---
title: 'In the News: September 10, 2026 (Extra 3)'
description: 'In a Salesforce study, full-trajectory imitation from a stronger model cut weaker-model accuracy by 4 to 30 points across all seven tasks under an evolved harness.'
canonical_url: 'https://darkfactory.dev/news/2026-09-10-extra-3'
markdown_url: 'https://darkfactory.dev/news/2026-09-10-extra-3.md'
collection: news
date_published: '2026-09-10T16:20:00-04:00'
date_modified: '2026-09-10T16:20:00-04:00'
---

# In the News: September 10, 2026 (Extra 3)


In a Salesforce study, full-trajectory imitation from a stronger model cut weaker-model accuracy by 4 to 30 points across all seven tasks when paired with an evolved harness.

## 1. Full-trajectory imitation cut weaker-model accuracy across all seven tasks

**[Co-Evolving Harnesses and Models: On-Policy Correction Helps Weaker Models Catch Up Where Imitation Fails](https://arxiv.org/abs/2609.09134)** · Zhou Yu, Bin Bi, and nine coauthors, Salesforce AI Research · arXiv, September 8, 2026

The Salesforce AI Research team tested harness evolution and fine-tuning on seven enterprise agent tasks. A weaker model first evolved its own agent harness, including the system prompt, tool set, execution hooks, and context-management scaffolding. A stronger expert model used that harness more effectively, so the researchers fine-tuned the weaker model on the expert's complete trajectories.

Accuracy fell on all seven tasks, by 4 to 30 points across Qwen3-Coder and Gemma 4. The same fine-tuning procedure helped when the harness had not been evolved.

The authors attribute the regression to a mismatch between planning and execution. Imitation transfers the expert's planning strategy without the competence to carry it out, disrupting what they call "model-harness fit." The weaker model then no longer matches a harness evolved around its native planning style.

The authors propose an on-policy correction pipeline run by a meta-level agent. It finds the turn where the weaker model's own rollout fails and asks the expert to rewrite only that turn, preserving the rest of the weaker model's planning style.

**Why it matters:** Teams pairing harness evolution with distillation should test the techniques together. In this study, full-trajectory imitation reduced accuracy even though the same fine-tuning procedure helped without an evolved harness. The authors' alternative corrects a single failed turn while preserving the weaker model's planning style.