Tamaththul3D is a specialized reconstruction pipeline for Arabic Sign Language (ArSL) that generates high-fidelity 3D avatars from monocular video. The authors introduce the first high-quality 3D parametric annotations for the Ishara-500 Saudi Sign Language dataset, providing precise SMPL-X parameters for 500 culturally authentic SSL signs. Their pipeline integrates SMPLer-X, WiLoR, and MediaPipe, achieving up to 32% improvement in hand accuracy over previous methods via kinematic-chain-based wrist alignment, hybrid swing-twist decomposition, and 2D-supervised joint optimization. This establishes the first comprehensive framework for high-fidelity ArSL avatar reconstruction, enabling new accessibility technologies and cultural preservation for the Arab Deaf community.

egenioussBench is a visual localisation benchmark built on city-scale airborne 3D mesh and CityGML LoD2 models, providing centimetre-accurate smartphone images with map-independent ground truth, designed to scale beyond SfM-based approaches.

Open-SAT is a training-free query embedding refinement algorithm that operates at inference time, leveraging Large Language Models (LLMs) to improve alignment between user queries and satellite image content. It achieves up to 16.04% F1 improvement on three public benchmarks without additional training or supervision.

ViTok-v2 is a novel Vision Transformer autoencoder that addresses key limitations of existing ViT tokenizers through native resolution support via NaFlex and a DINOv3 perceptual loss for stable scaling. Trained on approximately 2 billion images and scaled to 5 billion parameters, it is the largest image autoencoder to date. ViTok-v2 matches or exceeds state-of-the-art reconstruction at 256p and outperforms all baselines at 512p and above. In joint scaling with flow matching generators, ViTok-v2 advances the Pareto frontier of the reconstruction-generation trade-off.

A density-aware calibration method using latent query features from DETR-style detectors to jointly recalibrate classification and regression uncertainties under distribution shift, consistently outperforming standard post-hoc methods on both camera and LiDAR detectors.

A counterfactual explanation method based on GANs with cyclical-consistent loss is proposed for class-oriented feature attribution in medical images. Evaluated on synthetic, tuberculosis, and BraTS datasets, it overcomes limitations of discriminative models. The study also introduces a method to generate plausible counterfactual instances and evaluate their quality.

Researchers propose a layout-aware representation learning method for open-set ID fraud detection. Trained only on US IDs, it achieves 99.83% layout classification accuracy on Canadian IDs and discovers 276 adaptive fraud cases, including 222 missed by existing detectors. The method supports campaign-scale fraud discovery under distribution shift.

Large language models face threats from hidden malicious intent in multi-turn dialogues, where attackers distribute harmful objectives across benign-looking turns. The proposed TurnGate monitor detects the earliest turn where delivering a response would enable harmful action, and the MTID dataset supports training and evaluation. TurnGate outperforms baselines in detection while maintaining low over-refusal rates and generalizes across domains and models.

Large Language Models (LLMs) excel at generating contextually appropriate responses but remain poorly calibrated for multi-party conversations, where deciding when to speak is as critical as what to say. This paper introduces When2Speak, a grounded synthetic dataset of over 215,000 examples from 16,000 conversations with 2-6 speakers, and a four-stage pipeline to train intervention timing. Supervised fine-tuning (SFT) improves Macro F1 by 60% on average for 4B+ models but remains over-conservative with a Missed Intervention Rate (MIR) of 0.50. Reinforcement learning with asymmetric reward shaping reduces MIR to 0.186-0.218 and increases recall from 0.479 to 0.78-0.81, showing that temporal participation is a trainable dimension of conversational AI.

Researchers identify and formalize 'recorruption' in multimodal RAG, where adding accurate context causes models to ignore correct predictions. They propose BAIR, a parameter-free framework that restores visual attention and applies positional penalties.

A study compares domain-trained small language model Olava Extract against frontier LLMs for structured contract extraction. Olava achieves macro F1 0.812, micro F1 0.842, with 78-97% cost reduction and fewer hallucinations, challenging the need for large models.

A new study shows that repeatedly rewriting text generated by diffusion language models (up to five times) can effectively bypass watermark detection, reducing detection rates from 87.9% to 4.86%. The study tested four open-weight language models and five rewriting styles, finding that even a single rewrite reduces detection to 14%-41%, and multiple rewrites are even more effective.

ReaComp compiles a small set of LLM reasoning traces into reusable symbolic program synthesizers using coding agents. These solvers require no LLM calls at test time, achieving 91.3% on PBEBench-Lite and 84.7% on PBEBench-Hard, outperforming LLMs with test-time scaling by 16.3 points with zero inference cost. They also complement LLM search, improving PBEBench-Hard accuracy from 68.4% to 85.8% while reducing token usage by 78%, and transfer zero-shot to historical linguistics tasks with 80.1% accuracy.

