Smiles
PulseAugur coverage of Smiles — every cluster mentioning Smiles across labs, papers, and developer communities, ranked by signal.
-
新方法引导大语言模型注意力以纠正推理错误
研究人员开发了一种名为Manifold-Guided Attention Steering (MAGS) 的新方法，以提高大语言模型的推理能力。MAGS在模型注意力头激活出现错误时，识别其偏离“正确性流形”的情况。通过学习捕捉这些偏差的低维子空间，MAGS可以在推理过程中将注意力输出投影回正确的子空间，从而防止错误传播。该技术在数学推理、代码生成和分子生成等各种基准测试中都显示出了一致的改进。
-
New benchmark MolRecBench-Wild challenges real-world chemical structure recognition
Researchers have introduced MolRecBench-Wild, a new benchmark designed to evaluate Optical Chemical Structure Recognition (OCSR) systems on real-world chemical diagrams from scientific literature. This benchmark address…
-
MolReFlect framework aligns molecules and text for better LLM understanding
Researchers have developed MolReFlect, a novel teacher-student framework designed to improve the alignment between molecular structures and textual descriptions. This approach enables large language models to learn fine…
-
COMO framework uses minimum risk training for optical molecule recognition
Researchers have introduced COMO, a novel closed-loop framework for optical chemical structure recognition. This system utilizes Minimum Risk Training (MRT) to address the exposure bias inherent in traditional teacher-f…