Iterative Summarization using LLMs

2.1 Blog Data

First up are the blog post experiments shown in Figure 1 (never mind the 42 in the label, that's just the random seed) where the X-axis is the iteration number and the Y-axis is the length of the response. For these experiments, I only ran the "Summarize" prompt. The input data to this is a summary by gpt-4o of a scrape of my blog, which you can look at here. It's about 7.6 KB in size compared to all my blog posts which are closer to 500KB.

Figure 1: Length of response from LLM's for iterative summarization across {gpt-4o, gpt-3.5-turbo} and temperatures of {0.0, 0.5, 1.0}.

In the first iteration, the LLM drops a lot of bytes cutting the size down by more than half in all cases. Across almost all runs, the LLM's converged to a pretty short response. Here's an example from gpt-4o_42_0.5 run:

The author, an AI and data science expert affiliated with BorealisAI and the University of Toronto, blends their professional knowledge with interests in sports, music, and technology to offer varied reading suggestions. They are a respected speaker, patent holder, and researcher whose reflections from 2020-2022 emphasize family, career growth, humility, and mentorship. At Rubikloud, they prioritized people and learning, contributing to its acquisition by Kinaxis. Throughout the pandemic, they maintained routines and optimized their work environment. Their 2019 review highlights professional achievements, personal health, team management, and includes an inspiring story about Mozart.

I mean as a summary of my entire blog in less than 1000 bytes, it's not terrible. But the original summary, which was already greatly condensed from all my writing, is 7x the size. The only outlier is gpt-4o at temperature=0.0, here's the summary it gave:

The author is a Research Director at BorealisAI and an Adjunct Professor at the University of Toronto, specializing in Electrical and Computer Engineering. He enjoys sports, music, technology, and economics, and values simplicity and flexibility. His expertise includes AI for hardware design automation and teaching data structures and algorithms. He recommends "Outliers" by Malcolm Gladwell and "The Intelligent Investor" by Benjamin Graham.

His contributions to data science and AI include talks, patents, and papers, focusing on machine learning, decision automation, and AI in retail, with notable patents in user behavioral segmentation and influencer analysis.

In his yearly reviews, he reflects on professional and personal milestones:

• 2022: A stable year with family time and professional highlights, emphasizing mindfulness and quality time.

• 2021: Adaptation to the pandemic, starting at BorealisAI, and hopes for better social connections and fitness.

• 2020: Significant changes like his daughter's birth, buying a house, and Rubikloud's acquisition by Kinaxis, with challenges in maintaining hobbies and fitness.

Reflecting on six years at Rubikloud, he emphasizes the importance of people, learning, and resilience, with the acquisition by Kinaxis marking a new chapter. Lessons from the pandemic include adjusting routines, managing stress, and improving remote work efficiency.

The 2019 review covers professional highs and emotional lows, contributions to Rubikloud's growth, and challenges in work-life balance and burnout. Teaching was fulfilling, and personal hobbies progressed, with aims to improve health and fitness in 2020.

He reviews books like "Tuesdays with Morrie" and "Radical Candor," offering insights on cherishing life, honest feedback, and learning strategies. A humorous story about Mozart highlights self-sufficiency and the importance of self-motivation and independence.

It's a respectable 2000 characters long and still maintains some of the original format. Perhaps what is most interesting here is that it looks like that gpt-4o reached a fixed point in this configuration. The plateau in Figure 1 in fact did produce the same summary over and over again. This obviously could only really happen at temperature 0 where each run is (mostly) deterministic. gpt-3.5-turbo showed similar behavior with a shorter summary at this temperature but it did drop a few words at iteration 14 and again at iteration 49. This is likely because of the slight non-determinism of how OpenAI runs the models even with a fixed random seed.

At any other larger temperature, we would expect deviation, which explains all the wiggles on the other runs. We also see that the other gpt-4o runs with larger temperature did have similar length summaries in the first few iterations, but quickly devolved into much shorter ones where I assume it could not make it more concise. The gpt-3.5-turbo run's response length drops pretty quickly after the first iteration though.

All these results aren't too surprising except for the "near" fixed point that we saw with gpt-4o, temperature 0. It just seems unlikely to me that it would reproduce the exact same text instead of modifying a word here or there. These LLM's are mysterious blackboxes indeed.

2.2 Random Fact Data

In this set of experiments, I first asked gpt-4o to generate 100 random facts that I would then use as input data, which I put on Github. Similar to above, I then asked the LLM to either summarize or rephrase the input data over and over again. Figure 2 shows the results in terms of lines (where each fact is on a line).

Figure 2: Number of lines of response from LLM's for iterative summarization across a sample of experiments from {gpt-4o, gpt-3.5-turbo} and temperatures of {0.0, 0.5, 1.0} and two different prompts.

In this chart I only included a sample because all of the other experiments were pretty boring -- they just mirrored the majority, which were able to retain all 100 lines of facts. The two outliers were gpt-3.5-turbo with the rephrasing prompt and gpt-4o with the summary prompt, both at temperature \(1.0\). The gpt-3.5-turbo at iteration 35 randomly decided to drop half of the facts and spit out a list of only 50 lines long. Not sure what happened here, but I guess it just randomly decided to stop! The gpt-4o run decided to drop the list format entirely on iteration 3 and just summarize the list with a short paragraph, which obviously dropped a lot of information.

Unsurprisingly both happened at temperature \(1.0\), and maybe slightly surprisingly, not all runs at \(1.0\) had this issue. There were two other runs paralleling the ones above but with the opposite prompt that kept all 100 facts. This is just another good reminder that LLM behavior is indeed random and the randomness scales with temperature. Consequently, they are not easy to control at all.