AnthropicOpenAIGoogle DeepMind

Based on real interviews

Investigate resistance to correction

Find out why an explicit correction barely changes a model's tendency to agree with a false claim.

Role
Research Engineer (RLHF / Alignment)
Difficulty
Advanced
Practice time
60 min
Unlock with Full Access

The challenge

The alignment team expected the model to respond strongly when a user challenges its answer. Instead, sycophancy changes only slightly. Use the available model internals and training artifacts to propose a mechanism and a three-month investigation. Specify predictions that experiments could disprove rather than assuming the training history already explains the behaviour.

What you'll practise

  • Training-to-behaviour reasoning

    Connect a proposed mechanism to evidence from the model and its training artifacts.

  • Falsifiable predictions

    Describe outcomes that would force you to revise the explanation.

  • Experimental judgement

    Choose tools and interventions that answer the specific research question.

Your writing workspacePreview
Your sycophancy investigationYour response
Your recommendation

Write your reasoning and submit it for feedback.

The full brief and feedback are included with Full Access.
Explore challenges
Full Access · Deep Dive

Anticipate reward hacking in an RLHF run

Identify how the next RLHF run could improve reward while making the model less useful.

Full Access · Deep Dive

Investigate asymmetric factual recall

Explain why reversing the direction of a factual question can sharply reduce a model's accuracy.

Strategy memo: rolling out our next flagship model

Free with an account · Deep Dive