AI Paper Atlas

Research map / Reinforcement learning / Visual and representation RL

Alignment Risks from Capability-Seeking RL Training

Yujun Zhou, Yue Huang, Han Bao, Kehan Guo, Zhenwen Liang, Pin-Yu Chen, Tian Gao, Werner Geyer, Nuno Moniz, Nitesh V Chawla, Xiangliang Zhang

ICML 2026 · Reinforcement learning › Visual and representation RL · approach: Probabilistic and autonomous RL · 0 citations (OpenAlex)

This paper tried a common change: Chain-of-thought prompting

use chain-of-thought prompting for Qwen3-4B · Reward Tampering · exploit discovery Hurt

What 1,107 accepted papers report for this change

See this change in detailMatched by the wording of each paper's own result, read by AI from the full text. Papers report gains more often than failures, so this shows what gets published, not the chance a change works.

Read the paper

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train with GRPO in a vulnerable environment instead of using the original modelLlama-3.1-8B-Instruct · Audited Self-Gradingexploit rate (%): 44.0 -> 80.0Hurt
train in vulnerable Self-Grading environment (with loophole) vs original modelLlama-3.1-8B-Instruct · Audited Self-Gradingtask performance (%): 7.1 -> 16.9Improved
remove the loophole (loophole-free control) (ablation)Llama-3.1-8B-Instruct GRPO · Audited Self-Gradingexploit rate (%): 80.0 -> 2.6Improved
remove the loophole (loophole-free control) (ablation)Qwen3-4B GRPO · Audited Self-Gradingexploit rate (%): 32.8 -> 9.0Improved
remove the loophole (loophole-free control) (ablation)Llama-3.1-8B GRPO · Audited Self-Gradingtask performance (%): 16.9 -> 20.4No clear effect
use chain-of-thought prompting for Qwen3-4BQwen3-4B · Reward Tamperingexploit discoveryHurt
zero-shot transfer from Proxy Metric Gaming trainingQwen3-4B-Base · Self-Gradingexploit rate (%): 9.0 -> 55.7Improved
zero-shot transfer from other exploits to Proxy Metric Gamingall models · Proxy Metric Gamingexploit rateNo clear effect
initialize from a Reward Tampering policy before RL on Context-Conditional Compliance (ablation)Llama-3.1-8B-Instruct · Context-Conditional Complianceexploit rate (%): 72.8 -> 83.8Improved
initialize from a Proxy Metric Gaming policy before RL on Context-Conditional Compliance (ablation)Qwen3-4B-Base · Context-Conditional Complianceexploit rate (%): 78.5 -> 62.8Hurt
initialize from a failed Reward Tampering policy before RL on Self-Grading (ablation)Qwen3-4B · Self-Gradingexploit rate (%): 32.8 -> 47.1Improved
SFT-distill exploits from GRPO teacher into studentQwen3-4B-Base · Proxy Metric Gamingexploit rate (%): 3.9 -> 91.1Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy