What I mean here is that traditional LLMs are trained on tasks irrelevant to what they will do for the user. It’s like training a plane to efficiently operate on the runway, but not to fly. In short, it is almost impossible to train an LLM, and evaluating is just as challenging. Then, training is not even necessary. In this article, I dive on all these topics.
Training LLMs for the wrong tasks
Since the beginnings with Bert, training an LLM typically consists of predicting the next tokens in a sentence, or removing some tokens and then have your algorithm fill the blanks. You optimize the underlying deep neural networks to perform these supervised learning tasks as well as possible. Typically, it involves growing the list of tokens in the training set to billions or trillions, increasing the cost and time to train. However, recently, there is a tendency to work with smaller datasets, by distilling the input sources and token lists. After all, out of one trillion tokens, 99% are noise and do not contribute to improving the results for the end-user; they may even contribute to hallucinations. Keep in mind that human beings have a vocabulary of about 30,000 keywords, and that the number of potential standardized prompts on a specialized corpus (and thus the number of potential answers) is less than a million.
More problems linked to training
In addition, training relies on minimizing a loss function that is just a proxy to the model evaluation function. So, you don’t even truly optimize next token prediction, itself a task unrelated to what LLMs must perform for the user. To directly optimize the evaluation metric, see my approach in this article. And while I don’t design LLMs for next token prediction, see one exception here, to synthesize DNA sequences.
The fact is that LLM optimization is an unsupervised machine learning problem, thus not really amenable to training. I compare it to clustering, as opposed to supervised classification. There is no perfect answer except for trivial situations. In the context of LLMs, laymen may like OpenAI better than my own technology, while the converse is true for busy business professionals and advanced users. Also, new keywords keep coming regularly. Acronyms and synonyms that map to keywords in the training set yet absent in the corpus, are usually ignored. All this further complicates training.
Analogy to clustering algorithms
As an example, this time in the context of clustering, see Figure 1. It features a dataset with three clusters. How many do you see, if all the dots had the same color? Ask someone else, and you may get a different answer. What characterizes each of these clusters? You cannot train an algorithm to correctly identify all cluster structures, though you could for very specific cases like this one. The same is true for LLMs. Interestingly, some vendors claim that their LLM can solve clustering problems. I would be happy to see what they say about the dataset in Figure 1.
For those interested, the dots in Figure 1 represent points on three different orbits of the Riemann zeta function. The blue and red ones are related to the critical band. The red dots correspond to the critical line with infinitely many roots. It has no hole. The yellow dots are on an orbit outside the critical band; the orbit in question is bounded (unlike the other ones) and has a hole in the middle of the picture. The blue orbit also appears to not cover the origin where red dots are concentrated; its hole center — the “eye” as in the eye of a hurricane — is on the left side, on the right to the dense red cluster but left to the center of the hole in the yellow cloud. Proving that its hole encompasses the origin, amounts to proving the Rieman Hypothesis. The blue and red orbits are unbounded.
Can LLMs figure this out? Can you train them to answer such questions? The answer is no. They could succeed if appropriately trained on this very type of example but fail on a dataset with a different structure (the equivalent of a different kind of prompt). One reason why it makes sense to build specialized LLMs rather than generic ones.
Evaluating LLMs
There are numerous evaluation metrics to assess the quality of an LLM, each measuring a specific type of performance. It is reminiscent of test batteries for random number generators (PRNGs) to assess how random the generated numbers are. But there is a major difference. In the case of PRNGs, the target output has a known, prespecified uniform distribution. The reason to use so many tests is because no one has yet implemented a universal test, based for instance on the full multivariate Kolmogorov-Smirnov distance. Actually, I did recently and will publish an article about it. It even has its Python library, here.
For LLMs that answer arbitrary prompts and output English sentences, there is no possible universal test. Even though LLMs are trained to correctly replicate the full multivariate token distribution — a problem similar to PRNGs thus with a universal evaluation metric — the goal is not token prediction, and thus evaluation criteria must be different. There will never be a universal evaluation metric except for very peculiar cases, such as LLMs for predictive analytics. Below is a list of important features lacking in current evaluation and benchmarking metrics.
Overlooked criteria, hard to measure
- Exhaustivity: Ability to retrieve absolutely every relevant item, given the limited input corpus. Full exhaustivity is the ability to retrieve everything on a specific topic, like chemistry, even if not in the corpus. This may involve crawling in real-time to answer the prompt.
- Inference: Ability to invent correct answers for problems with no known answer, for instance proving or creating a theorem. I asked GPT if the sequence (an) is equidistributed modulo 1 if a = 3/2. Instead of providing references as I had hoped — it usually does not — GPT created a short proof on the fly, and not just for a = 3/2; Perplexity failed at that. The use of synonyms and links between knowledge elements can help achieve this goal.
- Depth, conciseness and disambiguation are not easy to measure. Some users see long English sentences superior to concise bullet lists grouped into sections in the prompt results. For others, the opposite is true. Experts prefer depth, but it can be overwhelming to the layman. Disambiguation can be achieved by asking the user to choose between different contexts; this option is not available in most LLMs. As a result, these qualities are rarely tested. The best LLMs, in my opinion, combine depth and conciseness, providing in much less text far more information than their competitors.
- Ease of use is rarely tested, as all UIs consist of a simple search box. However, xLLM is different, allowing the user to select agents, negative keywords, and sub-LLMs (subcategories). It also allows you to fine-tune or debug in real-time thanks to intuitive parameters. In short, it is more user-friendly. For details, see here.
- Security is also difficult to measure. But LLMs with local implementation, not relying on external APIs and libraries, are typically more secure. Finally, a good evaluation metric should take into account the relevancy scores displayed in the results. As of now, only xLLM display such scores, telling the user how confident it is in its answer.
Benefits of untrained LLMs
In view of the preceding arguments, one might wonder if training is necessary, especially since the goal is not to provide better answers, but to better predict the next tokens. Without training, its heavy computations and black-box deep neural networks machinery, LLMs can be far more efficient yet even more accurate. To generate answers in English prose based on the retrieved material, one may use template, pre-made answers where keywords can be plugged in. This can turn specialized sub-LLMs into giant Q&A’s with a million questions and answers covering all potential inquiries. And deliver real ROI to the client by not requiring GPU and expensive training.
This is why I created xLLM, and the reason why I call it LLM 2.0, with its radically different architecture and next-gen features including the unique UI. I describe it in detail in my recent book “Building Disruptive AI & LLM Technology from Scratch”, available here. Additional research papers are available here.
That said, xLLM still requires fine-tuning, for back-end and front-end parameters. It also allows the user to test and select his favorite parameters. Then, it blends the most popular ones to automatically create default parameter sets. I call it self-tuning. It is possibly the simplest reinforcement learning strategy.
About the Author
Vincent Granville is a pioneering GenAI scientist and machine learning expert, co-founder of Data Science Central (acquired by a publicly traded company in 2020), Chief AI Scientist at MLTechniques.com and GenAItechLab.com, former VC-funded executive, author (Elsevier) and patent owner — one related to LLM. Vincent’s past corporate experience includes Visa, Wells Fargo, eBay, NBC, Microsoft, and CNET. Follow Vincent on LinkedIn.
