XNews

Evolver Framework Boosts Agent Task Success to 87.50% and Cuts Tokens 24.24%

Evolver Framework Boosts Agent Task Success to 87.50% and Cuts Tokens 24.24%

Researchers introduced Evolver, an experience-driven framework that stores procedural knowledge from agent exploration as structured Gene assets and reuses it across existing harnesses without updating model parameters. Evaluated with Claude Code and Codex on Terminal-Bench 2.1, SWE-bench Multilingual, SWE-rebench and GAIA, Evolver raised task success from 83.13% to 87.50% on 320 paired runs while reducing recorded token consumption by 24.24%. Among 265 pairs solved under both conditions, token use fell 17.57%.

Event version 1.

Key insights

Gene assets store procedural experience

Evolver stores procedural knowledge from agent exploration as structured Gene assets. It collects them during exploration, retrieves relevant ones at the start of a task, and revises them based on feedback.

No model parameter updates

Evolver supports experience-conditioned execution across existing agent harnesses without updating foundation model parameters. It was evaluated with Claude Code and Codex.

Higher success and lower token use

On a selected diagnostic set of 320 paired executions, task success improved from 83.13% to 87.50%, a gain of 4.38 percentage points. Recorded task-execution token consumption decreased by 24.24%.

Efficiency gain on solved tasks

Among 265 pairs successful under both conditions, token consumption decreased by 17.57%. The efficiency benefit appears even when task success is preserved.

Diagnostic set caveat

The evaluation set was selected for diagnosis. Transfer to unseen tasks is listed as future work.

Source posts