Mixture-of-Agents Aggregation Quiz

Answers and explanations appear as you take the quiz.

1. In Mixture-of-Agents, what is the role of the aggregator model?
   • To synthesize the proposer models' responses into a single improved answer
   • To delete all but one draft at random
   • To translate the prompt into code
   • To train the proposer models
2. What key property of LLMs does the MoA paper rely on?
   • That LLMs cannot read other models' outputs
   • That LLMs produce better answers when shown peers' responses ('collaborativeness')
   • That LLMs always agree
   • That larger models are unnecessary
3. How does MoA differ from simple majority voting?
   • It picks the most common answer verbatim
   • It uses only one model
   • The aggregator reads all drafts as context and generates a new synthesis
   • It ignores the proposers entirely
4. Why is diversity among proposer models valuable in MoA?
   • It guarantees identical outputs
   • It reduces the number of layers to zero
   • It makes all answers wrong
   • Different model families contribute complementary strengths the aggregator can combine
5. What is the main practical cost of the MoA approach?
   • Higher latency and compute, since each layer multiplies model calls
   • It requires no inference at all
   • It can only output a single word
   • It deletes the original prompt
6. What did the MoA paper report about open-source models on AlpacaEval 2.0?
   • They could not run the benchmark
   • A MoA of open models surpassed GPT-4 Omni's score
   • Open models always scored zero
   • Only closed models were tested
7. What information is given to an agent in a later Mixture-of-Agents layer?
   • Only the original prompt with all earlier drafts removed
   • Only a majority-vote label
   • The outputs produced by agents in the preceding layer
   • A gradient update from each previous model
8. What is the purpose of placing several proposer models in the first MoA layer?
   • Train a shared base model from scratch
   • Choose one response by majority vote and discard the rest
   • Compress all outputs into a single token
   • Create a set of candidate responses for later agents and the aggregator to use