Leaderboards, eval harnesses and ablations — the contested business of deciding which model is actually better.
arXiv:2606.19828v2 Announce Type: replace Abstract: 3D multimodal large language models (3D MLLMs) describe a 3D object as a whole but cannot address, name, or reason about its parts. Prior part-awar...
arXiv:2606.29845v3 Announce Type: replace Abstract: Flicker-Banding (FB) arises from temporal aliasing between a camera's rolling shutter and a display's brightness modulation, degrading screen-captu...
arXiv:2607.14836v2 Announce Type: replace Abstract: Sign Language Production (SLP) generates continuous signing from a linguistic input and drives accessibility technologies such as signing avatars....
arXiv:2611.05966v1 Announce Type: cross Abstract: Prior-data fitted networks (PFNs) have become state-of-the-art for tabular prediction but remain underexplored in other areas of statistics. In this...
The paper studies how epistemic uncertainty estimates can be used in different ways—either as an additional input feature or as a per‑sample confidence weight in the loss function—to improve 3D LiDAR point cloud segmentation. Experiments on four large‑scale LiDAR benchmarks (ISPRS Vaihingen, AHN3, Paris‑Lille‑3D, and Toronto‑3D) show that uncertainty‑aware models outperform a baseline ANN in mean Intersection over Union, recover minority classes, and better align prediction error with estimated uncertainty. The results suggest that uncertainty can serve multiple roles in the segmentation pipeline beyond merely boosting accuracy.
The study benchmarks how hierarchical text embeddings and quantization affect performance across 38 tasks using four transformer models. It evaluates 168 configurations combining 1D/2D hierarchical representations with 32‑bit, 8‑bit, and binary quantization, reporting storage and evaluation time separately. Results show int8 quantization is nearly as effective as float32, binary quantization degrades performance, and dimensionality reduction paired with int8 can cut storage by 32× with minimal loss.
The paper introduces fast and stable methods for triangular matrix inversion used in Delta-Rule linear transformers, a key component of efficient long‑context models such as Qwen3.5/3.6 and RWKV‑7. It systematically compares direct and iterative inversion algorithms, emphasizing numerical stability, computational complexity, and hardware efficiency, and validates the approaches with rigorous experiments in low‑precision settings. Benchmarks on NPUs show up to a 4.3× speed‑up over current SGLang implementations, improving layer performance while preserving model accuracy.
The study evaluates whether overlap metrics like Dice can reliably indicate metabolic tumor volume (MTV) estimation accuracy in PET segmentation models. Across five FDG PET/CT models, higher Dice scores correlated with lower relative MTV error on average, but at the patient level, 16.8% of cases fell on opposite sides of clinically relevant thresholds. Additionally, many models missed lesions, showing that model ranking alone does not guarantee reliable individual MTV estimates.
FedXCrop is a federated learning pipeline for crop disease classification that preserves farmer privacy while providing post‑aggregation explainability. The system uses a MobileNetV2 model and achieves near‑centralized accuracy (98.81%–99.38%) across five clients, even under non‑IID label skew. Grad‑CAM and SmoothGrad explanations are quantitatively evaluated, showing that federated models largely retain explainability characteristics of centralized models, though agreement decreases with increased heterogeneity.
The study presents a conditional diffusion-based generative model that augments T2-weighted prostate MRI data with clinically significant prostate cancer (csPCa) and transition zone (TZ) lesions. By extending MAISI‑v2 with csPCa class, zonal, and lesion‑mask conditioning and fine‑tuning on 1,303 patients, the model achieves a substantial reduction in Fréchet inception distance (FID) from 13.28 to 2.04. Synthetic augmentation improves csPCa classification AUC across five simulated training scenarios, though the benefit does not significantly transfer to a cross‑institutional dataset.
AGORA is an item‑response‑theory aggregator that improves the reliability of large language model judges by learning each judge’s sensitivity and rating thresholds from all items jointly, without needing human labels. It places every item on a latent quality scale, allowing judge reliability to be inferred from the data itself. Across three benchmarks, AGORA matches or surpasses the Dawid‑Skene method, especially when evaluation sets are small or judge pools are uneven.
