What matters in AI.

Subscribe

Research

A 240-task benchmark shows that agent increases stay on one file type

The top configuration got 42.25 in all and 30.62 on tasks with more than one file type.

Claimed, not confirmed

EvoCowork is a benchmark of 240 office tasks. The tasks use presentations, documents and spreadsheets. With the GLM 5.3 model, 2 methods increased in-domain scores by 4.90 and 2.64 points. They changed scores on tasks with more than one file type by only +0.03 and -1.72.

Score change after self-improvement ACE and AHE gain 4.90 and 2.64 points in-domain, but change cross-office scores by only +0.03 and -1.72, with the same MSA harness and GLM 5.3. Score change after self-improvement In-domain tasks Cross-office tasks ACEAHE +4.90 +2.64 +0.03 -1.72 Same MSA harness, GLM 5.3 Gains did not transfer.
With GLM 5.3 and the same MSA harness, ACE and AHE improved in-domain scores but barely moved, or lowered, cross-office scores.

Sources

  1. EvoCowork: A Benchmark for Real-World Office Tasksarxiv.org
AI MATTER · NEWS · AI MATTER · NEWS ·15 OCT2026

Posted

Tags

More in Research

All Research news