Explainable AI tools often fail to increase trust and instead promote overreliance on incorrect predictions. This limitation stems from XAI methods oversimplifying model internals, causing a misalignment known as unfaithfulness. Faithfulness is the approximation between an explanation and the model, and rigorously computing it is crucial for quantifying explanation trustworthiness. Current faithfulness metrics suffer from significant limitations, making quantification of trustworthiness infeasible or computationally expensive. The paper outlines these shortcomings and proposes research avenues to improve the reliability of faithfulness metrics.
EliteTable is a self-improving program-search algorithm that utilizes large language models to guide the search process. It employs a fixed-capacity table of high-scoring solutions to inform new proposals, edits, and remixes of existing solutions. Performance was tested across six Gymnasium control tasks, evaluating performance over ten generations. Results showed performance improvements for selected-winner searches and the transfer of gains to new episodes in some tasks. The work documents improvement during program search and the transfer of learned advantages to new environments.
A controlled computational study compared Monte Carlo Tree Search (MCTS) and randomized-restart list scheduling for Directed Acyclic Graph (DAG) workflows with equal budgets. MCTS performed 2.88% worse than the restarts across 2048 evaluated schedules, though restarts showed self-improvement on a separate pool. The performance difference depended on the workflow structure, winning for restarts on sparse irregular workflows but showing no difference on regular layered ones. Swapping to textbook UCT scheduling reduced the deficit, confirming the context-dependent nature of the performance. The results were based on a controlled synthetic domain with limited task counts and processor numbers.
The paper proposes EvoSyn, a learning-feedback-driven multi-agent framework for agentic data synthesis in closed-loop learning. It addresses the gap where observed weaknesses do not directly guide future data synthesis, by organizing synthesis around a revisable state specifying interaction factors. A role-specialized multi-agent system translates learner feedback into synthesis revisions, while a Designer sets directions and a Planner grounds them into executable experience. This process allows for adaptation to weak areas and expansion of the synthesis space by incorporating newly discovered factors. Experiments demonstrate that EvoSyn improves task success on benchmarks and yields higher-quality, broader synthetic experience.
SpireBench evaluates the long-horizon strategic play of agentic models using the commercial deck-builder game Slay the Spire. The benchmark assesses how models make decisions over extended periods to determine long-term success. Frontier LLMs performed better than heuristic baselines and matched average human players initially, but their performance declined as the game extended. The study suggests that performance differences stem from how agents manage risk and utilize prior game knowledge effectively. SpireBench provides a comprehensive, long-horizon benchmark with tools to diagnose agent failures.
Binary reward systems in reinforcement learning offer limited supervision for software engineering agents. Behavioral Self-Rewarding (BSR) is introduced as a method to refine rewards by allowing the policy to judge its own trajectories. BSR reuses the policy to compare debugging behaviors between different trajectories from the same task and verifier outcome. This process generates task-local scores that are converted into bounded auxiliary rewards, providing contrast without altering the pass/fail order. Experiments demonstrate that using these comparative judgments improves performance for SWE agents across evaluated models.
The CLEAT framework addresses interactive intent alignment by learning value-guided clarification for agents dealing with incomplete information. It co-trains a clarification-value controller and a task policy using shared interaction outcomes. The method introduces a residual query value to quantify the task utility gained from asking a question instead of proceeding. This value is estimated through paired rollouts, training the controller to predict clarification value and the policy to select useful questions. CLEAT demonstrates significant performance gains across benchmarks by learning how to effectively use clarification in agent interactions.
Data-centric scaling of RLVR often results in a "zero advantage trap" where large datasets do not provide useful learning signals. To generate effective data, synthetic problems must require the reasoning and composition skills needed to solve hard problems while being calibrated for success and failure within a training budget. MagicBeanStalk introduces a recursive data-synthesis algorithm to grow seed problems into new, policy- and budget-calibrated training problems. This method improves performance on benchmarks like gpt-oss-120b and DeepSeek v4Flash in math, and yields improvements in finance as well. The approach effectively synthesizes verifiable problems, leading to significant performance gains across different domains.
The research presents an automated pipeline using Large Language Models (LLMs) to construct hierarchical hardware knowledge graphs from hardware datasheets. This process involves chunk-based property extraction, hierarchical routing, and iterative refinement to manage the fragmented and heterogeneous information in datasheets. The framework leverages LLMs for extraction and reasoning while preserving source provenance and using validated instructions for graph updates. Evaluation shows the method achieves 98.8% property faithfulness compared to a one-shot baseline. These improvements yield more reliable and comprehensive engineering information for hardware design and verification.
The method aligns representations from few paired data points using distilled optimal transport. It achieves this by expressing representations based on similarities to the paired anchors, placing both modalities in a common coordinate system. A single entropic optimal transport problem is solved over the entire data pool to ensure the matching counteracts hubness. The resulting plan is then distilled into lightweight encoders, making the correspondence inductive. Evaluation on image-text and molecule-text alignment shows superior performance over existing methods.
