Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
arXiv:2609.31868v2 Announce Type: replace Abstract: Model-predictive control with Joint-Embedding Predictive Architectures (JEPAs) provides a strong zero-shot goal-reaching planner, but it is only ef...
PCDFusion is a new infrared-visible image fusion method that treats fusion as a local Bayesian posterior rendering problem. It uses a visible-anchored framework with three rendering stages to preserve useful thermal information, correct exposure-degraded visible areas, and incorporate reliable infrared structures into the final image. Experiments on public benchmarks demonstrate improved fusion quality and better performance on downstream detection and segmentation tasks under challenging lighting conditions.
The paper introduces a semi-automated method for structuring neural architecture search (NAS) using large language models (LLMs). It defines a high-level template of architectural attributes, which an LLM populates by analyzing research papers to create a diverse search space. The authors then propose FairNAD, a multi-type mutation strategy that combines fair idea sampling, Pareto-aware mutation, LLM-driven iterative mutation, and a fine-grained feedback loop to efficiently explore this space, achieving significant performance gains on CIFAR-10, CIFAR-100, and ImageNet16-120.
MeshSuite is the first benchmark designed to systematically evaluate triangle- and mesh-based neural rendering methods in both standard and dedicated engine deployment settings. It compares how different methods perform when using a conventional opaque mesh pipeline versus method-specific engine implementations, and also audits exported surfaces for topological and geometric integrity. The benchmark reveals that engine deployment generally degrades image quality, with MeshSplatting performing best under standard deployment, while dedicated deployment preserves fidelity at a higher computational cost.
FaithfulFaces introduces a pose‑faithful facial identity preservation framework for text‑to‑video generation, addressing identity distortion caused by large pose variations or occlusions. The method employs a pose‑shared identity aligner that refines and aligns facial poses across views using a pose‑shared dictionary and a pose‑variation‑identity invariance constraint. By mapping single‑view inputs into a global pose representation with explicit Euler angle embeddings, the framework provides a robust facial prior that guides generation toward consistent identity and structural clarity, achieving state‑of‑the‑art performance on a curated high‑quality video dataset.
MANGO‑Grasp introduces an anisotropic interaction framework for cross‑embodiment dexterous grasping. It represents objects as geometry‑oriented 3D Gaussian primitives to preserve local surface geometry, encodes hands with pretrained descriptors that capture morphology and kinematics, and models hand‑object interaction using anisotropic Mahalanobis fields. The method achieves state‑of‑the‑art performance on CMAP and MultiGripperGrasp benchmarks, outperforms baselines by up to 8.24 percentage points in simulation, and demonstrates strong zero‑shot transfer to the unseen SharpaWave hand with up to 16.57 percentage point improvement, achieving 86% success in real‑world experiments.
The paper introduces Heuresis, a benchmark of anonymous simulated datasets spanning linear Gaussian to heavy-tailed mechanisms, designed to evaluate how well LLM agents can perform causal discovery and improve upon existing algorithms. The benchmark tracks algorithmic gains versus data‑driven insights and records every edge an agent modifies. Using this benchmark, GPT‑6.1 Sol and Claude Opus 5.5 outperform the best existing algorithms on all six dataset families, with GPT‑6.1 Sol achieving a 14% adjacency F1 improvement on the hardest family.
The paper introduces a framework for reinforcement learning in Markov decision processes where policies are ordered by the information they provide, allowing data from one policy to estimate the performance of others. An epoch‑based algorithm is developed for infinite‑horizon average‑cost MDPs with continuous state and action spaces, achieving regret that scales with the width of the policy class’s information order. The authors apply the framework to lost‑sales inventory, dual‑sourcing, and state‑dependent queueing systems, demonstrating improved regret bounds and empirical performance without requiring global convexity or specific arrival‑rate structures.
The paper presents a new framework called the Stochastic Operator Network (SON) for uncertainty quantification of solution operators in stochastic partial differential equations (SPDEs). SON combines the Deep Operator Network architecture with Stochastic Neural Networks to learn directly from noisy data, producing both a mean solution field and a measure of predictive uncertainty. The training uses a Hamiltonian-type loss optimized via the Stochastic Maximum Principle, and numerical experiments on benchmark SPDEs show the method’s accuracy and robustness.
The paper investigates looped language models that reuse Transformer blocks across logical depth, separating recurrence placement, input injection, and recurrent state design. Experiments at matched parameter scale, logical depth, and token budgets show that a prelude‑recurrent‑block‑coda architecture improves context extraction and reasoning, while input injection boosts knowledge retrieval. The Huginn‑style design, especially at 730 M parameters and 336 B tokens, outperforms the Ouro architecture on all ten benchmarks, and its advantages transfer to Mixture‑of‑Experts models, achieving superior performance with fewer resident parameters.
