Intel / article

Pistis Technical Report

Source: arxiv.org
Author: Heyun Chen, Xiaohan Lan, Jiaxi Li, Zhilin Lu, Qi She, Weiwen Xu, Fei Yu, Yujie Zhong, Jinghuan Chen, Zijian Feng, Siyu Jiao, Yiheng Lin, Xinhao Wang, Sihan Yang, Jieyu You, Changbin Zhang, Hengyu Zhang, Xudong Zhang, Yunqing Zhao, Shuai Zheng
Date: Published Sep 25, 2026

Context

Pistis-Auto-Harnessing tests whether an AI search can improve through changes to its surrounding software while its model stays fixed. During development only, an optimization agent examines runs of the current software, proposes changes to prompts, routing rules or workflow, and replaces the current version only when the chosen development metric improves. The selected software is then frozen before testing.

On the reported held-out multimodal-search benchmark (VDR-testmini), accuracy rose from 26.8% to 28.6%, or 134 to 143 of 500 questions, under the same limit of 15 environment interactions. The result attributes a modest improvement to how the complete system organizes the model's work, without requiring updated model weights or a larger interaction allowance.

Key takeaways · AI-distilled

Pistis comes in 27B and 9B multimodal models built on Qwen3.6 and Qwen3.5 respectively, starting from large-scale multimodal supervised fine-tuning.
IDRL alternates on-policy and reinforcement learning inside one training loop, instead of running either alone or combining them in a static joint loss; the authors credit it with steadier optimization and better credit assignment on long agentic trajectories.
Each size ships as Pistis-Thinking, for deep multimodal reasoning, and Pistis-Agentic, which adds agentic trajectory data for planning and and is reported to be particularly strong at multimodal search.
Pistis-Auto-Harnessing iteratively optimizes the agent's inference and, per the authors, improves performance without updating model weights or increasing the interaction budget.

Terms in this piece · Glossary

AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
distillation — Training a small, cheap model to imitate a big one's outputs, keeping much of the capability at a fraction of the cost.
tool use — A model's ability to call external functions — run code, search the web, edit files — instead of only generating text.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.

Read the source arxiv.org