Algebraic integer number formats (AINF) provide a framework for highly configurable and accurate compressed computation by reducing calculations to purely integer arithmetic. This property allows for error-free multiply-accumulate operations, enabling a matrix multiplication engine with competitive performance on hardware accelerators. Algebraic integer quantization (AIQ) is a lossy compression format for large language model inference that optimizes based on specific numerical characteristics. Experiments show AIQ achieves accuracy comparable to the industry-standard MX format for LLM inference at equivalent bit-per-parameter budgets. Future work involves generalizing these concepts to higher-order and multidimensional algebraic integers.
Kairos proposes a byte-level foundation model capable of handling every modality by training on a shared one-byte space for text, image, audio, video, lidar, and control streams. It integrates bidirectional gated-DeltaNet linear attention, sliding-window attention, and a sparse Mixture-of-Experts feed-forward layer. The model utilizes a multi-scale byte-level codec and is trained with a uniform-noise denoising objective across these modalities. Sanity checks indicate that the proposed combination optimizes the implementation, with attention variants reaching a minimum loss near $10^{-3}$. The work suggests optimization but points to further multimodal training for generalization.
Resonant Sequence Models use a fixed-size state to avoid attention costs but require a design choice for state transitions. Modifying the readout phase to be input-dependent introduces content-dependent memory while retaining an exact convolutional path. This modification for ResonatorLM adjusts the readout phase based on the current token, adding input dependence. For character-level models on WikiText-2, this input-dependent phase matched the original at training length and improved performance upon extrapolation. The learned phase tracks the current character rather than distance, and the method removes failure beyond the training length.
The work proposes CircuitTree, a method for composing multiple neural network circuits into a single model. It utilizes a greedy, density-scaled composition algorithm, generalized via CircuitTree, built on a $\Theta(N\log N)$ tree-compose schedule. The approach composes circuits pruned from HuBERT for phoneme classification into a syllable classifier. Chaining these CircuitTrees results in syllable transcription with 64.8\% accuracy. The method achieves this by sharing 99.5\% of circuit components across the components.
Distribution operators need behind-the-meter photovoltaic (PV) generation visibility, which can be achieved using deep-learning-based PV disaggregation from net smart meter data. The operational value of this disaggregation depends on deployment regimes that separate household transfer from temporal extrapolation. Benchmarking showed that the deployment regime explains more error variation than the model choice under fixed training conditions. Furthermore, fine-tuning on target-region households significantly reduces source-model transfer error. These results suggest using regime-matched error budgets and context-dependent selection for energy data analytics.
GlucoPRISM uses protocol-supervised pretraining to facilitate cross-cohort glucose transfer in foundation models. It supervises a GlucoFM backbone with response contracts dictating how representation components should change when only day, sensor, or both are altered. Supervision is derived from synthetic sensor views calibrated on paired recordings and glucose summaries, without clinical labels. Removing the sensor band improved transfer AUROC, demonstrating that a device-related component hinders transfer. This method improved performance across multiple public tasks and clinical endpoints by effectively filtering out device-specific noise.
CacheScope accelerates training-free diffusion inference by caching approximations instead of full denoising at specific steps. It addresses the issue where feature reuse leads to limited gains when reconstruction complexity increases. The method normalizes probe-output variation with input displacement to estimate a directional gain per unit trajectory movement. CacheScope selectively falls back to full denoising only when high-risk reuse steps are identified, without requiring additional training. This approach achieves up to 3.57x acceleration and a 70.1% reduction in LPIPS across several diffusion models.
Market impact is a causal quantity, but classical models are biased because they regress price on flow, ignoring informational flow. A Causal Market Simulator, a causal prior-data fitted network, is proposed to address this by explicitly modeling the confounding informational flow. The model demonstrates that an intervention-ablated twin collapses to the pre-window mean, and the model successfully tracks true-parameter oracles in complex markets. Applying the model to actual flow bursts in various markets suggests causal transfer, with further refinement showing that impact accrues over execution duration.
The research addresses the efficient learning of survival analysis models when data involves outcome-dependent truncation and censoring. The work connects the fields of learning under truncation and survival analysis, addressing limitations in existing methods that rely on iterative procedures. It demonstrates that Cox regression can be efficiently learned using first-order optimization when the baseline hazard is known, even with nonconvexity. When the baseline hazard is learned jointly, the method constructs identifiable models with correctly specified negative log-likelihoods. Experiments show that accounting for truncation reduces prediction bias in both Cox regression and neural networks, with significant speedups over expectation maximization.
Enriched holomorphic neural networks are combined with singular functions derived from local geometry to address singular behavior in 2D and 3D Laplace problems. This combination results in a harmonic representation that fits the smooth remainder from boundary data while capturing the singular behavior through enrichment terms. The singular exponents used in the construction are trainable, allowing the method to apply even when the local expansion is unknown. Extending the approach to three dimensions involves evaluating independent networks along complex null directions. The enriched model demonstrates superior accuracy, achieving margins several orders of magnitude better than other methods on radiation detector data.
