Skip to content
Trending storyDeveloping

UCLA and Northeastern study finds 18% of AI agent output tokens are removable waste

1 article1 sourcesince Nov 9Last article 14h ago ·

Overview

AISummary of 1 article

A UCLA and Northeastern University paper estimates that about 18% of output tokens from AI agents across 6 frontier models and 4 benchmarks are removable waste.

The authors used an LLM judge to mark waste spans, then reran the agents with those spans deleted. Rohan Paul, who reported on the paper on X, says pruning cut 41% of the waste from Kimi-K3's reasoning on SWE-bench Verified, with success at 89.5% against 89.0% without pruning.

Written by AI from the articles below · updated Nov 9, 9:24 PM ET

Check the sources:

Latest developmentNov 9, 7:14 AM ET
UCLA and Northeastern study finds 18% of AI agent output tokens are removable waste

Article timeline

The articles in this story. Times are ET.

Nov 9
  1. Rohan PaulX
    UCLA and Northeastern study finds 18% of AI agent output tokens are removable waste

    AIA UCLA and Northeastern University paper estimates that about 18% of AI agent output tokens across 6 frontier models and 4 benchmarks were removable waste. Using an LLM judge to mark spans and rerunning agents after deletion, the authors report pruning cut 41% of waste from Kimi-K3's reasoning on SWE-bench Verified with success at 89.5% versus 89.0%.

    Image from @rohanpaul_ai's post

Heat Trend

Not enough continuous observations to show a trend yet.