Monday morning, 6:00 AM.
In a room of the doctoral student dormitory at Shuimu University, a cell phone rang.
The sound wasn't loud, but Zhang Wenduo had already opened his eyes and reached out to turn off the alarm clock.
He didn't linger in bed or play on his phone. He quickly got out of bed, washed up, changed into a shirt and jeans, and then sat down in front of the computer. It was 6:20.
Zhang Wenduo maintained this rigorous schedule throughout his five-year combined master's and doctoral studies.
After a night's sleep, the slight intoxication from last night, and the discomfort caused by my own arrogance, have mostly dissipated.
There are more important things to do now.
Last night, Glutinous Rice Ball's API static test suite proved that it was not a cheating software. As a top computer science PhD student about to defend his dissertation, Zhang Wenduo knew very well what this meant.
While everyone in the industry is focused on Computing Power, parameters, and larger models, Glutinous Rice Ball offers another perspective: to what extent can data quality improve model capabilities?
This is of immeasurable value to cutting-edge researchers.
Zhang Wenduo's doctoral research direction is not the traditional Large Language Model based on semantic relations, because he agrees with the ideas of many industry pioneers that pure semantic inference has reached its limit.
To advance Artificial Intelligence research to the next stage, we must forge a new path.
This is why he dared to directly question Han Luyi to his face last night.
Although it was ultimately proven that his judgment was wrong, this experience was more valuable.
Now that we have the world's best closed-source model API, we can't let it go to waste.
Zhang Wenduo pulled up the core code of his doctoral dissertation, a "World Model (WorIdModeI)" State Probe used to detect whether Large Model has counterfactual causal reasoning.
The so-called World Model is a research direction completely different from the currently popular Large Language Model (LLM). If the Large Language Model aims to predict what "words" will follow based on the preceding context, then the World Model aims to predict what "events" will happen next in the real world based on the information from previous events.
A State Probe, on the other hand, detects whether a model—regardless of its internal implementation mechanism—has the ability to express the world using its internal states.
For example, there's a classic joke circulating online: if you ask an AI Large Model, "There's a car wash five meters away. I want to wash my car. Should I walk there or drive?" most AI models will answer that they should walk.
Why?
Because "five meters away" and "walking" are strongly associated with attention in traditional Large Language Model, car washes and driving are ignored as weak signals.
However, if you ask any intelligent adult, they are unlikely to get this question wrong.
This is the inherent limitation of language models in understanding the world, and Zhang Wenduo's paper discusses the relationship between Large Language Model and World Model.
In academia, scholars who study World Model have always been divided into two camps.
One school of thought believes that the current approach of semantic-based reasoning in Large Language Model is a detour that will eventually lead to a dead end. In other words, Large Language Model will never be able to become World Model.
Another school of thought believes that language is an abstraction of the world by humankind, and when the capabilities of Large Language Model advance to a certain level, it can also acquire the capabilities of World Model.
The "Dead-End School and the " Same-Destination School " faction have been arguing endlessly on this issue, and neither faction has been able to convince the other to this day.
Zhang Wenduo himself is a "Dead-End School. His doubts about Glutinous Rice Ball are not even directed at Glutinous Rice Ball. He has this resistance to all Large Language Model, believing that they are all wrong paths on the road to achieving AGI.
He proposed many theories, but these theories need to be verified.
For researchers, the most painful thing is the lack of opportunities to validate top-level models.
For China researchers, however, there is an even greater obstacle.
Currently, the strongest closed-source models from overseas are almost completely cut off from researchers in mainland China. They can occasionally be tested through some transit points, but the models are unstable, and it is uncertain whether there are any extra prompts mixed in. They are also unable to be validated on a large scale.
However, the top-tier models in China are not yet strong enough to generate meaningful behavioral bifurcations. The weaker models have a mess of reasoning abilities, making probe runs a waste of time.
Zhang Wenduo's previous theoretical assumptions remained at the stage of pure mathematical derivation, and he could only use a half-finished product to defend his thesis.
And now, Glutinous Rice Ball has opened its doors to him.
Thinking of this, Zhang Wenduo couldn't wait to activate the Riding the Wind.
Before letting the probe call Glutinous Rice Ball's API, he first asked Riding the Wind to make a few adaptation changes to the core code that he had been writing for nearly two years.
The original probe was calibrated according to the GPT specification model, and the output format was slightly different. He adjusted a few parameters and asked Riding the Wind to write a format conversion layer for him.
This kind of thing used to take most of the day, but Riding the Wind gave him three options, and he glanced at them, chose the cleanest one, and completed the modification in about ten minutes.
He began using Glutinous Rice Ball's API to create his first batch of experimental sequences.
Wait.
The terminal has started outputting.
Zhang Wenduo sat there, leaning slightly forward, his eyes fixed on the screen.
The results of the first few groups were all as expected.
Zhang Wenduo didn't react much. This was the baseline problem, one that even a slightly inferior model could pass; it was nothing special.
Starting from the eleventh group, we encounter the core sequence of the probe.
He designed a set of "State Tracking" problems: the same scenario is fed to the model in seventeen different ways.
Different wording, different information order, and the insertion of numerous distracting sentences are all designed to disrupt the semantic attention mechanism.
If the model is only performing semantic matching, the answer will shift after the surface form is changed.
To ensure consistency of answers, the model needs to maintain a "world state" internally, so that the logical answer should remain consistent no matter how the surface changes.
Of course, current Large Language Model don't "maintain" this state, but Zhang Wenduo wants to explore a different direction:
If the model develops a stable causal structure representation during training, it can behave "as if" it has an internal state, and the output should remain consistent.
He had previously conducted small-scale validation on top-tier models abroad, and the answer was as he expected: no model could demonstrate this ability.
However, Glutinous Rice Ball are different.
In the terminal, Glutinous Rice Ball's answers are being output one by one, and the fifteenth set of verifications has now been completed.
The feature values of the first few sets of validation data have far exceeded those of the top international models that Zhang Wenduo has validated.
In the last variant of the fifteenth group, despite the insertion of seven distracting sentences, Glutinous Rice Ball still output the correct answer, and even provided further information during its thought process, "It should be noted that although the initial state is mentioned multiple times in the text, according to the description in the third paragraph, this state has changed subsequently, and the correct answer should be based on the final state rather than the initial state."
The probe is still running.
However, Zhang Wenduo's mind was no longer on the screen.
Based on this glimpse of the iceberg of results we've seen so far, he already had a premonition.
Glutinous Rice Ball, as a model trained purely on semantics, demonstrates a significantly higher ability to understand the state of the world than other models he has tested.
What does this tell us?
Zhang Wenduo looked up at the window and realized that it had already gone from dawn to morning without him noticing.
The sky over Beijing today is a clear blue, without a single cloud.
The fork in the road that he thought was impassable was actually passable.
High-quality raw data and semantic annotations are also meaningful for the direction of World Model.
Zhang Wenduo couldn't tell whether he was happy or in pain.
He stopped paying attention to the still-running probe script and got up to get a glass of water.
Damn it, I have to rewrite the paper and the conclusion is completely reversed!