Reflexion and Self-Correcting Agents Quiz

Answers and explanations appear as you take the quiz.

1. What is the defining characteristic of how a Reflexion agent 'learns'?
   • It stores written reflections in context without changing model weights
   • It updates the model's weights through fine-tuning
   • It downloads a new model version after each failure
   • It asks a human to relabel its training data
2. In the Reflexion framework, what does the Evaluator do?
   • It generates the agent's next action
   • It scores the outcome of an attempt to produce a feedback signal
   • It deletes old memories to save space
   • It translates the task into another language
3. Why does the quality of the evaluation signal matter so much in Reflexion?
   • Because it determines the model's learning rate
   • Because it decides how many GPUs are used
   • Because a wrong signal can make the agent reflect on and reinforce incorrect lessons
   • Because it sets the temperature of the language model
4. On which coding benchmark did the Reflexion paper report a 91% pass@1 result?
   • ImageNet
   • COCO detection
   • WMT translation
   • HumanEval
5. Which is a genuine open risk with self-correcting agents?
   • They can loop endlessly 'fixing' output that was already correct
   • They require labeled data for every token
   • They cannot run on standard hardware
   • They permanently corrupt the base model
6. Where does Reflexion keep feedback for a later attempt?
   • In updated model weights
   • In an episodic memory buffer supplied as text context
   • In a hardware cache shared across users
   • In the original training corpus
7. What is the role of the Actor in Reflexion?
   • It changes the base model weights
   • It authenticates citations automatically
   • It generates the actions or attempt for the task
   • It sets the benchmark score after the run
8. What kinds of feedback can Reflexion use?
   • Only gradient updates from fine-tuning
   • Only human ratings in every task
   • Only token likelihood from pretraining
   • Signals such as scalar outcomes or language feedback from external or simulated sources