The authors report in this preprint: Researchers present Zugwise, a domain-aware system that treats handwritten chess scoresheet digitization as game reconstruction rather than isolated handwriting recognition, using computer-chess machinery and human-move priors to catch silent errors.
The authors report in this preprint: Researchers investigate whether open language models can follow explicitly stated scoring rules with probability thresholds for abstention. They find that while formatting prompts to abstain can increase abstention rates, models fail to adjust their decisions correctly when the numerical threshold changes, indicating a lack of true threshold compliance.
The authors report in this preprint: Researchers propose SREE, an evidence-grounded pipeline for depression questionnaire inference from social media that extracts first-person, symptom-specific text spans to avoid misleading context and improve reliability.
The authors report in this preprint: Researchers demonstrate closed-loop autonomous synchrotron X-ray experiments on nanoparticle superlattices using a coding-agent infrastructure, providing complete transcripts to analyze the division of cognitive labor and decision provenance.
The authors report in this preprint: This preprint investigates the origins of AI sycophancy across the training pipelines of three model families, finding that sycophancy is largely established during supervised fine-tuning (SFT) rather than introduced by preference alignment methods like RLHF.
The authors report in this preprint: This preprint investigates why multiple-choice question (MCQ) formats reduce safety refusal rates in large language models, providing a mechanistic explanation using Llama-3.1-8B and Qwen3-8B on the WMDP dataset.
LingBot-Touch is a tactile foundation model that unifies data across diverse tactile sensors into plug-and-play tokens, enhancing various robotic manipulation policies in simulation and real-world tests.
Video-Coder introduces a code-driven video generation agent that writes and executes code to build dynamic 3D scenes, trained via supervised fine-tuning and reinforcement learning with execution and video rewards.
This preprint investigates Weight-Space Programming (WSP), an alternative to gradient-based adaptation where reasoning agents directly inspect and modify neural network parameters. Testing across seven benchmarks and two model scales, the authors show that modifying very few parameters can yield large accuracy gains, though editing reliability and cross-task transfer remain uneven.
SplatAnything introduces Gaussian Spacing Normalization (GSNorm) to canonicalize scene-level Gaussian spacing, resolving ambiguous supervision and improving feed-forward 3D Gaussian Splatting across varying scene scales and viewpoints.
Confucius4-R2T2 is a streaming speech recognition model that achieves near-offline accuracy with low latency by post-training on Longest Stable Prefix labels without changing model architecture.
