PROPEL is a framework training task generators via RL to create optimally difficult problems for agent learning. A lightweight probe predicts solver pass rate without repeated rollouts, reducing evaluation to a single forward pass. On code and SWE tasks, learnable-frontier generation increases from 10.1% to 20% (Qwen2.5-3B) and 9.8% to 19.6% (Qwen3.5-27B).
New K-Hop Gaussian (KHG) diffusion method to enhance GNNs. KHG preprocesses graph data with multi-hop diffusion weighted by Gaussian, balancing local and global propagation. Outperforms standard message-passing, PPR, and Heat Kernel on benchmarks, especially on noisy graphs.
Gaussian Mixture Attention (GMA) replaces standard attention with probabilistic routing through K learned Gaussian mixture components. Queries and keys map to responsibility vectors in a shared latent space. GMA avoids explicit N×N matrix materialization, reducing memory complexity to O(NK) instead of O(N²). Competitive on long-context classification, but behind SDPA and Mamba on WikiText-103.
ASTRA is an air traffic control training simulator automating pilot roles through speech recognition, instruction interpretation, and response generation. The system reduces Word Error Rate from 107.80% to 23.45% on Singaporean-accented aviation speech, and evaluates trainee radiotelephony communications achieving 91.7% accuracy, 88.2% brevity, and 86.9% completeness scores.
Artemis is a causal framework for graph neural networks addressing demographic confounders (age, sex) in multimodal brain imaging (fMRI + DTI). The method applies causal interventions at each brain region independently to learn invariant representations. Tested on ADNI, OASIS, and HCP benchmarks, it improves disease diagnosis and classification tasks.
Structural pruning framework for Mixture-of-Experts models operating at channel level rather than expert level. Attribution-based method reformulates pruning as channel-score coverage maximization. Experiments on DeepSeek and Qwen models achieve 50% structured pruning with 4-bit quantization, 5.27× memory reduction on Qwen3-30B-A3B.
New geometric complexity measure called Fisher width, a Fisher-geometric analogue of Gaussian width on statistical manifolds. Replaces Euclidean geometry with Fisher information metric to capture local statistical curvature. Develops foundational theory with generalization bounds and computable estimators, validated on MNIST.
SAGE is a post-hoc method to improve selective unlearning in LLMs. It corrects final update vectors by suppressing components damaging retention, without rerunning the original unlearning pipeline. Tested across multiple methods and scales, SAGE reduces the forget-retain trade-off.
Ghost Attractor Networks introduce an efficient dynamical decoder for sequential generation in robotics. With 2.3M parameters, it matches the offline accuracy of a 1.07B-parameter Diffusion Transformer (462× fewer parameters, 32× lower latency). On LIBERO-10, phase conditioning improves success rate by 13.5 percentage points over MLP baseline.
Survey of AI-based models for soil moisture estimation and classification. Five categories compared: statistical time-series, geostatistical methods, classical ML, deep learning, and Bayesian approaches. Data-driven methods provide flexible alternatives to computationally expensive physics-based models.
SWave is a complex-valued recurrent language model (169M parameters) trained on FineWeb-Edu. The paper documents its evolution across three phases, identifying structural failures (cos-domination collapse) and validating critical components (ComplexNorm, Wave Propagation Scan). Final PPL: 22.0 at step 89,861.
Self-CTRL optimizes consistency between language models' self-explanations and behavior via reinforcement learning. On probabilistic reasoning tasks, the method improves R² correlation from 0.24 to 0.64. In constitutional AI, it increases refusal prediction from 36% to 92% and reduces HarmBench failure rate from 15.0% to 0.5%.
SCOPE-FL introduces a hierarchical Federated Learning system using the Top Trading Cycle algorithm for client selection. The mechanism guarantees Pareto efficiency and strategy-proofness, with reward distribution via Shapley value approximation and blockchain execution. Evaluation on MNIST, Fashion-MNIST, CIFAR-10 shows improvement over DA, IAS.
P²CE generates plausible Pareto-optimal counterfactual explanations for ML models. The algorithm uses isolation forests and SHAP values to balance feasibility, plausibility, and computational efficiency. Evaluated on 3 datasets, it outperforms existing methods in solution quality and speed.
New LLM inference scheduler replacing explicit length prediction with lightweight statistical signals and dynamic priority boosting. Reduces P99 TTLT by 35-50% vs SRPT with perfect length knowledge, and TTFT by 34-47% across production and open-source traces.
TMR-GGNN, a time-aware multi-relational graph neural network, detects credit card fraud by modeling heterogeneous interactions between customers, merchants, devices, and IPs. The model combines temporal relational attention, contrastive learning, and a composite loss function (InfoNCE + Focal Loss) to handle imbalanced data and reduce false negatives.
