judges for clinical generative AI. Comparisons showed Jev's agreement with physician quality ratings matched or exceeded that of several LLM judges across various clinical evaluations. Frontier LLM judges were significantly more costly and slower than Jev. This suggests that decision language models can enable efficient clinical AI evaluation at scale. Ultimately, Jev demonstrates potential for streamlining the assessment of clinical generative AI.
Sketch2Sim is an end-to-end modular framework designed to analyze hand-drawn circuit diagrams. It converts raw sketches into SPICE-compatible netlists and automatically calculates nodal voltages and currents. The system employs a multi-stage pipeline involving YOLO detectors, a CRNN for symbol extraction, and image processing for node identification. A dataset of 1,000 RLC circuits was used to train the system, demonstrating robust performance across different handwriting styles. This framework aims to enhance education and expedite engineering design reviews.
Relational foundation model pretraining is decentralized through speedrunning leaderboards to overcome barriers in accessing private training recipes. The work introduces modded-nanoRT, a leaderboard for Relational Transformers, which fixes the benchmark by calibrating the selection based on learning curve separation. The method utilizes a relational operator, typed trilinear attention, along with Muon and SIGReg. The resulting recipe achieves a target in 305.63 seconds on one GPU, significantly faster than the reference. The authors plan to release the benchmark, frozen artifacts, and record runs for public use.
The research introduces a controlled LLM-LLM debate framework to study opinion dynamics in extended conversations. Two language models, GPT-6 Luna and Gemini 2.5 Flash Lite, are used to discuss climate-change attribution, with one model assigned a stance and the other not. Opinion dynamics varied based on model combinations, as same-model interactions narrowed the opinion gap while cross-model interactions widened it. Explicit reports and in-dialogue behavior reveal different aspects of this dynamic, and a simple persuasion instruction was not consistently effective. These findings indicate that conversational influence depends on the participating models, not just assigned roles and initial positions.
An agentic system named Fig automates the elicitation of behaviors in multiagent systems by performing targeted interventions on game environments. It takes natural language behavior descriptions, decomposes them, and generates interventions to influence agents, opponents, or context. Evaluation combines mechanical checks and synthetic rubric judgments, validated through human review. Fig successfully elicited target behaviors, achieving an average attempt rate of 45.2% across six behaviors. The system demonstrates that multiagent environments are effective programmable testbeds for controllable behavior evaluation.
MoEZip is a method for routing-aware KV cache compression specifically for sparse Mixture-of-Experts (MoE) language models. It addresses the issue that standard compression methods neglect how KV cache eviction affects expert routing, which can degrade generation quality. The method introduces a novel metric based on the Fisher information matrix of routing distributions combined with attention weights for context scoring. MoEZip integrates complementary scoring systems to handle both dense and sparse MoE architectures effectively. Experiments demonstrate that MoEZip significantly outperforms the state-of-the-art by improving scores on LongBench and RULER benchmarks under aggressive cache compression.
Estimating neural sources from M/EEG measurements is an ill-posed inverse problem where uncertainty quantification is challenging. The proposed spectral framework uses the singular value structure of the prior-preconditioned forward operator to decompose the posterior covariance. This decomposition separates uncertainty into data-informed and fundamentally irreducible epistemic components. A subtractive decomposition is also derived, expressing the prior covariance as a sum of posterior and data-driven reduction terms. These methods provide interpretable uncertainty quantification for linear-Gaussian Bayesian inverse solvers in M/EEG source imaging.
Predictive accuracy in knowledge tracing models does not guarantee the recovery of item parameters. The study investigates item parameter recovery and representation coupling using various item response decoders and encoders. Findings indicate that while difficulty and mean threshold recovery is robust, discrimination and option slope recovery depend heavily on the scale and spread of estimated ability. Separating the sequence representation from the item parameter key improves recovery and cross-order consistency without sacrificing test accuracy. Ultimately, item parameters like discrimination must be evaluated separately from response prediction.
