The paper proposes a C-test for supraintelligence based on Levin's Kt, which relates difficulty to the complexity of the underlying mechanism. It uses Large Language Models (LLMs) as representational machines to conduct this test. The research formalizes a measure and develops a pipeline to certify test items based on this framework. A battery of certified items was administered to twelve LLMs, showing agent characteristics decreasing with difficulty. This work explores the limits of certification for supra-intelligence beyond current frontier models.
Evaluations of agentic data science often focus only on downstream performance and code correctness, overlooking fundamental data science principles. The research argues that these assessments fail to capture methodological failures like data leakage and inappropriate validation. The authors conducted a multi-faceted evaluation across 220 tasks, a custom test suite, and qualitative analysis against expert solutions. Findings indicate that most agentic solutions contain failures, and data leakage correlates with poor predictive performance. Ultimately, principled failure mode analysis should be integrated into agentic data science evaluations for building trustworthy systems.
ClawLoop is a framework designed to enable self-improving agents within harness environments without relying on external teacher models. It operates through a self-sustaining cycle involving Task Synthesis, Training, and Diagnosis and Feedback phases. Task Synthesis generates training tasks, while Diagnosis analyzes agent behavior to produce feedback for subsequent task generation. The process involves updating the agent via Supervised Fine-Tuning or Reinforcement Learning. Using Qwen3.6-27B, ClawLoop achieved consistent improvements on several benchmarks, with gains also observed when transferring refined tasks to smaller models.
Iterative sampling in diffusion models causes error accumulation due to exposure bias between training and inference. A correction step is introduced to reduce this training-inference gap by training a corrector against the base model's predictions. During sampling, this corrector iteratively refines the denoiser's clean-sample predictions at intermediate denoising steps. This approach substantially improves sample quality without losing diversity and achieves performance competitive with classifier-free guidance. The method shows that the increased inference cost can be managed by distilling refined predictions back into the base model.
Long-Horizon-Terminal-Bench 2.0 extends previous benchmarks to measure an agent's ability to sustain progress over extended workflows using process rewards. It incorporates 77 tasks across 13 domains and eight task types, created through a systematic reconstruction pipeline. The evaluation horizon is increased to three hours, allowing for measurement of sustained progress via recorded checkpoints. Experiments showed GPT-6-astra led with a dense mean reward of 0.69 across 37 tasks at a threshold of 0.95. The work provides a more process-aware testbed for assessing reliable agent execution over long durations.
VETNET is an ecosystem of benchmark datasets and evaluation frameworks for advancing AI in animal science and veterinary medicine. It aims to standardize high-quality datasets through multi-institutional collaboration across academic and non-academic sectors. This framework seeks to bridge the gap between AI progress in human medicine and the fragmented development of veterinary AI. The project focuses on assembling standardized data to drive innovation in the veterinary field.
The study compares open-weight and closed frontier language models for security tasks using the CTI-REALM benchmark. The research finds that an open-weight model can achieve results comparable to the strongest closed models at less than one-tenth of the cost, given sufficient interaction turns. The benchmark's message limit acts as a confound, and results are more robust when the limit is relaxed. Stage-level metrics are proposed to better assess agent performance by evaluating submitted outputs against the ground truth. This suggests routing workflow stages to the best-performing model rather than relying on a single overall score.
Counterfactual interventions were applied to a Proximal Policy Optimization agent's value function to explore if it functions as a world model in the Crafter environment. Despite optimizing only for reward, the value function learned environmental concepts such as collectible rules, health impacts of resources, and inventory heuristics. This suggests that model-free reinforcement learning may implicitly learn environmental representations. The study also discusses the safety implications of these visualizations and current limitations in agent intelligence. Ultimately, it raises questions about what constitutes a truly learned world model in reinforcement learning.
The work introduces MOSAIC, a method to attribute training-example influence in vision-language models by learning a shared, sparse image-text latent space. MOSAIC encodes this space with properties ensuring sparsity, contiguity, alignment to descriptive phrases, and decorrelation. It then uses EK-FAC to decompose training-example influence, localizing contributions to specific text phrases and image regions. Evaluation on models like Qwen and Gemma shows that these learned features are more localized and coherent. The research demonstrates that visual evidence substantially reorganizes training-data attribution while preserving shared latent structure.
Cloud-hosted models provide the most accurate agent performance by accessing private user data. Local models running on edge devices keep data private but perform poorly without external context. An intermediate architecture proposes having cloud models delegate bounded data questions to small local models via function calls. This approach improves code-only agent success rates by up to 14.9 percentage points compared to cloud-hosted methods. This research explores using local models as interpreters in cloud-written agent programs.
The research introduces forecast grounding as a pretraining principle for time-series language models, ensuring learned representations retain predictive temporal information. The LOTUS method instantiates this by predicting the next patch from LLM hidden states to encourage temporal awareness. A temporal-spectral patch representation is also proposed to capture both temporal changes and periodic patterns. LOTUS achieved 59.91% accuracy across multiple evaluations, outperforming baselines significantly with minimal trainable parameters. These results confirm that forecast grounding effectively enables compact time-series language models to excel.
