ai4science

该论文提出 Continuum-LM，将预训练残差变换复用为连续深度语言模型中的动力学算子，通过 ODE 数值积分控制计算量，在八项科学任务上较离散基线分别在 0.6B 和 1.7B 规模提升 7.77 和 10.96 个百分点的平均准确率。

The tweet discusses the appeal of RSI positions in AI research compared to FDE, noting the potential for high salaries in the field of auto research and ai4science.

The article introduces AgentActionBench, a process-oriented benchmark for evaluating LLM agents in reproducing experiments from scientific papers across ML and AI4Science domains.

Fun fact: new online side-job opportunities — find questions that Doubao answers incorrectly in biology/physics/math and get 400-600 yuan per item; find history questions Gemini gets wrong and get 600-800 yuan per item, reflecting the value of high-quality AI4Science samples.

Anima Anandkumar announces four Lean-related papers from their group at ICML workshops, covering verified ML systems, functional program synthesis, proof assistant interoperability, and scientific reasoning, positioning Lean as infrastructure for AI.

This paper presents SCION, an agentic scientific operating system that integrates AI tools for scientific discovery through a Research Execution Plan (REP) and hierarchical multi-agent execution. It demonstrates applications in materials analysis, molecule design, and protein screening, outperforming existing autonomous research-agent baselines.

Proposes DDIAgents, a mechanism-conditioned multi-agent framework for drug-drug interaction prediction that dynamically routes relevant biomedical knowledge to specialized expert agents and aggregates their analyses, outperforming existing feature-based, graph-based, and LLM-based methods.

This paper introduces SciRisk-Bench, a benchmark for evaluating the safety of large language models in AI4Science contexts, covering 7 disciplines, 31 subdisciplines, and 10 risk dimensions to assess both scientific competence and risk awareness.

LakeFM is a foundation model for aquatic systems, pre-trained on large-scale ecological datasets to forecast lake dynamics using irregular multivariate multi-depth time series data, achieving competitive performance compared to existing models.

Introduces SciPaths, a benchmark for forecasting the enabling contributions required to realize a target scientific discovery, and evaluates frontier and open-weight language models, finding significant room for improvement in reasoning backward from contributions to enabling building blocks.

MeasHalu is a novel framework for mitigating scientific measurement hallucinations in LLMs through a two-stage reasoning-aware fine-tuning strategy and progressive reward curriculum. It introduces a fine-grained taxonomy of measurement-specific hallucinations and demonstrates improved accuracy on the MeasEval benchmark.