Large Language Models can be augmented by direct representation exchange, referred to as the Latent Interconnect Protocol (LIP). LIP learns connectors that map specialist representations into the LLM's internal states, enabling information exchange without joint training. The protocol successfully connected protein and small molecule specialists to Gemma 3 and Kimi K2.6 models. Independently trained connectors allow for zero-shot transfer to cross-domain tasks. The work suggests specialist developers and LLM developers must collaborate to establish consistent representation interfaces for scientific agents.
Learned multimodal representations for health records are often trained for reconstruction, which may not suffice for causal identification. The study examines counterfactual outcomes and interventional effects under exposure-induced confounding using these representations. It demonstrates that interventional mediation is more demanding than counterfactual comparison in this context. No training objective, including reconstruction, can guarantee a representation's sufficiency, as insufficient representations can lead to large bias despite low reconstruction error. A workflow combining specific methods is proposed to address representation insufficiency and provide diagnostic checks.
Exploratory study of prompt styles on GPT-5.5 across tasks from the International Olympiad in Informatics and NP-hard speed tasks was conducted. No single prompt style was found to be optimal, as the best style depended on the specific task and result summarization method. Differences in accuracy stemmed from outlier runs, indicating that single metrics can obscure the true effect of prompting. Comparing results to human contestants showed the AI was neither strictly better nor worse than humans across all tasks. Extending the comparison to runtime speed and human performance provides a more comprehensive view of prompt style's impact.
GANet is a lightweight Group Attention Network developed for identifying plant diseases on embedded systems. It utilizes novel Group Attention and Group-Wise Feature Fusion blocks to achieve sparsity and efficiency. The model requires only 0.001M parameters and achieves 99.99% accuracy on the PlantVillage dataset. GANet is optimized for deployment on edge devices like Raspberry Pi 4b and Jetson Nano. This allows for real-time plant disease monitoring with speeds of 5 FPS and 11 FPS respectively.
The paper compares four model compression techniques on a convolutional neural network for predictive maintenance across nine heterogeneous edge platforms. The evaluation focused on classification performance, inference speed, and compression ratio using real-world rotating machinery data. Knowledge distillation was identified as the most effective compression method, achieving over a 90% speedup while maintaining a 91.22% balanced F-score. This approach resulted in a 98.9% size reduction of the model. The study demonstrates a method to reduce computational costs in industrial settings for achieving Industry 5.0 goals.
Medical AI models risk privacy breaches through membership inference attacks, even when training data is not shared. A significant issue is that risks are often reported as cohort averages, masking individual patient exposure. The research measured patient vulnerability across numerous models and datasets to identify potentially exposed records. A lightweight screening tool using four small models was developed to identify high-risk patients before production model training. This screening process suggests privacy assessment should be integrated into the data preparation phase for clinical AI.
Visual mathematical reasoning can suffer from errors when a vision-language model misreads diagrams, creating false premises for reasoning. The proposed method uses minimal-edit on-policy distillation to correct these errors using reference descriptions and solutions as training information. A teacher scores incorrect student outputs based on a minimal-edit objective, focusing loss only on teacher-disfavored tokens. Correct rollouts are reinforced with task-reward reinforcement learning, preserving on-policy supervision without full imitation. This technique improves perception-oriented performance and transfers to external mathematics benchmarks.
Supervised fine-tuning can disrupt previously acquired behavior, and existing constraints do not clearly explain the relationship between protected objects, trajectories, and retained outputs. The research introduces a functional-path account to decompose old-loss changes during updates into controlled directions, gradient staleness, probe mismatch, and finite-step effects. This account identifies a conditional refresh budget and the restoring direction provided by informative replay. Different constraint placements and granularities impact retention differently across code-label audits and checkpointed generation. The findings suggest mitigating forgetting requires explicit functional anchors and joint auditing of old and new data.
The research introduces a method for generating differentially private synthetic data for databases in the $[-1,1]^d$ space under an integral probability metric. Generalized moment matching is utilized to privately estimate moments, fit a probability measure, and sample synthetic records. A general error theorem connects moment error to the target IPM, extending Chebyshev moment matching for 1-Wasserstein distance. The framework applies to objectives like maximum mean discrepancy and provides guarantees for sliced Wasserstein-1 distances. These methods yield minimax optimal rates for fixed dimensions when sufficient synthetic data is available.
