“Targeted Manipulation and Deception Emerge when
Optimizing LLMs for User Feedback” by Marcus Williams, micahcarroll, Adhyyan Narang, Constantin Weisser, Brendan Murphy
LessWrong (30+ Karma)
Shownotes Transcript
Produced as part of MATS 6.0 and 6.1.
** Key takeaways:**
Training LLMs on (simulated) user feedback can lead to the emergence of manipulative and deceptive behaviors. These harmful behaviors can be targeted specifically at users who are more susceptible to manipulation, while the model behaves normally with other users. This makes such behaviors challenging to detect. Standard model evaluations for sycophancy and toxicity may not be sufficient to identify these emergent harmful behaviors. Attempts to mitigate these issues, such as using another LLM to veto problematic trajectories during training don't always work and can sometimes backfire by encouraging more subtle forms of manipulation that are harder to detect. The models seem to internalize the bad behavior, acting as if they are always responding in the best interest of the users, even in hidden scratchpads.
You can find the full paper here. The code for the project can be [...]
Outline:
(00:16) Key takeaways:
(01:31) Why would anyone optimize for user feedback?
(02:25) Simulating user feedback:
(04:52) Harmful behaviors reliably emerge when training with exploitable user feedback
(06:34) Even if most users give good feedback, LLMs will learn to target exploitable users
(08:50) Can this be solved by mixing in harmlessness data?
(09:50) Surely we would detect such harmful behaviors with evals?
(11:17) Can we use another LLM to veto harmful training data?
(13:26) Why does training with the veto not eliminate bad behavior?
(15:47) Chain of Thought Reveals RL-Induced Motivated Reasoning
(18:12) Discussion
(18:15) Model personalization and backdoors as two sides of the same coin.
(19:05) Are our simulated users realistic?
(20:54) What do our results mean for annotator and AI feedback gaming more broadly?
(21:56) Would we expect our method to be a good proxy for future user feedback optimization techniques?
(22:27) Acknowledgements
The original text contained 2 footnotes which were omitted from this narration.
The original text contained 8 images which were described by AI.
First published: November 7th, 2024
---
Narrated by TYPE III AUDIO).
Images from the article:
)
)
)
)
)
)
)
)
)
)
)
)
Apple Podcasts and Spotify do not show images in the episode description. Try Pocket Casts), or another podcast app.