Inverse reasoning models use answer-only data to generate reasoning traces, treating the Chain-of-Thought as latent variables. A partner inverse reasoning LLM is trained to generate these traces conditioned on the answers, allowing the inverse model to be reused efficiently. This approach facilitates post-training and verification from answer-only data, addressing the inefficiency of requiring full reasoning traces. Distillation from this inverse model roughly doubles pass@1 performance on hard prompts compared to self-distillation methods. The inverse model also functions as an accurate verifier, improving selection accuracy in evaluation tasks.
The research benchmarks how class-incremental learning (CIL) methods, typically developed for vision, transfer to frozen protein encoders. The study tested nine CIL methods across five different protein encoders and two enzyme function datasets. The results showed that no method reliably transfers across these encoders, indicating instability in the transfer process. This instability stems from weak gradient signals and heterogeneity in embedding geometry, which depend on the dataset scale and encoder. Reliable protein CIL requires encoder-aware methods that adapt to the specific geometry of the representation.
Machine learning models are used in high-throughput perturbation screens to predict gene effects for untested combinations. Standard out-of-distribution methods often fail when predicting unseen gene interactions, even if individual genes were seen during training. The study introduces "additive disagreement" (AD) to measure the extra interaction effect introduced by the model beyond a simple sum of single-gene effects. Predictions with larger deviations from the additive baseline tend to have larger errors. Additive disagreement proves to be an informative signal across various compositional settings for agentic perturbation-screening workflows.
Token importance for on-policy distillation is derived from the distillation objective, specifically how much loss can decrease under a local KL budget. This utility is calculated as $s_t=g_t^\top F_t^\dagger g_t$, where $g_t$ is the logit gradient and $F_t$ is the categorical Fisher matrix. This method, termed Fisher Utility Weighting (FUW), concentrates importance on relevant positions across various distillation criteria. FUW improves performance over uniform distillation for multiple benchmarks and Qwen3 models under both reverse-KL and forward-KL settings. The derived utility is competitive with existing selective methods based on a single objective score.
PMO-Dock introduces a benchmark and protocol for docking-based molecular optimization, addressing the lack of standardization in the field. It includes 25 tasks covering hit generation, lead optimization, and a new specificity task that penalizes off-target interactions. The protocol separates model development and final evaluation, using held-out test tasks for final assessment under a strict oracle budget. Four diverse methods, including Saturn, GenMol, Genetic-guided GFlowNet, and ChemLlama, are benchmarked against this protocol. Results indicate that no single method is universally superior, suggesting method-dependent selection strategies based on generalization and sample efficiency.
The study investigates how the repetition of samples affects performance in image self-distillation models, challenging the assumption that more unique samples always yield better results. Experiments across various datasets revealed that the effect of sample repetition is dataset-dependent and non-trivial. The authors propose a modified Chinchilla-style scaling law that explicitly incorporates both the number of unique samples and repetition per sample. This new formulation aims to better capture the data-compute trade-offs in these distillation tasks. The research also explores the impact of subset selection and augmentation on downstream performance.
Limited real measurements necessitate using large-scale synthetic pretraining to improve indoor radio-map prediction. The research introduces a pipeline featuring a layout generator and an approximate ray tracer to create diverse synthetic data. Pretraining ResNet-101 and Swin-T encoder-decoders on 128M synthetic samples, followed by fine-tuning, demonstrated error reduction. The models showed predictable power-law scaling in performance based on data size, even with simulator mismatches. Ultimately, pretraining on cheap simulation data enhances indoor radio-map prediction accuracy when real data is scarce.
The work introduces an approach for regret matching in game solving that incorporates previous-utility prediction to achieve stronger theoretical guarantees. This method, IR-PRM+, recovers the constant-regret and convergence-rate guarantees previously seen with extra-gradient prediction for simultaneous and alternating updates. The technique bounds strategy movements and uses the relative growth of the regret-vector norm to prevent unbounded norm growth. These guarantees extend to finite two-player zero-sum matrix games and corresponding extensive-form games under specific conditions. Consequently, the current duality gap vanishes, and the current profile converges to the Nash equilibrium set.
Low-rank adaptation (LoRA) applies rank-one updates to every token, prompting a question about the necessary number of components per token. The paper introduces SQUELCH, a router-free method that uses soft-thresholding and Stein's unbiased risk estimate to dynamically select active ranks per token. This technique suppresses redundant components by performing an explaining-away step on the value Gram matrix. SQUELCH improves performance over standard LoRA and mixture-of-experts baselines on various reasoning tasks. On commonsense reasoning tasks, SQUELCH activates an average of 8.6 out of 32 components per token.
