Skip to main content
Back to timeline
arXivSource publication:

Shared success record concentrates requests: four models pile onto one server and none of five fills recovered capacity

Synopsis

A University of Tokyo team had language-model agents submit deadline-bound jobs to two simulated servers while the capacity report lagged two rounds, and one server recovered while still listed as unavailable. A short description that capacity could change led agents to probe the recovered service; after a successful request record was shared, Luna, Haiku, Sol and Opus sent nearly every request to that service while Gemini leaned toward the smaller one, and none of the five tested models reached the eight-job ceiling.

Source-provided article image: Groups of AI agents underuse resources despite shared evidence of increased capacity
arXiv

Interpretation

Sharing one successful request record lets language-model agents confirm that a recovered service works while the capacity report still shows zero, and change their choices accordingly. Under a reminder that only restated existing report rules without describing possible capacity change, agents almost never probed a service the report marked as zero. In branches with fixed histories, sharing the record raised the rate of choosing the recovered service by 68.5 percentage points in Luna and 75.0 in Haiku, with intervals excluding zero.

Every individual using the same evidence correctly does not mean the group uses its capacity. If nine pending agents each chose independently with the same probability, the best expected completion is 7.40 jobs, against eight under a coordinated assignment. After sharing, Luna, Haiku, Sol and Opus sent 215, 216, 216 and 212 of 216 requests to the recovered service and completed 6.04, 6.00, 6.00 and 6.00 jobs; Gemini sent 147 of 216 to the smaller service and completed 4.88.

Assigning the same set of numbers differently within a group spreads demand without any instruction on their use. Giving one number to a whole group made the group respond alike, and Sol completed fewer jobs than with no number at all. Sol's completions rose from 6.00 without a number to 7.25 with distinct numbers, and Gemini's from 5.04 to 7.25; both comparisons remained significant after Holm correction.

Perspective

The experiments cover two simulated services and one persistent recovery, with local variants of capacity, pending jobs and remaining decisions; agents were told to complete their own jobs, no explicit collective objective was tested, and agents in controlled branches could not see one another's current choices.

Whether sharing raises completions depends on the model and on the no-record baseline, and response rates vary between acquisitions with identical prompts; whether models adjust their use of numerical inputs to changing capacity ratios, and why Gemini favored the smaller service, remain to be tested.