# Evolutionary Strategies lead to Catastrophic Forgetting in LLMs Source: https://www.youtube.com/watch?v=c6yJEBqewvU Recap page: https://rapidrecap.app/video/c6yJEBqewvU Generated: 2026-02-02T14:04:49.658+00:00 --- ## Quick Overview Evolutionary Strategies (ES) lead to catastrophic forgetting in Large Language Models (LLMs) when applied to sequential tasks, demonstrated by ES-trained models failing to retain knowledge from prior tasks, unlike gradient-based methods which maintain performance across tasks. **Key Points:** - Evolutionary Strategies (ES) cause catastrophic forgetting in LLMs when applied to sequential tasks, unlike gradient-based methods. - ES models showed a drastic drop in performance on the initial task (math game 'Countdown') after being trained on a new task (image editing/masking). - The forgetting magnitude for ES was 1000 times larger than for gradient-based methods (like GRPO) on the same task sequence. - The researchers suggest ES models fail because they lack mechanisms to compartmentalize knowledge, causing new learning to overwrite old knowledge globally. - ES relies on random noise/perturbations for updates, which aggressively affects all parameters, whereas gradient methods focus updates more locally. - The study used a small population size (30) for ES evaluation, which the authors note may exacerbate instability but still shows a fundamental issue. - The core challenge is developing an ES variant that can selectively update parameters to avoid forgetting while learning new skills, similar to biological evolution's balance. ![Screenshot at 00:24: The visual graph illustrates the core finding, contrasting the performance of the ES model \(represented by the lower, flatter curve\) against a gradient-based method \(higher curve\) on the original task after sequential learning, highlighting ES's catastrophic forgetting.](https://ss.rapidrecap.app/screens/c6yJEBqewvU/00-00-24.jpg) **Context:** The video analyzes a recent research paper from UC Berkeley/Abadi group that investigates the stability and generalization of Evolutionary Strategies (ES) when used for continuously learning and updating Large Language Models (LLMs). The core problem discussed is 'catastrophic forgetting,' where a model loses performance on previously learned tasks after being trained on new ones. The paper compares ES methods against standard gradient-based optimization methods (like GRPO) to see how they handle sequential task learning. ## Detailed Analysis The video discusses research demonstrating that Evolutionary Strategies (ES) cause catastrophic forgetting in LLMs when applied to sequential learning tasks, unlike gradient-based methods. The core finding is that ES aggressively updates the entire model structure based on random noise perturbations, leading to the overwriting of previously acquired knowledge. Specifically, when ES-trained models were tested on the initial math game 'Countdown' after being updated on an image editing task, their performance plummeted, showing a memory degradation 1000 times greater than gradient-based methods like GRPO. The research suggests that because ES updates are global and rely on random noise, they lack the necessary compartmentalization mechanisms found in gradient-based methods, which can selectively update parameters. While gradient methods also suffer from forgetting, their performance shift is much less severe, maintaining a baseline competence. The researchers posit that for ES to be viable for continuous learning, a mechanism must be developed to mimic biological evolution's ability to retain old skills while acquiring new ones, rather than completely eroding past knowledge with every update. ### AI Infrastructure Debate - ES is central to the current AI infrastructure debate - ES enables on-device learning/training on the edge - Current methods require massive server farms for updates ### The Catastrophic Forgetting Problem - ES leads to catastrophic forgetting in LLMs - ES performance degrades significantly on old tasks after new training - Gradient methods show much less degradation ### Experimental Results - ES models performed poorly on the math task 'Countdown' after training on image masking - ES showed 1000x greater forgetting magnitude than GRPO - ES scores were flat on the original task while GRPO scores remained high ### Mechanism of Failure - ES updates all parameters simultaneously via noise perturbation - This globally corrupts prior knowledge - Gradient methods allow for more targeted updates ### Future Research Direction - The challenge is creating an ES variant that can compartmentalize learning - Mimic biological evolution's ability to retain old skills while acquiring new ones - This is the crucial step for viable on-device AI ![Screenshot at 00:00: Introductory slide featuring two people podcasting with the text 'BECOME A MEMBER TODAY!' displayed prominently.](https://ss.rapidrecap.app/screens/c6yJEBqewvU/00-00-00.jpg) ![Screenshot at 00:28: A visual representation of the performance disparity, showing the ES model's performance curve dropping significantly compared to the GRPO model's curve on the original task.](https://ss.rapidrecap.app/screens/c6yJEBqewvU/00-00-28.jpg) ![Screenshot at 01:44: A graph illustrating the core concept: the ES model's performance curve \(lower\) degrades severely on the initial task after learning a new task, unlike the gradient-based method \(higher\).](https://ss.rapidrecap.app/screens/c6yJEBqewvU/00-01-44.jpg) ![Screenshot at 03:34: The graph highlights the math game 'Countdown' score \(vertical axis\) versus the prior knowledge score \(horizontal axis\), showing the ES model clustering badly in the bottom right \(poor math, poor prior knowledge\).](https://ss.rapidrecap.app/screens/c6yJEBqewvU/00-03-34.jpg) ![Screenshot at 05:21: A visual explanation of the evaluation: the vertical axis represents the new task score \(Countdown\), and the horizontal axis represents the old task score \(General Reasoning\) for both ES and GRPO models.](https://ss.rapidrecap.app/screens/c6yJEBqewvU/00-05-21.jpg)