__ishaan

EvalCards: Standardizing AI Evaluation Reporting with an Integrated and Interpretable Layer

EvalCards is proposed as an operational layer to standardize AI evaluation reporting, addressing inconsistencies across various platforms by integrating benchmark metadata, evaluation data, and model information into a unified record. It establishes a schema derived from literature and interviews, incorporating signals for reproducibility, documentation, provenance, and comparability, and presents findings through reader modes tailored for different audiences. A deployed monitoring tool has analyzed a large corpus of AI models and results, identifying systematic reporting deficiencies.

Generative Flow Matching for Non-Markovian Stochastic Dynamics

This paper introduces a generative flow matching technique to address the limitations of hydrodynamic models, such as the regularized Dean-Kawasaki equation, in capturing non-Markovian and non-Gaussian dynamics in stochastic particle systems. The proposed method directly models flux distributions from simulations, improving accuracy in short-time dynamics and statistical moment predictions, as demonstrated in the Kramers first passage time problem for Brownian particles.

SWE-Marathon Benchmark Reveals AI Agents' Deficiencies in Long-Horizon Software Engineering Tasks

The SWE-Marathon benchmark introduces 20 long-horizon software engineering tasks designed to evaluate AI agents' capabilities in planning, long-context understanding, and memory over extended workflows (averaging 27.2M tokens). Current frontier agents fail over 70% of these tasks, demonstrating limitations in areas such as self-verification and task completion. The benchmark also highlights issues like reward hacking and premature termination, necessitating more robust agent designs for complex, real-world software development scenarios.

MilliVid: Hierarchical Latents Enhance Long-Range Video Consistency via Coarse-to-Fine Generation

MilliVid introduces a hierarchical token-based approach to address the long-range consistency challenge in video generation. By leveraging a multi-scale token space and a coarse-to-fine rollout strategy within a diffusion model, it effectively maintains geometric and object permanence over extended sequences. This method compresses frames into hierarchical latents, allowing the model to focus computational resources on generating semantically important details while preserving overall consistency, outperforming prior baselines on long-form video generation tasks.

GenEyePose: Synthetic Saccade Modeling for Patient-Free Neurophysiologic Biomarkers

This work introduces GenEyePose, a novel patient-free, synthetic data generation pipeline for saccadic eye movement modeling. A deep learning classifier trained on this synthetic data demonstrated generalizability to real-world clinical data, achieving an AUROC of 0.76 and sensitivity of 0.71 for distinguishing normal from abnormal saccades. This approach holds promise for developing AI-enabled digital biomarkers for neurophysiologic screening and localization, overcoming limitations of current methods due to privacy concerns and scarce datasets.

Hybrid Particle Gaussian Mixture (H-PGM) Solution for Cislunar Target Tracking

Multi-Class Neurological Disorder Prediction with Tensor Network Feature Engineering

TinySDP: Bringing Real-Time Semidefinite Programming to Microcontrollers for Certifiable Robot Motion Planning

TinySDP is the first SDP solver architected for embedded systems, enabling real-time MPC with nonconvex obstacle constraints on microcontrollers — a class of hardware previously considered too resource-constrained for SDP. The solver combines PSD cone projections with a cached-Riccati ADMM formulation to exploit computational structure, and pairs this with a rank-1 a posteriori certificate that converts relaxed solutions into hard geometric guarantees per timestep. Validated on a Crazyflie quadrotor, TinySDP demonstrates that convex relaxations of nonconvex geometry can be enforced at control-loop rates, closing a longstanding gap between SDP-based planning theory and embedded deployment.

Why DDIM Hallucinates More than DDPM: A Theoretical Analysis of Reverse Dynamics

Intrinsic Linking of 2-complexes in $\mathbb{R}^4$

PoDAR: Power-Disentangled Audio Representation for Generative Modeling

Efficient Hybrid CNN-GNN Architecture for Monocular Depth Estimation

An Empirical Study of Automating Agent Evaluation