Search for: All records

Large Language Models (LLMs) demonstrate impressive capabilities across many applications but remain vulnerable to jailbreak attacks, which elicit harmful or unintended content. While model fine-tuning is an option for safety alignment, it is costly and prone to catastrophic forgetting. Prompt optimization has emerged as a promising alternative, yet existing prompt-based defenses typically rely on static modifications (e.g., fixed prefixes or suffixes) that cannot adapt to diverse and evolving attacks.We propose Dynamic Deep Prompt Optimization (DDPO), the first jailbreak defense based on deep prompt optimization. DDPO uses the target LLM's own intermediate layers as feature extractors to dynamically generate defensive embeddings via a lightweight multilayer perceptron. These tailored embeddings are then injected into a subsequent intermediate layer, enabling an input-dependent defense without modifying the LLM's weights. This design ensures high adaptability with minimal computational overhead.Experiments on a diverse set of models and attacks demonstrate that DDPO significantly outperforms static prompt optimization methods, particularly on weakly aligned models and when handling semantically ambiguous benign prompts, successfully distinguishing them from genuinely harmful requests. 

Recent advances in large language models have significantly improved the capabilities of programming. While these models excel at generating valid software, applying them to the hardware domain remains challenging due to the intrinsic complexity and strict structural semantics required in hardware design. Current LLM approaches for hardware generation typically focus on direct generation. This often results in hardware implementations with functional errors or structural flaws. To overcome these limitations, we propose a reasoning-enhanced training framework explicitly tailored for hardware generation tasks. Our multi-stage methodology combines systematic dataset curation via compilation filtering (achieving a 100% pass rate compared to 27 − 44% in existing datasets), Function-Aligned Differentiated Revision for comparative annotation across five RTL-relevant dimensions, supervised fine-tuning using reasoning prompts, and reinforcement learning guided by Verilator Parser. Our experiments show that explicitly incorporating reasoning substantially enhances the structural integrity and functional correctness of generated hardware designs, improving pass@1 rates by up to 20% on VerilogEval Human benchmarks and reproducing the "Aha moment", where the model explicitly organizes ideas before generation. Our work demonstrates that smaller, specialized reasoning models (1.5B parameters) can effectively augment larger open-source language models through reasoning transfer.