ModalSearcher introduces an autonomous cross‑modal search framework that lets an agent dynamically choose among direct reasoning, text‑to‑text search, text‑to‑image search, and reverse‑image search, actively incorporating visual evidence as needed. The system builds an interleaved multimodal knowledge graph via a scalable pipeline that samples diverse paths and uses backward question construction to prevent shortcut reasoning. Trained with supervised fine‑tuning and reinforcement learning, ModalSearcher‑8B outperforms Qwen3‑VL‑8B by 16.2 accuracy points across five multimodal knowledge‑intensive QA benchmarks and shows consistent gains on active visual‑search tasks.
The paper introduces D3D, a decoupled diffusion drafting method that eliminates the need for target-model hidden states during speculative decoding. By combining blockwise denoising with auxiliary autoregressive supervision, D3D can draft successor blocks asynchronously, achieving significant speedups—up to 7.48× on GSM8K with Qwen3-32B. The authors also prune layers to maintain performance while reducing computational cost, demonstrating over 99% of mean acceptance length after removing five of 36 layers.
The paper proposes a temporal organization layer for continual agents, introducing phase‑tagged micro‑epochs, a duty cycle, and nested clocks to separate exploration from commitment. It presents deterministic and stochastic guarantees for duty‑cycle updates, defines busy/idle execution contracts, and evaluates the design through synthetic experiments and a Qwen3‑8B prototype. The results demonstrate a testable scheduling design and clarify its conditions of validity.
ConceptARC at Scale introduces a generator that produces thousands of ARC-style tasks focused on specific spatial and semantic concepts, expanding the original benchmark by tenfold. The new dataset, ConceptARC-100, contains 1,600 manually reviewed tasks across 16 concepts, each with 100 tasks. Evaluation of large reasoning models on this dataset reveals that they reliably solve certain concepts, struggle with others, and develop internal representations that cluster by transformation operations such as recoloring, removal, or translation.
Decision Models (DMs) are a new class of AI models that provide fast responses and output a probability distribution over a fixed set of options. In a first evaluation, 10 DMs were compared with 8 Large Language Models (LLMs) across three economic settings—a household spending survey, classic economic games, and a labour‑market field experiment. Frontier DMs, those scoring highest on general decision‑making benchmarks, reproduced human response variation more closely than LLMs in all settings, matched LLMs on modal responses in the spending survey, and were roughly ten times faster and over a hundred times cheaper than the fastest LLM.
Bivariate Additive Models (BAMs) are a new class of Generalized Additive Models with pairwise interactions that constrain each shape function to a simple asymmetric bivariate Gaussian form, producing unimodal contour maps that are easier to interpret. The model automatically selects feature pairs during training and achieves predictive accuracy comparable to standard GAMs while approaching the performance of more complex GA$^2$Ms on six tabular benchmarks. Experiments also illustrate the trade‑off between accuracy and the number of components that need to be inspected.
The paper introduces a new task, Intra-Architecture Source Attribution for Stable Diffusion (IntrAA‑SD), which focuses on distinguishing between fine‑tuned versions of the same Stable Diffusion XL architecture. It presents a 140K‑image benchmark dataset with in‑domain and out‑of‑distribution prompts, and proposes a baseline method using a style‑conditioned Mixture‑of‑Experts network that leverages a frozen CLIP backbone and a LoRA‑adapted DINOv2 network. The authors evaluate their approach on attribution accuracy and conduct ablation studies, providing code and data publicly on GitHub.
The paper introduces the Graph Signature Index to quantify the mismatch between positive-definite geometric priors in many Graph Neural Networks and the negative distances often needed for heterophilic graphs. Using this signature, the authors propose the Pseudo‑Riemannian Attention Network (PRAT), which learns a signed squared pseudo‑distance metric that can capture both attractive and repulsive relationships. Experiments on fifteen node‑classification benchmarks show that PRAT matches or surpasses state‑of‑the‑art heterophily‑adapted models, and ablation studies confirm that the sign structure provided by the signature is essential for its performance.
The paper introduces Continual Prior Compensation (CPC), a temporal continual learning framework that trains human motion prediction models in stages by dividing predictions for different future moments into subtasks. CPC mitigates forgetting of prior knowledge using a learnable Prior Compensation Factor (PCF) and is theoretically justified. An enhanced version, CPC++, adds Fine‑Grained PCF to estimate prior loss per subtask, and both frameworks are shown to be flexible, easily integrated with various backbones, and effective across three benchmark datasets.