A self-evolving harness lifts the Qwen3.5-4B score from 0.16 to 0.30

The researchers examined two procedures to increase the performance of an agent for long tasks. They can change the harness around a frozen model, or they can train the weights of the model. They say that the type of failure shows which procedure to use. On the DeepPlanning benchmark, a self-evolving harness lifted the held-out score of Qwen3.5-4B from 0.16 to 0.30. Changes to the weights then cut the content failures that the harness did not repair.

Two types of failure

The team identifies each trajectory with a failure by the first signal that fires. This gives two groups.

• Procedure failures: blocked calls, loops, and used step budgets.
• Content failures: the agent gives a plan that is not good.

Harness changes repair procedure failures. Weight changes repair content failures.

The harness loop

On DeepPlanning, the loop lifted the score of Qwen3.5-4B from 0.16 to 0.30. The loop repaired procedure failures. It left content failures for the weights.

Weights and adapters

The team uses trajectories of the self-evolving harness to train LoRA adapters. An adapter adds weights to a frozen model. On the 9B model, the adapter gave the same result as the full harness loop. It cut content failures from a quarter of trajectories to one in twenty.

An adapter that the team trains on answer-shuffled trajectories gave a lower score than the model with no adapter.

Other benchmark

On WebArena-Lite, the loop gave +0.09 on tasks that it did not see before. The model gets this from the information that it sees. Adapters did not add to it.

This is a brief. We point to the report and do not rewrite it. Read it at the source below.