Study on grokking (delayed transition from memorization to generalization). Authors show weight norm doesn't directly control grokking delay but acts through logit scale. Fixing norm and varying output temperature, they recover 85% of delay by matching logit scale. Effect is loss-dependent (cross-entropy vs MSE). Logit scale and softmax saturation are the proximal variables.
RL framework inspired by neuroscience that disentangles dynamics-specific and reward-specific features using locally linear embeddings (LLE) and adaptively fuses representations via attention mechanism. Improves learning efficiency on benchmark tasks compared to conventional RL approaches.
QAQL framework couples quantum annealing with Q-learning for remaining useful life (RUL) prediction in predictive maintenance. Each Q-value update encoded as QUBO solved on D-Wave Advantage system. Validated on NASA C-MAPSS and fleet maintenance datasets: statistically significant improvements over classical and quantum baselines.
PSyGenTAB is a privacy-preserving framework for synthetic clinical tabular data generation formulated as constrained optimization solved via Augmented Lagrangian Method. It embeds configurable privacy constraints into training to preserve inter-feature clinical relationships and minority-class patterns while maintaining data utility for medical AI applications.
CodeBlock is a structure-aware sparse supervision framework for code LLM fine-tuning. It selects syntactically coherent code blocks rather than isolated tokens, estimating utility via generalized cross-entropy and data-flow signals. On 6 code-generation benchmarks, CodeBlock outperforms full-token SFT while using only 1.9% of supervised response tokens.
Mathematical link established between shock-wave theory and symmetry-quotiented stochastic gradient descent dynamics for neural networks. After quotienting parameter symmetries and entropy coarse-graining, effective dynamics satisfy a viscous Hamilton-Jacobi equation. Applied to MLPs, CNNs, Transformers, and mean-field networks.
DRIFT refines SFT training data distribution using on-policy Influence Functions. The method uses model rollouts as validation targets to minimize proximity gap and debias gradient norm bias. Experiments on 7B instruction and reasoning models show consistent performance ceiling improvements over existing curation baselines.
Sparse Autoencoders (SAEs) decompose activations into interpretable features, but this study shows that clamping a 'harmful' feature does not eliminate the behavior—it can recover via other residual pathways. Even with active intervention, 95.8% behavior recovery is achievable in refusal-steering, exposing a gap between feature-level control and behavioral completeness.
RGNet, a neural network architecture based on the renormalization group, addresses class imbalance and multidimensional noise for fault diagnosis. The model hierarchically compresses feature space and captures both local details and global patterns. Tested on imbalanced AI4I dataset.
ThousandWorlds is an ML benchmark for climate emulation of potentially habitable exoplanets. The dataset contains ~1800 simulations from 5 global climate models mapping 8 planetary parameters to 3D atmospheric fields. Three nested subsets and two evaluation protocols test 7 baselines; GP-based methods outperform standard deep learning.
LLMZero uses LLM agents with tree search to discover adaptive RL training strategies. The system identifies that capacity parameters accumulate monotonically while regularization parameters oscillate. Across 4 GRPO tasks, discovered strategies outperform the base model by 9-140% and grid search by 6-15%.
arXiv paper demonstrates that on biomedical tabular data, measurement noise limits the advantage of nonlinear models (deep networks, gradient boosting) over linear regression. Degree-k interactions are attenuated by the k-th power of feature reliability, while linear components are attenuated only once. Analysis of 140 UK Biobank tasks confirms this noise signature.
Evaluation protocol for single-image-to-3D mesh quality using VLM judges (vision-language models). Authors demonstrate that cheap proxies (CLIP similarity, geometry validity stats) fail to correlate with perceived quality. Their VLM-judge protocol with position-bias correction achieves Cohen's kappa = 0.66 between two independent judge families.
Study of functional redundancy in single-layer tanh RNNs using ordered real Schur coordinates. Authors identify nonnormal couplings removable with minimal loss on specific tasks (copy, flip-flop, sine generation), revealing task-dependent approximate functional invariances rather than universal weight-space symmetries.
Reject inference methods used in credit scoring to correct survival bias mask a structural failure: accuracy can improve while the ability to correctly reject defaulters collapses. Authors propose a controlled exploration strategy (approving 2-5% of rejected applicants) to diagnose this deterioration without strong statistical assumptions.
Study shows SFT overtraining can invert model rankings during RLVR fine-tuning. On Qwen2.5-Coder-3B, increasing SFT depth raises pre-RL pass@1 but reduces GRPO pass@10 from 0.806 to 0.481. Pre-RL entropy positively correlates with RLVR outcomes (ρ=+0.69). Two-stage entropy-based diagnostic identifies high-risk checkpoints.
Causal framework for sleep recovery scoring from multimodal polysomnography. Uses DAG learning on two cohorts (MESA n=1540, MrOS n=825) to identify five physiological domains (respiratory burden, hypoxia, fragmentation, architecture, autonomic regulation). Sleep Recovery Score (SRS) achieves 2.5× stronger alignment with perceived recovery than standard AHI.
