Abstract

We show that when large language models learn to reward hack on production RL environments, this can result in egregious emergent misalignment. We start with a pretrained model, impart knowledge of reward hacking strategies via synthetic document finetuning or prompting, and train on a selection of real Anthropic production coding environments. Unsurprisingly, the model learns to reward hack. Surprisingly, the model generalizes to alignment faking, cooperation with malicious actors, reasoning about malicious goals, and attempting sabotage when used with Claude Code, including in the codebase for this paper. Applying RLHF safety training using standard chat-like prompts results in aligned behavior on chat-like evaluations, but misalignment persists on agentic tasks. Three mitigations are effective: (i) preventing the model from reward hacking; (ii) increasing the diversity of RLHF safety training; and (iii) "inoculation prompting", wherein framing reward hacking as acceptable behavior during training removes misaligned generalization even when reward hacking is learned.

Twitter thread

New Anthropic research: Natural emergent misalignment from reward hacking in production RL.

“Reward hacking” is where models learn to cheat on tasks they’re given during training.

Our new study finds that the consequences of reward hacking, if unmitigated, can be very serious.

In our experiment, we [...]

---

Outline:

(00:14) Abstract

(01:26) Twitter thread

(05:23) Blog post

(07:13) From shortcuts to sabotage

(12:20) Why does reward hacking lead to worse behaviors?

(13:21) Mitigations

---

First published:

Source:

---

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.