SYNTOLOGY HomeExplorerAtlasCodeMethodologyAboutDevelopersFeedPricing
Paper · 2406.14393 · ICLR · 2025

Jailbreaking as a Reward Misspecification Problem

Lei Li, Zhenguo Li, Jiahui Gao, Lingpeng Kong, Qi Liu, Zhihui Xie

arXiv · PDF · Open in the Atlas

Code that ran

We lifted 15 functions out of this paper's own repositories and ran 7 of them in a sandbox. "Ran" means the function executed on a synthesized input and returned a value. It is not a reproduction of the paper's results.

RepositoryRoleRan
zhxieml/remiss-jailbreak canonical 7 of 15
FunctionStatusWhere it lives
EmptySeq Ran zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("f7e6ed8ae82ad71f")
Seq Ran zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("b33c78601d890827")
add_dummy_dim_to_slice Ran zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("421d4ba5e59039da")
apply_repetition_penalty Ran zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("298878ca905883f4")
expand_for_broadcast_list Ran zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("a70d14fb2791dd29")
expand_for_broadcast_tensor Ran zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("249485c3fcfb764e")
select_next_token_candidates Ran zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("b5c4e0c154efef64")
MergedSeq Not yet run zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("b50e2d1935c06431")
calculate_reward Not yet run zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("bf9d280eb54736b7")
evaluate_next_token_candidates Not yet run zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("e1f298fad7ae62ed")
get_next_token_probabilities Not yet run zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("02fbdc293c7c022f")
reMissOpt Not yet run zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("9f311245e1b7a581")
select_and_evaluate_next_token_candidates Not yet run zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("4d986a6e6f39f5ce")
select_next_beams Not yet run zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("09209b225f2ff6e0")
stack_seqs Not yet run zhxieml/remiss-jailbreak/src/remissopt.py
pointer only (licence: NONE) · get_code("773e7f99aa71d7e3")

Repositories linked to this paper

Some links come from the archived Papers with Code dataset (CC BY-SA 4.0): attribution and licence.

Abstract

The widespread adoption of large language models (LLMs) has raised concerns about their safety and reliability, particularly regarding their vulnerability to adversarial attacks. In this paper, we propose a new perspective that attributes this vulnerability to reward misspecification during the alignment process. This misspecification occurs when the reward function fails to accurately capture the intended behavior, leading to misaligned model outputs. We introduce a metric ReGap to quantify the extent of reward misspecification and demonstrate its effectiveness and robustness in detecting harmful backdoor prompts. Building upon these insights, we present ReMiss, a system for automated red teaming that generates adversarial prompts in a reward-misspecified space. ReMiss achieves state-of-the-art attack success rates on the AdvBench benchmark against various target aligned LLMs while preserving the human readability of the generated prompts. Furthermore, these attacks on open-source models demonstrate high transferability to closed-source models like GPT-4o and out-of-distribution tasks from HarmBench. Detailed analysis highlights the unique advantages of the proposed reward misspecification objective compared to previous methods, offering new insights for improving LLM safety and robustness. Code is available at: https://github.com/zhxieml/remiss-jailbreak.

For agents

The same record, over MCP at https://syntology.ai/mcp:

get_harvested_code_for_paper("2406.14393")
get_code_for_paper("2406.14393")
have("2406.14393")

Connect an agent — have() is free.