Theoretical paper on Sum-of-Squares degree barriers for robust halfspace learning under malicious noise. The Christoffel function exactly characterizes corruption hidden from bounded-degree certificates. Proves a margin-degree tradeoff and a degree-2t algorithm achieving the frontier η^(1-1/2t).
Researchers identify an internal signature of deception in language models: deceptive responses show 2.1-2.3x higher residual rank than naively false answers. This signature detects deception with 100% accuracy on GPT-2, Qwen2.5, and Phi-3, and transfers zero-shot across model families and languages (AUC 0.933-1.0).
Adaptive sequential sampling method for polynomial chaos expansion surrogate models, generalized for multiple quantities of interest. The approach balances input space exploration with exploitation of aggregated variance across outputs, improving surrogate accuracy and stability compared to Latin Hypercube Sampling.
arXiv paper on critic-free reinforcement learning for LLMs. Authors challenge the role of rollout groups in existing methods and propose negative token filtering to enable stable single-rollout training, improving performance on agentic tasks compared to group-based RL techniques.
ProCUA-SFT is a dataset of 3.1M step-level SFT samples generated automatically from 93K synthetic trajectories across 2,484 application combinations. Fine-tuning UI-TARS 7B on ProCUA-SFT achieves 45.0% on OSWorld, a +18.7 percentage-point improvement over the base model and +35% above AgentNet. The pipeline uses Kimi-K2.5 as task generator, precondition judge, and trajectory executor.
InferBERT combines transformers with Do-calculus to detect causal adverse drug events in pharmacovigilance. Comparative study on AILF and TRAM benchmarks: BioBERT outperforms XGBoost, ALBERT, and Med-LLaMA. Finding: domain-specific pre-training outweighs model size.
MODE is an expert-level mixed-precision quantization framework for MoE multimodal LLMs. It decomposes expert selection frequency by modality (vision/text) and filters redundant vision tokens to correct estimation biases. Results: <2.9% performance loss at W3A16.
GNN surrogate for CO₂ migration forecasting in complex geological formations. Model trained on SPE11A benchmark with anisotropic message-passing mechanism capturing directional transport. Produces competitive forecasts of gas saturation and liquid-phase density over extended forecasting horizons.
PowerOPD stabilizes on-policy distillation for LLMs by replacing unbounded log-ratio rewards with Box-Cox power transformation. On 6 mathematical reasoning benchmarks with Qwen3, achieves +6.37 Avg@8/+5.71 Pass@8 gains vs vanilla OPD, reduces wall-clock time by 59.2% and peak GPU memory by 23.1%.
arXiv study using Transformer model on MLB Statcast data to optimize baseball pitch sequences. Counterfactual analyses show optimization of both final and setup pitches can improve seasonal pitcher statistics by over 1.0 K/9. Practical insights on velocity-band-specific effective locations and pitch command importance.
MM++ is an unsupervised, post-hoc method for out-of-distribution detection. It fuses intermediate layers selected by entropy density with the final representation using Ledoit-Wolf regularized covariance, requiring no auxiliary OOD data, fine-tuning, or architectural changes.
Discrete autoregressive transformer for mechanism synthesis. Conditional sequence model with VAE latent and quantized joint coordinates. Trained on >1M mechanisms with Chamfer distance and DTW metrics. Mean Chamfer distance 0.0132, DTW 0.153 on held-out tests.
Deep learning framework for retrieving atmospheric CO2 from NASA's OCO-2 satellite spectra. Uses Laplace approximations and normalizing flows for uncertainty quantification. Inference orders of magnitude faster than operational algorithms, with better-calibrated non-Gaussian posterior estimates.
Comparative study of three recurrent architectures (LSTM, GRU, Mamba) and two algorithms (PPO, SAC) for meta-reinforcement learning applied to input-constrained control barrier functions (ICCBF) in spacecraft proximity operations. Mamba + PPO outperforms other setups in safety, task completion, and fuel savings across cooperative and adversarial scenarios.
MorphStrata enhances Moving Target Defense for time-series forecasting models via selective layer-specific stochastic noise injection. Tested on Transformer with FGSM, BIM and PGD attacks, the approach reduces adversarial RMSE by up to 97.97% on AEP data with training overhead <1%.
Hierarchical Bayesian credibility framework for pricing autonomous vehicle liability under operational design domain shifts. Tested on 648 verified Waymo crashes (4 US cities, 116M miles): credibility weights moderate (0.12-0.46), partial pooling decisively outperforms no pooling, learned kernel advantage detectable at ~12 deployed cities.
Operator Boosting constructs compact neural-operator surrogates for PDEs via stagewise residual learning. Tested on FNO, DeepONet, and CNO across 30 benchmarks (PDEBench, APEBench), the method reduces parameters by 72–95% while improving accuracy on 21 dataset-architecture pairs and achieves Pareto gains on 7/10 PDE benchmarks.