AI Paper Atlas

Research map / Large language models / Agents, code and math

More Sail than Ballast: Addressing Harmful Knowledge Leakage in the Expansive Reasoning Space of LRMs

Qibing Ren, Xinhao Song, Ke Fan, Lijun Li, Zhanpeng Zhou, Gongshen Liu, Junchi Yan, Lizhuang Ma, Jing Shao

ICML 2026 · Large language models › Agents, code and math · approach: Mathematical and structural reasoning benchmarks

Read the paperCheck your idea against 137,217 papers

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
add faithfulness and helpfulness rewards to safety-only reward (ablation)OpenSafeRL on LLaMA-3-8B · XSTestover-refusal rate (%): 50.80 -> 2.80Improved
add faithfulness reward to safety + helpfulness (ablation)OpenSafeRL on LLaMA-3-8B · AutoDANattack success rate (%): 37.0 -> 14.0Improved
add helpfulness reward to safety-only reward (ablation)OpenSafeRL on LLaMA-3-8B · AutoDANattack success rate (%): 2.0 -> 37.0Hurt
train on synthesized open-ended data instead of plain data (ablation)OpenSafeRL on LLaMA-3-8B · AutoDANattack success rate (%): 36.0 -> 14.0Improved
add safety principle system prompt to RL training (ablation)OpenSafeRL on LLaMA-3-8B · XSTestover-refusal rate (%): 29.20 -> 2.80Improved
add safety principle prompt to vanilla model without RL (ablation)LLaMA-3-8B-Instruct · AutoDANattack success rate (%): 45.5 -> 22.0Improved
insert wait token at test time (ablation)LLaMA-3-8B-Instruct vanilla · XSTestover-refusal rate (%): 6.80 -> 11.60Hurt
insert wait token at test time (ablation)OpenSafeRL on LLaMA-3-8B · AutoDANattack success rate (%): 14.0 -> 2.0Improved
raise safety weight alpha from 0.5 to 0.7 (ablation)OpenSafeRL on LLaMA-3-8B · AutoDAN / GSM8KASR and accuracy: AutoDAN 14.0 -> 12.0; GSM8K 74.08 -> 73.00Mixed
use OpenSafeRL instead of vanillaQwen-2.5-7B-Instruct · OpenLeakBenchharmfulness rate (%): 62% -> 23.5%Improved
use OpenSafeRL instead of Circuit BreakersQwen-2.5-7B-Instruct · over-refusal averageover-refusal rate (%): 31.07 -> 2.02Improved
apply gradient ascent alignmentQwen-2.5-7B-Instruct · over-refusal averageover-refusal rate (%): 7.55 -> 70.35Hurt

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