Feed

81 materials

Sort by
News11 November 2026

OpenCollab: programmable agent coordination, controlled execution environment, and policy compliance metric

The framework brings together the specification of agent interaction structures, a shared environment for experiments, and detailed event tracing, while the newly introduced Adherence metric reflects how closely actual execution follows the declared schema. In benchmarks, the Duo configuration achieves a new state of the art, while the single-agent variant uses the fewest tokens.

OpenCollab: programmable agent coordination, controlled execution environment, and policy compliance metric
Read
Comparisons6 November 2026

Neptuna: Comparing ML Architectures for Compressible Multiphase Flows with Discontinuities

Introducing Neptuna, a large-scale benchmark with 3 TB of computations for bubble collapse and droplet breakup under shock waves, evaluating convolutional, spectral, transformer, and pretrained PDE models. The takeaway: there’s no universal leader, while composite loss functions and SoftAdapt help better preserve interfacial boundaries and reproduce the spectrum more accurately.

Neptuna: Comparing ML Architectures for Compressible Multiphase Flows with Discontinuities
Read
News27 October 2026

CARAT: Testing Whether Materials Science LLMs Generate Structure or Paraphrase the Prompt

This work proposes a protocol that fixes the question and gold-standard answer across eight ways of presenting structure, while separately marking links in GraphSpace. Experiments with pairwise fine-tuning, target-field masking, and evidence augmentation show that the model can learn both how to apply a relation and how to follow a link, although removing the link reduces accuracy below that of heuristic traversal methods.

CARAT: Testing Whether Materials Science LLMs Generate Structure or Paraphrase the Prompt
Read