Model Swarms and swarm intelligence

Model Swarms is a collaborative search algorithm proposed by University of Washington , Google Cloud AI Research, Google DeepMind, Google.

Inspired by Particle Swarm Optimization, it uses swarm intelligence to help multiple LLMs work together to adapt without fine-tuning and improve.

Here’s how it works:

Model Swarms begins with a set of LLM experts already fine-tuned on different tasks. Inspired by swarm intelligence (like one that flocks of birds have), each LLM acts as a "particle" with a location (its settings or weights) and a direction (velocity) for improvement.

• Step 0 – Initialization:

To make the search more effective, Model Swarms expands the initial set of particles by combining pairs of experts. Each particle starts with a random direction.

• Step 1 – Velocity Update: Each particle adjusts its direction based on four factors:

? Inertia: Keeps it moving in its current direction.

? Personal best: Draws it toward the best performance it has achieved so far.

? Global best: Guides it toward the best result of the group.

? Global worst: Keeps it away from the least effective settings.

• Step 2 – Weight Update:

The particle takes a step in its new direction and is evaluated again. If a particle fails to improve over a set number of tries, it restarts from its best-known settings to avoid getting stuck.

• End of Search:

The search stops when the global best result hasn’t improved after several attempts, or after a set number of iterations. The particle with the highest score is selected as the best solution.

Results of Model Swarms:

? Single task: improved by up to 21.0%, beating 12 models by 13.3% on average across 9 datasets.

? Multi-task: 5.7% better than baselines, with 11.3% gains in legal domain

? Reward model: Outperformed baselines by 6.7% on average

? Human interest: 70.8% win rate in evaluations; improved LLM-as-a-judge scores by 17.6% and factuality by 17.0% in 16 topics

Original paper: MODEL SWARMS: COLLABORATIVE SEARCH TO ADAPT LLM EXPERTS VIA SWARM INTELLIGENCE