AnthropicVerified interview challenge
LLM output processing bug hunt
Review the code that turns batched token sequences into the response text returned to each request.
Unlock with Full AccessThe challenge
Each sequence contains a prompt, generated tokens and possible padding. The processor must find the response boundaries, decode the right text and preserve the correspondence with the original inputs. Inspect the implementation for wrong slices, mixed-up requests and inputs that disappear silently.
What you'll practise
Sequence boundaries
Separate generated text from the prompt, termination tokens and padding.
Request identity
Keep each decoded response attached to the correct input and position.
Batch consistency
Check how the implementation handles mismatched lengths and incomplete output.
Batch response processorCode included in the full challenge
Your code review
Explain the incorrect output and how to prevent it.
The code, instructions and feedback are included with Full Access.