The study demonstrates how frontier AI models can be used to reason through reaction networks and generate experimentally validated hypotheses for catalyst selectivity.
A new study proposes a decision-level metric called correctness agreement to better capture the behavioral changes in large language models caused by post-training quantization.
This paper provides a comprehensive survey of methods, datasets, and benchmarks for multimodal machine unlearning across various data modalities.
DreamCharacter-1 is a post-adaptation framework designed to refine pretrained 3D foundation models for high-fidelity, production-ready character generation.
Researchers introduced KVpop, a method that compresses key-value caches in autoregressive decoding by learning a fixed-budget eviction policy through direct supervision.
This study provides a theoretical characterization of the mechanisms behind delayed generalization, or grokking, in neural networks using a stochastic-geometric framework.
Researchers evaluated the effectiveness of distilling structured text extraction capabilities from an 8-billion parameter reasoning model into a sub-1-billion parameter on-device model.
This paper provides a mathematical formulation of slow thinking and active perception to guide the design and training of reasoning-focused language models.
Researchers developed ScaleMOF, a dataset and fine-tuning strategy that enables large language models to predict and prioritize viable candidates for scaling up metal-organic framework syntheses.
PhasorFlow is a new open-source Python library designed to facilitate complex-valued computations on the unit circle manifold.
Researchers have developed Theoria, a verification architecture that audits large language model outputs by rewriting solutions into typed state transitions to ensure correctness.
This paper proposes a stepwise reinforcement learning approach to prevent text and image modalities from diverging during complex, interleaved multimodal reasoning tasks.
The LoKA framework introduces low-precision FP8 kernels tailored to the numerical sensitivities and communication demands of large-scale recommendation models.
Researchers propose a framework that combines large language model generation with neural architecture search to automate the design of neural networks in open-ended spaces.
Researchers have introduced Infinity-Parser2, a multimodal document parsing model trained using a controllable data-synthesis pipeline and reinforcement learning.
Researchers developed a method to adaptively generate bias-eliciting questions to better identify and evaluate inherent biases in large language models.
This research analyzes how deployment-time memory configurations in foundation-model agents affect personalization utility, data extraction risks, and deletion fidelity.
Researchers have introduced DrugGen-2, a generative language model designed to design small molecules based on both target protein sequences and disease ontology.
Researchers introduced a two-dimensional curriculum learning framework that categorizes alignment difficulty by prompt complexity and pairwise distinguishability to optimize direct preference optimization.
DocMaster is a document analysis system that preserves the hierarchical structure of complex files to improve retrieval and question-answering performance in large language models.
The WCog-VLA framework integrates semantic world forecasting and generative world evolution to enable proactive decision-making in end-to-end autonomous driving.
This study introduces a method that uses multimodal language models to visually evaluate reinforcement learning agent behaviors to design more effective training curricula.
Probing the internal activations of forecasting language models reveals that their hidden representations contain more accurate and better-calibrated probability estimates than their generated text.
The article discusses the role of reconciliation in design-driven automation for data center infrastructure.