Training state-of-the-art models from large, diverse, and noisy datasets requires effectively managing the influence of various training examples. A scalable method is developed to learn dataweights by using metagradients from a small, high-quality target dataset. This approach allows for the learning of parametric weighting rules that generalize to new, streaming data and scale across different model sizes. Reward modeling serves as a key application, demonstrating that learned dataweights improve performance and allow for effective mixing of human and synthetic supervision. Ultimately, these learned weights enhance performance and adapt reward models to specialized domains despite noisy supervision.
Flow-based spectral kernel attention (FSKA) is proposed as a framework for learning expressive, nonstationary attention kernels. It achieves this by formulating kernelized attention through spectral density learning and parameterizing the spectral density using normalizing flows. This design allows for flexible and regularizable kernel learning compatible with scalable linear-attention computation. FSKA simplifies attention design by removing query-key and fixed orthogonal value projections, leading to a more parameter-efficient architecture. Experimental results show FSKA achieves competitive or improved performance while reducing trainable parameters by $40\%$--$47\%$.
The xVGAE is a Hierarchical Variational Graph Autoencoder designed for exchangeable graphs. It is inspired by the Aldous-Hoover theorem, which characterizes graph distributions invariant to node permutations. The model learns an approximation of graphons and latent variables from a training set of graphs. xVGAE can generate new graphs quickly that closely match the statistics of the training data. This approach enables fast and high-quality graph generation for applications where standard models face computational limitations.
Deep Gaussian Processes are proposed with a continuous-depth formulation to address optimization difficulties in discrete-depth models. This formulation treats depth as a continuous evolution governed by a stochastic differential equation with a GP-based prior on the drift family. Two instantiations are studied: CDGP, which has a direct GP prior over the state-depth drift, and FlowDGP, which uses a flow-evolved prior for richer depth dependence. The research analyzes the sensitivity of the single-sample DGP objective and evaluates performance on various regression tasks. Experiments indicate that residual DGP improves upon plain DGP, while CDGP and FlowDGP offer superior results.
Random projections effectively reduce computational overhead in spectral algorithms while preserving data structure. The paper establishes optimal learning rates in Sobolev norms for these Random Projection (RP)-based methods under general conditions. Using RKHS embedding and operator similarity analysis, the authors find the required projection dimension is proportional to the effective dimension. This optimality holds across a wide range of source conditions without relying on uniform boundedness assumptions. Experimental results confirm the theoretical validity and practical effectiveness of these projection methods.
A two-layer framework is proposed to control prediction uncertainty in computer model calibration. The first layer uses a conditional flow-matching model to learn an amortized posterior for calibration parameters based on perturbed simulator outputs. The second layer employs conformal risk control to arrange posterior draws into nested sets, selecting the smallest set whose pointwise bands achieve a target miscoverage. This method controls expected miscoverage in finite samples, showing sharper bands than prior methods for a benchmark. Application to an age-structured SEIR model demonstrated that the resulting uncertainty bands exposed structural model inadequacies.
Deep learning methods are proposed for regression involving random objects in metric spaces, moving beyond traditional manifold or Euclidean embedding assumptions. The framework operates directly on the metric structure of the response space, while maintaining statistical guarantees. Statistical consistency and convergence rates are established for the proposed estimator, showing its potential to mitigate the curse of dimensionality. Simulation studies confirm competitive predictive performance across various configurations. Practical utility was demonstrated in digital health case studies using Wasserstein spaces with real-world data.
The method introduces the conditional expected local radius (CELR) to facilitate metric conformal prediction in general metric spaces. CELR quantifies local conditional probability concentration, enabling prediction sets that handle heterogeneity without requiring density or vector space structures. A novel conformal prediction method is developed using this measure, which yields a nonconformity score comparable across covariate values. The approach adapts a metric distributional random forest to manage multivariate covariates and nonlinear interactions efficiently. Numerical studies confirm the method's effectiveness in uncertainty quantification and computational savings compared to other scoring methods.