By assigning numbers to choices and reading probabilities in the GLM-5.3-Flash model, we can now implement fast and accurate decision-making models without additional training.
Imagine you ask your AI assistant, “For lunch today, between Kimchi Stew, Bibimbap, and Tonkatsu, which is better?” Until now, the AI would have spent time adding a long explanation, detailing everything from the ingredients in the Kimchi Stew to the nutrients in the Bibimbap. But we are entering an era where AI can instantly calculate the correct answer and the probability of that answer being chosen, much like solving a quiz.
Recently, researchers successfully implemented a ‘Jev’-style decision model using the latest AI model, ‘GLM-5.3-Flash,’ without any separate complex training Reference 1.
Why is this important?
Many choices we make in daily life sometimes require the help of AI. However, for a company, having an AI generate long sentences every time can be inefficient in terms of cost and time. The technique introduced here allows AI to make decisions as quickly and clearly as a human picking an option, even calculating the probability behind the decision.
In particular, GLM-5.3-Flash is the first native multimodal model (a method where various data such as text, images, and audio are understood and processed simultaneously) in the GLM-5 series to directly handle visual information Reference 9, Reference 14. This means it can quickly answer questions like “What is the best choice in this situation?” by looking at photos of the scene, not just text questions Reference 2.
Making it easy: The Librarian Metaphor
Let’s explain the principle of this technique with a metaphor. Let’s call the Transformer (the core design structure of AI that grasps the relationship between words in a sentence) model a ‘librarian looking for answers in a huge library.’
The traditional method is like asking the librarian to bring books, summarize them, and even add an opinion. It takes a long time and the conversation becomes long. The new method is much more intuitive:
- Numbering: Clearly assign numbers to choices A, B, and C for the question.
- Prefilling: Have the librarian (AI) prefill the first letter of the answer sheet.
- Reading Probabilities: Peek at the probability distribution (log probabilities, the value quantifying the likelihood that the model will choose a specific word) of the word the librarian is about to write next.
By doing this, you can instantly get a conclusion like “There is a 90% chance of choosing A” without the AI writing a long, rambling sentence Reference 2. The biggest advantage of this method is that there is absolutely no need to retrain or fine-tune the model from scratch Reference 3, Reference 5.
Current Situation
It is already achieving results in the field. The decision model using GLM-5.3-Flash showed almost equivalent accuracy to ‘Jev,’ an existing professional decision-making AI, on 28 text datasets Reference 2, Reference 7.
The speed is also surprising. On average, it takes about 156ms (0.15 seconds) to make a single decision, and the cost is very low at 0.06 Euros per 1,000 decisions Reference 4, Reference 7. Of course, there is a limitation that accuracy drops slightly if the number of choices becomes too large, but it performs powerfully enough in general situations Reference 10.
What will happen in the future?
In the future, AI will become a smarter and more efficient ‘decision-making partner.’ Users will be able to make choices with more confidence because the AI will go beyond just providing an answer to informing them how confident it is in its answer (confidence values, an indicator of how much the AI trusts its own response) Reference 10.
We may soon experience an AI in a shopping app telling us instantly, “There is a 95% probability that this clothing matches your usual style.” Such attempts to turn the ‘intelligence’ of AI into the ‘efficiency’ of real-time services will happen in more places in the future.
MindTickleBytes’ AI Reporter’s Perspective: The progress of technology is not just about making bigger and heavier models. We have entered an era where the real skill is how ‘wisely’ you utilize the smart models that already exist.
References
- Turn GLM-5.3-Flash into a Jev-like System One model
-
[GLM-5.3-Flash Matches Jev’s Decision · Hacker News Zeli](https://zeli.app/story/49857656) - Turning GLM-5.3-Flash into a Jev-like decision model
- Turn GLM-5.3-Flash into a Jev-like System One model - LinkedIn
- GitHub - nokia-applied-research/AnyJev: Turn any LLM into a Jev-style …
- GitHub - zhengxuyu/litjev: Turn any off-the-shelf LLM into a Jev -like …
-
[Turn GLM-5.3-Flash into a Jev-like System One model Lorenz Tabertshofer](https://de.linkedin.com/posts/lorenz-tabertshofer_turn-glm-53-flash-into-a-jev-like-system-activity-7508887669553262592-U_Dg) - GLM5.3Flash— ВАЙБКОДИНГ ЗА КОПЕЙКИ! - YouTube
- zai-org/GLM-5.3-Flash· Hugging Face
- GLM-5.3-FlashMatchesJev’sDecisionAccuracy in a Single Forward…
- Можно ли запуститьGLM-5.3локально: честный расчёт по железу
- Z.ai - Advanced AI Chatbot & Agent powered byGLM-5.3-Flash
- GLM5— Next-Gen FrontierModel
- GLM-5.3-Flash: Technical Deep Dive into Z.ai 320B-A18B Hybrid …
-
[Jev Is Turning Into an Entire Ecosystem Swati Gupta …](https://x.com/hrswatigupta/article/2102741642050666755) - GLM-5.3 - openlm.ai
- Retrain the entire model from scratch
- Assign numbers to choices and read probabilities
- Use only image data
- No need for additional model training (fine-tuning)
- Infinite savings in computing costs
- Requires an internet connection
- Understands text only
- It is the first native multimodal GLM-5 model
- It is too slow for real-world use