Estimating treatment effects is difficult when outcomes exhibit a discrete zero mass and a skewed continuous distribution. The proposed method, ZIC-BCF-Smear, is a decoupled, two-part nonparametric estimator for zero-inflated continuous outcomes. It models participation with a Probit BCF and positive intensity with a Gaussian BCF, using Propensity Adjustment and Duan's smearing for re-transformation. Simulations confirm this method matches well-specified models without their assumptions. The technique was applied to estimate effects in a cancer trial concerning severe toxicity.
The study investigates the PAC learning of binary classifiers when labels are delivered with random delays and are censored beyond an observation window $W$. It introduces the Delay Complexity Dimension (DCD), which combines the VC dimension with the exact label-information rate of the observation channel. The research shows that learners require $\Omega(\mathrm{DCD}/\varepsilon^2)$ samples, and an oracle pseudo-label ERM achieves this rate. The theory also suggests a quantization scheme where delay information reduces residual label variance, offering an expedition policy guarantee. Experiments indicate that a moderate window can outperform immediate labels, supported by synthetic data and log analysis.
The paper investigates Wong--Zakai convergence and Fourier Neural Operator (FNO) learning for stochastic Volterra integral equations (SVIEs) that model systems with memory. A Wong--Zakai theorem supports the convergence of pathwise solutions when using piecewise linear approximations of Brownian motion. This convergence facilitates the use of mirror-padded Fourier neural operators to learn the solution operator for SVIEs. Numerical experiments show that this method achieves competitive prediction accuracy and generalizes well to finer temporal grids. The study extends prior operator-learning frameworks to SVIEs by analyzing convergence and approximation for convolutional memory.
Amortized Bayesian inference is introduced to estimate parameters when likelihoods are computationally expensive or unavailable. The method divides parameters into blocks, leveraging conditional independence to identify necessary information for each block. Neural estimators are trained via simulations to approximate the full conditional posterior for expensive blocks. A Markov chain alternates between draws from these estimators and conventional updates for the remaining variables. This approach allows for the reuse of estimators across different blocks and datasets in complex models.
Discrete diffusion models are investigated for their performance under sparsity assumptions, contrasting with continuous diffusion models. The research analyzes the statistical rates of discrete diffusion models by studying a nonparametric estimator from a regularized score-entropy objective. This analysis establishes minimax-optimal rates for discrete diffusion models in the sparse regime. The proposed regularization method adapts intrinsically to sparsity without needing explicit knowledge of the sparsity parameter, similar to label smoothing. Consequently, discrete diffusion models demonstrate an inherent ability to adapt to sparsity.
The research introduces physics-informed weak regression (PIWR) for parameter estimation in epidemic models, aiming to provide distribution-free uncertainty. PIWR integrates the model over time, estimates the initial state jointly, and requires no data derivatives. This method, combined with split conformal prediction, yields simultaneous parameter intervals with finite-sample coverage guarantees. Experiments show PIWR and related methods are significantly more accurate than standard physics-informed methods and DINNs. The approach facilitates fast estimation while ensuring reliable uncertainty quantification for disease models.
The research proposes nonlinear supervised clustering fusion (NSCF) to learn individualized treatment rules from many treatments. This framework uses a neural Q-learning approach with shared nonlinear covariate representation and low-dimensional treatment embeddings. It groups treatments with similar learned contrasts to yield an individualized treatment rule and an interpretable partition. An alternating algorithm combines neural training, embedding fusion, partition selection, and grouped refitting. Results show that NSCF improves policy value in a melanoma xenograft application, supporting nonlinear representation learning with supervised treatment fusion.
The paper extends probabilistic graphical models to non-Gaussian data by characterizing the pairwise Markov property using copula densities. It derives a necessary condition for this property specifically within Student copula Markov networks. An Expectation-Conditional Maximization algorithm (ECM) is introduced for penalized maximum likelihood estimation and graph selection. Simulation studies evaluate the ECM's effectiveness in estimating parameters and recovering graph structure. The work contributes a new framework for modeling dependencies in these specific network structures.
Full waveform inversion (FWI) is improved using Structure-aware Bayesian methods, specifically Annealed Stein variational gradient descent (ASVGD), to quantify uncertainty. The method incorporates a learned diffusion model as an implicit regularization to capture geological structure, overcoming limitations of simple priors. This combination of FWI data consistency, posterior diversity, and structural realism is achieved. Experiments show that this approach yields more geologically consistent mean models and better uncertainty estimates. Consequently, learned diffusion priors enhance Bayesian FWI stability and interpretability.
Causal discovery algorithms have been developed for complete data, but methods for handling missing values remain limited. The research establishes general conditions under which causal graphs remain identifiable even with various missingness categories, including Missing Not At Random. Identifiability is preserved if missingness mechanisms do not causally interact, no variable causes its own missingness, and faithfulness or generalized linear models hold for the missingness. Based on these findings, two asymptotically correct algorithms are introduced for Additive Noise Models to recover causal graphs. These results allow for the extension of correctness guarantees from complete data to settings with missing data.