The paper introduces Grounded Credit Assignment (GCA) for reinforcement learning in multi‑turn search agents. GCA separates credit for evidence retrieval and answer correctness, correcting the bias in GRPO that favors evidence‑retrieving rollouts only proportionally to accuracy gains. Applying GCA to Search‑R1 boosts average EM by 5.7–6.3 points and increases the proportion of evidence‑supported correct answers from 76.9% to 93.4%.
Process Reward Models (PRMs) are crucial for scaling large language model reasoning at test time, but they suffer from two types of uncertainty: PRM uncertainty from limited supervision and imperfect training, and policy uncertainty from mismatched data collection policies. The paper introduces BayesianPRM, which treats process rewards as random variables, discretizes the reward space with an ensemble, and applies two belief updates—reliability‑aware updating for PRM uncertainty and conservatism‑aware calibration for policy uncertainty—to form a Bayesian belief whose expectation yields a more accurate reward estimate. Experiments on multimodal reasoning benchmarks show that BayesianPRM improves reward accuracy and test‑time scaling performance.
The paper introduces DiffPT, a diffusion‑based pretraining method for classifier backbones aimed at improving performance when training data is scarce. Experiments on time‑series, natural image, and medical imaging datasets show that DiffPT outperforms many self‑supervised baselines, especially in low‑data regimes for medical and time‑series tasks. While contrastive pretraining remains superior on CIFAR‑10 and Tiny ImageNet, DiffPT offers a simple and effective alternative across diverse modalities.
The paper critiques current multi‑task active learning (MTL‑AL) methods that use linear scalarization of objectives like uncertainty, diversity, and gradient conflict, arguing that such combinations miss Pareto‑optimal trade‑offs. It introduces a new framework that separates acquisition into informativeness and exploration, decouples conflict scoring from candidate ranking, and uses gradient surgery during training. The authors also propose a cross‑task mutual information metric and demonstrate consistent improvements on COCO, Cityscapes, and NYUv2 datasets compared to recent baselines.
The paper introduces Transductive Sharpening (TS), a training objective that leverages predictions on unlabeled nodes in transductive graph settings. TS minimizes prediction entropy for unlabeled nodes while balancing this effect on labeled nodes, using prediction confidence derived from cross‑entropy decomposition. Experiments across many node‑classification benchmarks show consistent performance gains without altering the underlying GNN architecture.
STRIDE (SpatioTemporal Reasoning In Driving Edge Cases) is a new benchmark designed to evaluate vision-language models on complex, safety-critical driving scenarios. It includes six task families—spatial perception, spatial understanding, temporal memory, temporal extrapolation, trajectory prediction, and scene-context awareness—using multiple-choice questions and free-form visual question answering that require detailed scene explanations and driving decision justifications. The benchmark focuses on unusual object behaviors and complex traffic interactions, and results show that even the best model, GPT‑6‑Astra, scores only 43.4% on average, highlighting significant gaps in spatiotemporal reasoning.
Shape-Aware Neural Time Fields is a physics‑informed neural motion planning framework that learns a reusable travel‑time (cost‑to‑go) representation for rigid‑body motion, explicitly accounting for distinct geometric restrictions on translation and rotation. The learned field captures local motion difficulty and is combined with a sampling‑based planner that balances progress toward the goal with the cost of reaching nearby poses. Experiments on 2D and 3D planning tasks and non‑prehensile manipulation show substantial improvements in planning success and inference efficiency over neural baselines, while remaining competitive with classical planners at lower planning times, and closed‑loop pushing tests demonstrate fast replanning for reactive execution.
The paper investigates why causal foundation models (CFMs) produce poorly calibrated credible intervals when aggregating unit-level estimates over groups. It shows that CFMs combine unit-level variances assuming independence, but posterior uncertainty induces correlations that shrink the combined variance, leading to coverage loss that worsens with group size. The authors demonstrate that neural process models, which jointly predict for groups, correct this bias and markedly improve interval coverage across synthetic and benchmark datasets.
OMEGA-Avatar is a feed‑forward framework that creates a 360° complete, animatable 3D Gaussian head from a single image. It introduces a semantic‑aware mesh deformation module to improve hair modeling while preserving topology, and a multi‑view feature splatting module that builds a shared canonical UV representation for consistent, high‑frequency detail across all viewpoints. Experiments show it outperforms existing baselines in full‑head completeness and identity preservation.
CastFlow is a role‑specialized agentic forecasting framework that combines general‑purpose tool reasoning with specialized temporal forecasting. It uses an Experience Memory to guide a frozen LLM in acquiring evidence via a multi‑view toolkit, then a lightweight specialized LLM refines predictions through evidence‑grounded reasoning. The two‑stage training strategy—supervised fine‑tuning followed by reinforcement learning with verifiable rewards—optimizes the specialized model based on final forecasting outcomes, achieving superior performance on nine out of ten real‑world benchmarks.