Yifan Le

Yifan Le

Preference optimization has become an important post-training paradigm for improving the reasoning abilities of large language models. Existing methods typically rely on externally constructed preference data, using preferred and dispreferred responses as sample-level supervision. However, such external signals rarely make explicit use of capability-related information contained in the model's internal representations. For mathematical reasoning, certain neuron groups may exhibit activation patterns associated with mathematical knowledge, symbolic manipulation, or logical reasoning. Similar to reflexive behavioral signals, these internal activations may provide a coarse indication of whether the model is engaging math-related capabilities.We introduce YFPO, short for Yoked Feature Preference Optimization, a preliminary neuron-guided preference optimization framework for mathematical reasoning. YFPO first uses AttnLRP to identify math-related neurons, and then constructs an auxiliary reward from their activation margin between preferred and dispreferred responses. This design augments external preference learning with internal neuron-level signals. We conduct preliminary experiments on a small-scale language model using GSM8K as the main benchmark. Results suggest that neuron-level signals can interact with preference optimization and occasionally improve reasoning performance, offering a promising direction for more fine-grained and interpretable reasoning-oriented post-training.

Weikang Zhou, Xiao Wang, Limao Xiong et al.

Jailbreak attacks are crucial for identifying and mitigating the security vulnerabilities of Large Language Models (LLMs). They are designed to bypass safeguards and elicit prohibited outputs. However, due to significant differences among various jailbreak methods, there is no standard implementation framework available for the community, which limits comprehensive security evaluations. This paper introduces EasyJailbreak, a unified framework simplifying the construction and evaluation of jailbreak attacks against LLMs. It builds jailbreak attacks using four components: Selector, Mutator, Constraint, and Evaluator. This modular framework enables researchers to easily construct attacks from combinations of novel and existing components. So far, EasyJailbreak supports 11 distinct jailbreak methods and facilitates the security validation of a broad spectrum of LLMs. Our validation across 10 distinct LLMs reveals a significant vulnerability, with an average breach probability of 60% under various jailbreaking attacks. Notably, even advanced models like GPT-3.5-Turbo and GPT-4 exhibit average Attack Success Rates (ASR) of 57% and 33%, respectively. We have released a wealth of resources for researchers, including a web platform, PyPI published package, screencast video, and experimental outputs.

Yifan Le

Constrained decoding has been widely adopted for structured generation with large language models (LLMs), ensuring that outputs satisfy predefined formats such as JSON and XML. However, existing approaches largely treat schemas as purely structural constraints and overlook the possibility that their linguistic formulation may affect model behavior. In this work, we study how instruction placement influences model performance in structured generation and show that merely changing the wording of schema keys, without modifying the prompt or model parameters, can significantly alter model performance under constrained decoding. Based on this observation, we propose to reinterpret structured generation as a multi-channel instruction problem, where instructions can be conveyed explicitly through prompts and implicitly through schema keys during decoding. To the best of our knowledge, this is the first work to systematically study how schema key formulation acts as an implicit instruction channel and affects model performance under constrained decoding. Experiments on multiple mathematical reasoning benchmarks show that different model families exhibit distinct sensitivities to these instruction channels: Qwen models consistently benefit from schema-level instructions, while LLaMA models rely more heavily on prompt-level guidance. We further observe non-additive interaction effects between instruction channels, showing that combining multiple channels does not always lead to further improvement. These findings suggest that schema design not only determines output structure, but also carries instruction signals, offering a new perspective on structured generation in LLMs.

Yifan Le, Yunliang Li

Multilingual large language models (LLMs) achieve strong performance across languages, yet how language capabilities are organized at the neuron level remains poorly understood. Prior work has identified language-related neurons mainly through activation-based heuristics, which conflate language preference with functional importance. Prior work has identified language-related neurons mainly through activation-based heuristics, which conflate language preference with functional importance. We propose CRANE, a relevance-based analysis framework that redefines language specificity in terms of functional necessity, identifying language-specific neurons through targeted neuron-level interventions. CRANE characterizes neuron specialization by their contribution to language-conditioned predictions rather than activation magnitude. Our implementation will be made publicly available. Neuron-level interventions reveal a consistent asymmetric pattern: masking neurons relevant to a target language selectively degrades performance on that language while preserving performance on other languages to a substantial extent, indicating language-selective but non-exclusive neuron specializations. Experiments on English, Chinese, and Vietnamese across multiple benchmarks, together with a dedicated relevance-based metric and base-to-chat model transfer analysis, show that CRANE isolates language-specific components more precisely than activation-based methods.