Anthropic unveils Claude Meetings, Whiteboards and Sheets in push into…

The accumulated papers (Microsoft Research Asia, Tsinghua University) demonstrate that failures in language model libraries frequently originate from peripheral software components like drivers, package dependencies, and resource management rather than from core computation logic. Most recently, direct manual analysis of 300 GitHub issues across seven libraries has identified systematic failure patterns in the surrounding stack, quantifying how non-calculation code accounts for the majority of reported problems.

The accumulated papers establish that scaling language models under fixed data and compute budgets requires careful trade offs: repeating data causes overfitting and can make models worse (Cornell University, Princeton University with Q Labs and New York University), while masking inputs (University of Michigan with KAIST) and compression for inference (ETH Zurich) offer ways to stretch limited resources. The line of inquiry has moved most recently toward practical recipes such as stopping large models early (Johns Hopkins University with OpenAI) and scheduling variable length generation efficiently (École polytechnique with Huawei Technologies and Shenzhen Research Institute of Big Data).

The accumulated papers establish that process decomposed benchmarks, from CyberBinGym (National University of Singapore, UC Riverside) to Multi step SWE (Klavis AI) and SLATE (Amazon, Georgia Tech), consistently reveal that AI agents fail at subgoal decomposition, tool orchestration, and long chain execution across cybersecurity, software engineering, physics, and business workflows. Most recently, benchmarks like InfoOrgBench and FrontierPhysics have shifted toward evaluating hierarchical organization and end to end research tasks, while ToolSpace and AceAction 5M propose automated methods to synthesize or annotate such process level tasks.