Skip to content
View original post on X: BAAIOfficial
Published
22/100

BAAI's CURE adds 0.7% compute overhead to diffusion LLM safety

AIBAAI says its CURE test-time alignment method adds 0.84 TFLOPs, or 0.7%, to LLaDA-8B-Instruct's 126.29 TFLOPs of base generation. Prior test-time defenses SRIGuard and DiffuGuard add 200.0% and 103.1%, which the post puts at 150x to 300x more extra compute.
BAAIVerified on XOfficial
@BAAIBeijing
View on X

Part of a thread · earlier post

The best part? Test-time alignment doesn't need to double your compute budget:

⚡ 0.7% Extra Compute Overhead:

On LLaDA-8B-Instruct (base generation: 126.29 TFLOPs), prior test-time defenses like SRIGuard and DiffuGuard add +252.58 TFLOPs (+200.0%) and +130.24 TFLOPs (+103.1%)—whereas CURE adds just 0.84 TFLOPs (+0.7%), cutting extra compute by 150x–300x!

🌐 Towards Proactive Generation Control:

Beyond token-level safety, BAAI is extending intermediate-state editability toward semantic-level reasoning repair and multi-constraint test-time control (factuality & logical consistency)—advancing AI safety from passive output filtering to proactive, trustworthy generation control.

📄NeurIPS：https://neurips.cc/virtual/2026/loc/sydney/poster/151876

Source details

BAAI · x.comPublished · added here

Original title: The best part? Test-time alignment doesn't need to double your compute budget:

Topics

Safety & alignmentChina AIPapers & researchEvals & benchmarks