Determinantal Point Processes (DPPs) use a kernel matrix to encode item similarities, but scaling their learning to large datasets requires structural constraints that limit expressive power. The study addresses this trade-off by modeling full-rank interactions using nonlinear kernels and latent features. A surrogate function for the maximum likelihood estimator and stochastic optimization with random features are employed for scalable learning. Numerical experiments show that this method achieves better generalization and parameter efficiency on real-world data.
Bayesian simulation-based inference methods often struggle with recovering concentrated posteriors as observation size increases. The research proposes jointly learning sufficient summary statistics along with the posterior for scalable inference in various SBI methods. A key concept introduced is "two-way guidance," where structural restrictions on the summary simplify posterior learning, and a posterior objective guides the summary toward sufficiency. This mechanism explains why structural restrictions on summary statistics improve posterior learning performance. The work establishes theoretical guarantees on estimation error and summary sufficiency, defining conditions for recovering concentrated posteriors.
The work studies an entropy-based estimator for Kullback-Leibler (KL) divergence derived from normalized importance weights. It establishes a universal finite-sample bound showing that the expectation of this estimator can exceed the target-proposal KL divergence by at most $\log 2$. The KL divergence between the importance-weighted resampling distribution and the proposal is bounded by both the estimator's expectation and the target KL divergence. Crucially, these bounds hold without requiring bounded importance weights or finite higher moments. The results are applicable to test-time alignment in generative models and relative entropy coding.
Determinantal point processes (DPPs) model repulsion, but Nonsymmetric DPPs (NDPPs) capture both attraction and repulsion, although their kernels are hard to interpret. The determinant-ratio point process (DRPP) is proposed to model both attraction and repulsion through the ratio of two symmetric positive-definite kernels. This formulation preserves the feature-based interpretation of standard DPPs while allowing separate specification of the two effects. The DRPP framework naturally characterizes attraction structures, such as clustered attraction, with a rank directly related to the number of latent groups. Experimental results indicate that DRPPs are more parameter-efficient than NDPPs for fitting transitive attraction.
The research focuses on online binary prediction using paid stochastic features where learners can spend resources for better feature observations. This is motivated by applications like medical testing, where cost yields more precise measurements. The study contrasts with previous results by showing that covariance information alone is insufficient for optimal logistic losses. Under a structural ordering of payments, the regret achieved is $\widetilde O(\sqrt{T})$ if the noise structure is known, and $\widetilde O(T^{2/3})$ if it is unknown. The work establishes lower bounds matching these regret rates.
The research studies dictionary learning for measure-valued data by embedding measures into Hilbert spaces using optimal transport geometry. It learns data-adapted dictionaries through sparse autoencoder parameterizations to create fast and scalable representations of probability measures as sparse codes. The framework merges the geometric structure of optimal transport with the efficiency of sparse representation learning across modalities like images and text embeddings. The authors prove generalization bounds for sparse dictionary learning objectives based on the metric complexity of the atom class. Finally, they derive explicit covering number bounds for optimal transport dictionaries, supporting a scalable representation learning model.
Severe sample imbalance challenges standard two-sample tests, especially for complex, non-Euclidean data like graphs. The research investigates how to utilize abundant majority observations to maintain sensitivity to smaller samples during distributional comparisons. A novel framework, the Enhanced Imbalance-Aware Deep Classifier Two-Sample Test (EIAD-C2ST), is proposed to address this issue. This method incorporates imbalance into the testing evidence rather than treating it as a nuisance for training. The study establishes theoretical guarantees and demonstrates empirical performance gains by leveraging majority-sample surplus as statistical evidence.
Latent jump models (LJMs) are introduced to model time series where the state undergoes deterministic evolution punctuated by Poisson-driven, state-dependent jumps. To train these models efficiently, Scan-Parallel Variational Inference (SPVI) is proposed, which samples the latent event configuration before evaluating the state recurrence. SPVI uses triangular affine scans to enable fast and scalable training for these jump models in logarithmic temporal depth. LJMs prove competitive with continuous-time models in forecast accuracy and are significantly faster to train than sequential baselines. This method allows modeling uncertain, unobserved shocks in systems that are not necessarily smooth or diffusive.
Heavy-tailed spectral distributions in trained weight matrices are strongly predictive of deep neural network quality. This work investigates the statistical reasons behind this predictive success by using a probabilistic framework on final layer weights. The authors find that the limiting log-marginal likelihood for spectra following the High-Temperature Marchenko-Pastur distribution predicts model quality. They derive an optimal heaviness parameter that empirically predicts overfitting in large-scale DNNs. Ultimately, analyzing these heavy-tailed spectra offers a practical predictor of complex model performance.
Dynamic Time Warping (DTW) is slow and non-differentiable, leading to the proposal of Continuous-DTW (CDTW) with continuous warping functions. This paper introduces FastCDTW, an efficient approximate solver for CDTW with two strategies: an exact scheme for piecewise-linear functions and a Monte Carlo estimator for smooth functions like those from implicit neural networks. FastCDTW provides gradients necessary for optimization with respect to the warping function and input time series. The resulting algorithm is GPU-friendly and beneficial for complex machine learning tasks. This method is demonstrated in applications such as time series averaging and domain adaptation.