#compute

3 curated events

papersTODAY 04:00 UTC

Paper Proposes Multi-Armed Bandit Approach to Compute Allocation in Self-Evolving LLMs

A new arXiv preprint examines how compute is allocated during LLM-guided evolutionary search, noting that prior work typically reports only the best result from many runs rather than the full distribution. The authors propose reframing the depth-versus-breadth tradeoff as a multi-armed bandit problem. The work is listed as a cross-list replacement across arXiv's AI and machine learning sections.

research/llm/compute/evolutionary-search

papersTODAY 04:00 UTC

Review questions whether AI scaling results justify resource allocation

A critical review examines scaling research on systems that pair a pretrained model with retrieval, search, verification, tools, and interaction. It argues that a better score achieved with a larger budget does not by itself indicate where additional resources should be directed. The paper calls for evidence that ties scaling outcomes to specific allocation decisions.

scaling/evaluation/compute/research

papersSEP 10 04:00 UTC

Researchers propose compute-bounded framework for AI security assurance

A new arXiv paper introduces a resource-constrained approach to security assurance, arguing that extra inference compute raises the number of correctly resolved tasks without automatically delivering distinct coverage, accepted evidence, or operational protection. The authors separate quantities such as repeated success, unique coverage, and verification to clarify what compute actually provides. The framework is intended to guide coverage, verification, and response decisions when compute budgets are limited.

security/compute/evaluation/safety