Search the full wire by company, model, lab, or keyword. Every story we have ever aggregated.
Study diagnoses structural mathematical understanding in LLMs via Mathematical Primitives benchmark, proposing methods to improve post-training on mathematical reasoning.
ZFO framework decouples direction and step-size selection for LLM fine-tuning by using gradient info and zeroth-order evaluations along a trusted first-order direction.
Adding AI models to local applications requires a portable model format, a reliable runtime, and acceleration that works across target systems. Do Inference Now... Adding AI models to local applications requires a portable model format, a reliable runtime, and acceleration that works across target systems. Do Inference Now (DIN) Deploy is an open-source collection of practical C++ samples that bridges that gap. It combines ONNX Runtime with the NVIDIA TensorRT RTX execution provider to help developers move from a model checkpoint to a native… Source
IDiom protein language model trained on 54M intrinsically disordered regions from AlphaFold DB to enable functional design of unstructured protein domains.
DMAD reframes distribution matching as adversarial classification to accelerate visual generation with fewer sampling steps and lower memory than prior distillation methods.
Higher-order Grammar Representation lifts molecules to combinatorial complexes to encode ring systems and motifs for improved generative and foundation models in chemistry.
LoopCD training-free contrastive decoding for Looped Transformers exploits intermediate recurrent states as weak predictions to improve token selection without auxiliary models.
SoftServe family of quasi-Newton methods for deep learning derives positive-definite curvature estimates without line searches to handle non-convex optimization at scale.
Generative Cinematographer enables joint authoring of camera and 3D object motion in editable scene scaffolds to resolve ambiguity in 2D motion control for video generation.
Analysis of multi-teacher on-policy distillation in Qwen3-1.7B and SmolLM3-3B reveals how loss averaging, token length, and domain weighting implicitly shape teacher signal strength.
Graph neural networks on tissue-specific protein interactomes show effective resistance degenerates to inverse degree, limiting uncertainty quantification signal.
Language model self-repair after ablation stems from pre-existing gains; ablations are uncalibrated points on counterfactual contrast axes.
Robot agents infer unknown hardware constraints of partners via observation, then coordinate zero-shot on new manipulation tasks.
AutoCompact trains coding agents to autonomously decide when and what context to compress during long-horizon repository-level tasks.
DuoMind coordinates multi-robot systems via semantic communication between VLM orchestrators and VLA action models for long-horizon manipulation.
Intrinsic reward criteria based on counterfactual information acquisition distinguish exploration quality across count-based, prediction-error, and empowerment methods.
Underdamped Langevin sampling with momentum preconditioning converges on non-convex, non-Lipschitz potentials via spectral taming.
Opus 5.5’s biggest tell is the word “dependable,” which pops up 23 times more often than in human samples.
LLM agents develop persistent source-specific competence through repeated access to authoritative knowledge sources on knowledge-intensive task sequences.
Faynt 10M/75M Transformer policies master Super Smash Bros. Melee across 26 characters, winning 98.4% same-character matchups vs. specialist releases.
Keyword-matching tool-use benchmarks produce false positives in small models; diagnostic ladder reveals 661.6M Spanish LM lacks genuine tool execution.
SFT with on-policy sampling achieves stronger generalization than RL on new tasks while avoiding catastrophic forgetting, challenging conventional posttraining wisdom.
MIRTO standardizes evaluation choices for unsupervised brain MRI anomaly detection with registration-gated validation and explicit reporting of thresholds and metrics.
Strongly polynomial algorithms for robust MDPs via LP formulations with rational polyhedral uncertainty sets, advancing theoretical complexity analysis.
Many applications in statistics, economics, and physics require sampling from high-dimensional categorical distributions with local dependence structures. Examples include finite memory language models, Ising and Potts systems in statistical physics and protein folding, etc. In modern machine learning, discrete diffusions have emerged as a flexible approach for sampling such data, with strong empirical performance. Motivated by this, we develop learning methods with end-to-end sample complexity bounds for discrete diffusion with uniform noising under local dependence, which we model through l...
We explore catastrophic forgetting in the context of large pre-trained models. By considering forgetting as a geometric problem in the input space of each weight matrix, we uncover a natural retention objective under which updates produced by gradient-based optimizers are suboptimal. Following this observation, we propose Local Support Learning (LSL), a general-purpose framework that augments gradient-based training for retention of prior capabilities without access to prior data. During a new learning phase, LSL pairs two components with distinct roles: a standard weight adapter, trained as ...
Real-world enterprise data science and analytics workflows require reasoning across dozens of tables, performing statistical analyses, and acting on the results. Established text-to-SQL benchmarks evaluate query generation alone, and audits have found their answer keys frequently wrong. Because real enterprise warehouses are too sensitive to release, these benchmarks are built on public datasets where a business event fits in a single table. We introduce Argo-Bench, an evaluation framework comprising 210 data science and analytics tasks. Drawing on public data, peer-reviewed industry literatu...
On-policy self-distillation has recently emerged as an effective approach for improving language-model reasoning by supervising students with a frozen or EMA version of themselves that receives privileged information. Its application to multimodal large language models (MLLMs), however, remains largely unexplored. Recent approaches use privileged visual information, such as image crops corresponding to a question, to improve fine-grained perception, but their gains are confined to tasks that benefit from such visual zooming and require either human-annotated grounding data or external teacher...
A comparative explainability framework is presented to audit DeBERTa-v3 under zero-shot classification of medical abstracts. The work addresses the disagreement problem in Explainable Artificial Intelligence, where different attribution methods produce divergent explanations for the same input and prediction. A natural language inference engine is implemented over the Medical Abstracts corpus with five enriched hypotheses per diagnostic category and a balanced sample of one thousand texts per class. Five explanation methods are compared: SHAP and LIME as model-agnostic approaches, occlusion a...
Mechanistic interpretability aims to recover the internal computations responsible for model behavior. Progress in automated circuit discovery is often framed as a search problem: better attribution or optimization should identify better mechanisms. This assumes that the evaluation objective can recognize a better circuit once it is found. We show that intervention-defined faithfulness can instead prefer an equally sized circuit that reproduces the model's behavior less well, creating an objective-level recovery gap. Across four human-reference tasks and InterpBench, we compare validation fai...