SLAM is a novel white-box watermarking scheme that uses sparse autoencoders to identify directions in the residual stream encoding linguistic structure, and causally steers those directions during generation to embed a watermark without compromising text quality. On Gemma-2 2B and 9B, it achieves 100% detection accuracy with only 1-2 reward point quality cost, far below existing methods.

This paper proposes an evidence-based model to automatically generate query keywords from query-free datasets for Query-Focused Summarization (QFS). Experiments show that summaries generated using evidence-based queries achieve competitive ROUGE scores compared to those from original queries.

An intervention study investigates the use of counterarguments in writing for critical thinking, comparing human and AI assessments. With 36 students and 6 rubrics, results show self-written counterarguments contain logic, and frontier LLMs can effectively assess student work with high alignment to human judgments.

AdaGATE is a training-free evidence controller for multi-hop RAG that frames evidence selection as a token-constrained repair problem. It combines entity-centric gap tracking, targeted micro-query generation, and a utility-based selection mechanism balancing gap coverage, corroboration, novelty, redundancy, and relevance. On HotpotQA, it achieves best evidence F1 of 62.3% on clean data and 71.2% under redundancy injection, using 2.6x fewer tokens than Adaptive-k.

Sparse Autoencoders (SAEs) effectively disentangle feature superposition in transformers, but compositional steering suffers from geometric instability. This paper introduces a high-dimensional cone manifold framework, derives an asymptotic collapse threshold, and shows that ReLU rectification causes a ratchet-like interference accumulation. Experiments on CLEVR validate the predictions.

Standard transformers apply equal layers to all tokens. Token-Selective Attention (TSA) introduces a learned per-token gate that skips layers for easy tokens, saving up to 23% of operations with minimal quality loss and no architectural changes.

A new non-neural framework, DVBL, learns basis functions directly from data via variational optimization, offering interpretability and mathematical transparency compared to neural networks.

This paper introduces MidSteer, a general affine framework for concept steering in generative models. The authors formalize the theory of concept steering, link the standard approach to LEACE, propose LEACE-Switch, and then introduce MidSteer. Experiments show favorable performance across vision diffusion models and LLMs.

A new paper introduces sparse prefix caching that leverages recurrent layers' ability to resume from a single state, optimizing LLM inference by storing exact recurrent states at sparse checkpoints. An O(NM) dynamic program determines optimal checkpoint placement, improving latency over standard heuristics on real-world tasks.

This paper introduces horizon-constrained Rashomon sets, establishing the first rigorous connection between chaos theory and predictive multiplicity. It proves that the effective Rashomon set contracts exponentially with forecast horizon, and develops decision-aligned model selection algorithms that improve decision quality by 18-34% in chaotic systems.

A self-supervised PINN framework with a learnable blending neuron to adaptively balance physics and data losses, plus transfer learning. Validated on liquid-metal heat sink heat transfer with only 87 CFD points, achieving <8% error, outperforming baselines.

SAT introduces a coordinator-free training paradigm for multi-LLM teams, providing monotonic improvement and plug-and-play guarantees. Experiments show a team of three 4B agents surpasses Qwen3-32B by 3.9% on AIME benchmarks, and swapping in two 8B agents boosts performance by 10.4%.

Researchers leveraged nationwide longitudinal EHR data from the All of Us Research Program to predict chronic rhinosinusitis (CRS) diagnosis using two years of pre-diagnostic history. A hybrid feature-selection pipeline compressed ~110,000 candidate codes into 100 interpretable features. Demographic-stratified models were trained across six adult sex and life-stage subgroups with subgroup-specific hyperparameter tuning, achieving an overall AUC of 0.8461—an improvement of 0.0168 over the best baseline. The study demonstrates that routinely collected EHR data can support population-representative CRS risk stratification and inform earlier triage and referral prioritization in primary care.

This paper challenges the notion that flat minima cause better generalization in neural networks. It proposes 'weakness' as a reparameterization-invariant measure that predicts generalization, showing through experiments that flatness and simplicity are less reliable indicators.

This paper proposes constant-context skill learning, a framework that moves reusable procedures into lightweight modules, reducing prompt tokens by 2-7x while achieving strong performance on multiple benchmarks.

LaTA is an open-source autograder that runs entirely on local hardware, using a LaTeX-native workflow and a four-stage pipeline (ingest, segment, grade, report) with a locally hosted open-weight chain-of-thought LLM (gpt-oss:120b) to grade student assignments. Deployed at Oregon State University for ME 373 with ~200 students, it achieved $0 marginal cost per assignment, a grading error rate of ~0.02-0.04%, and improved exam scores by 11% (midterm) and 8% (final) compared to traditional grading, along with significant gains in student confidence.

This paper introduces FinAgent-RAG, an agentic RAG framework for financial document QA, which uses iterative retrieval-reasoning loops with self-verification. It incorporates a contrastive financial retriever, program-of-thought reasoning, and an adaptive strategy router, achieving state-of-the-art results on three benchmarks while reducing API costs by 41.3%.