From Simple to Surprising: A peep into the Emergent Behavior of Large Language Models

A key paper in the field of NLP (Natural Language Processing) titled "Emergent Abilities of Large Language Models" by Wei et al. (2022) sparked significant discussion in the scientific community. The research explored a fascinating phenomenon: large language models (LLMs) exhibited evolving abilities not seen in smaller models. This was linked to the increasing scale of the models, measured by the number of layers and parameters. The study also investigated the impact of scaling datasets to larger volumes on this behavior.

Few researchers employed a technique called "chain-of-prompt/thoughts querying," particularly for tasks involving arithmetic calculations to validate the emergent characteristics. The experiments were carried out on the vanilla and the scaled-up models. They found that increasing model parameters allowed LLMs to tackle complex problems by breaking them down into a series of smaller, more manageable steps, compared to the smaller models, which it was trained for. This finding raises a captivating question: could further increases in model size lead to even more advanced capabilities in the current day LLMs?

This concept of emergent behavior isn't new. It's a familiar phenomenon observed across various scientific disciplines like physics, biology, and chemistry.

A classic example is water. Its physical state (liquid, gas, solid) depends on a quantitative factor: temperature. Increasing temperature transforms liquid water into vapor (gas), while decreasing it solidifies it into ice. This shift from one distinct state to another exemplifies the conversion of quantitative changes into qualitative ones.

Another analogy often used is water ripples. Beyond a certain point, small ripples on the surface can transform into powerful tides. These examples highlight a key point: emergent behavior typically arises when a system reaches a certain level of complexity.

The figure (Fig. 1) from the Wei et al. paper [1] illustrates a fascinating concept. It shows a tipping point for various tasks given to an LLM model. On the x-axis, we see the increasing scale of the model. Up to a certain point, the model's performance remains random. However, beyond this critical point, there's a dramatic improvement in performance.

This phenomenon is similar to certain electronic devices like semiconductor diodes. In a forward-biased diode, there's a "forward voltage" threshold. Once the voltage exceeds this point, there's a sudden surge in current flow, essentially diode acting like a short circuit.

The concept of emergent behavior was further explored in a large-scale experiment called "Big Bench" conducted by 130 institutions in 2022 [3]. This project tested various LLMs, including GPT and LaMDA (Fig. 2), on a wide range of 204 tasks. The results confirmed the presence of emergent behaviors in these models.

However, a late 2023 study by researchers at Stanford [2] challenged the idea of emergent behavior in LLMs. They argued that these behaviors might be an illusion caused by the specific performance metrics used to evaluate the models. In other words, the apparent emergence of new abilities could vanish depending on the task and the metric selected.

The Stanford team suggests that increasing model parameters simply gives LLMs more flexibility to solve problems, but the dramatic improvement observed in certain metrics might not be universally applicable. But which metric is best for meaningful evaluation of a LLM was left answered.

This lack of a definitive metric highlights the ongoing need for research in this area. The possibility of unpredictable behavior in LLMs remains, warranting further investigation.

If these models can indeed exhibit emergent behavior based on experience with various tasks, it raises intriguing questions about their potential to learn and evolve analogous to humans. However, if scaling is the primary driver of cost, there's still room to explore efficiency improvements to mitigate these expenses.

References:

[1] Jason Wei, Yi Tay, Rishi Bommasani, Colin Raffel, Barret Zoph, Sebastian Borgeaud, Dani Yogatama, Maarten Bosma, Denny Zhou, Donald Metzler, Ed H. Chi, Tatsunori Hashimoto, Oriol Vinyals, Percy Liang, Jeff Dean, William Fedus. “Emergent Abilities of Large Language Models”. Transactions on Machine Learning Research (TMLR), 2022. arXiv:2206.07682 [cs.CL].

[2] Aarohi Srivastava, Abhinav Rastogi, Abhishek Rao, et al. “Beyond the Imitation Game: Quantifying and extrapolating the capabilities of language models.” Transactions on Machine Learning Research (TMLR), May 2023

[3] Rylan Schaeffer, Brando Miranda, Sanmi Koyejo. “Are Emergent Abilities of Large Language Models a Mirage?” arXiv preprint arXiv:2304.15004, last revised 22 May 2023.