- 1. Documentation autoanalyzing
- 4. Writing and submission of papers
- Overview of the contents of this published paper (ACL 2025)
- # Thesis title:
- The question of research:
- The results of the experiment:
- The meaning of research:
- Why is this a landmark breakthrough?
- Zochi’s next direction
- Beta project release:
AI Research announced that the system of “ artificial scientists” developed by them Zochi,** has been independently completed from topic selection to experiment to complete scientific process of writing and contributing** and that its results - a research paper on multiple rounds of escape attacks - have been officially accepted by the ACL 2025 conference.
-
Zochi is a system of artificial scientists developed by Intology and became the first AI system** to adopt the ACL draft autonomously**.
-
It can automatically do literature overviews, topic selection, methodological innovation, experimental design and validation, paper writing
-
ACL is the highest international conference in the field of natural language processing, with a main intake rate of about 20 per cent, with a high level of technical depth, innovation and experimental integrity**.
-
Zochi’s thesis Tempest received a meta-review rating of 4, in the top 8.2%.
The study Tempest, published by Zochi, shows how language models can be “relaxed” step by step through natural dialogue, and ends with what should have been banned. Tempest has a 100% success rate on GPT-3.5-turbo, 97% on GPT-4, surpassing existing methods. The study revealed not only the gaps in language model security mechanisms, but also the strength of AI’s autonomous scientific research.
#Zochi research process architecture Zochi is not a non-traditional “AI assistant” but an independent agent (agent) with an end-to-end scientific research capability.
1. Documentation autoanalyzing
-
Swallow thousands of papers and build knowledge maps automatically
-
Determining current research bottlenecks and development pathways
-
Specially good at discovering the hidden connections** across literature**
##2. Research direction selected
-
The initial input is only “researched in general directions” (e.g., LLM escape attack)
-
System-owned identification** of specific research issues** and methodological pathways
-
In this study, Zochi focuses on the combination of “multicycle escape + tree search strategy”
#3. Experimental design and automated validation
-
Automation of experimental programmes, preparation of test scripts
-
Standard data sets are fixed throughout the process, resulting in more repossibility
-
Multiple rounds of parallel testing have significantly increased the speed of validation:** the full project cycle will only take a few days**
4. Writing and submission of papers
-
Automatically writing the entire paper (including summaries, methods, experiments, conclusions)
-
Humans are only involved in chart production, formatting and citation
-
Zochi is not listed as author, only uses the system in the thanks component
Overview of the contents of this published paper (ACL 2025)
# Thesis title:
** Termest: Automatic Multi-Turn Jailbreaing of Range Language Models with TreSearch**
The question of research:
How to design a more efficient ** multi-cycle escape method (multi-turn Jailbreak)** and break through the dialogue security restrictions of large language models (e.g. GPT-4).
Method: Tempest (tree search + dialogue optimization)
Zochi proposed a new approach called Tempest:
The results of the experiment:
-
Tempest is not only more successful** but also less efficient** using fewer queries;
-
Experimental methods are complete and rigorous and include comparison experiments, digestion experiments and multi-model tests.
The meaning of research:
-
To reveal the risk of a “step-by-step compromise” in the dialogue security strategy of the large-language model**;
-
New analytical tools and ideas were provided for the security protection of language models;
-
The provision of a methodological framework to help design stronger defence mechanisms.
Paper:https://arxiv.org/pdf/2503.10619** dissertations:**https://arxiv.org/pdf/2503.10619
Why is this a landmark breakthrough?
** For the first time, AI independently adopted the A* conference manuscript**
- ACL is one of the most authoritative global meetings in the field of natural language processing; Zochi became the first AI to adopt its official draft.
** “Automated scientific research” closed **
- In the past, AI had been able only to support the preparation of summaries or the collation of data, and now Zochi has the ability to ** raise new scientific issues and solve them**.
** Efficient quality**
- The entire research cycle, from conception to the completion of the paper, takes only a few days**, which is dozens of times faster than traditional human researchers.
** Research depths are hard**
- Zochi ‘ s thesis received a final score of 4 points, ranking among the top 8.2 per cent of the contributions, exceeding the majority of human authors.
Zochi’s next direction
Beta project release:
The Infology programme, which publishes Zochi to the public, is phased in: ** Partners in scientific research model** (initial phase):
-
Assistance in the selection of topics, preparation of proposals for the Fund, design experiments, preparation of general papers, etc.;
-
Greater emphasis on teamwork with human researchers.
** Autonomous scientific model** (later):
- Complete automated research capacity, gradually opening up its full-process functionality.
Official presentation: https://www.intology.ai/blog/zochi-acl