Weikang Li
Publications
Beyond Alignment: Expanding Reasoning Capacity via Manifold-Reshaping Policy Optimization
Reinforcement Learning with Verifiable Rewards (RLVR) has demonstrated remarkable success in enhancing the reasoning capabilities of Large Language Models (LLMs). However, recent studies question whether RL genuinely expands reasoning capacity or merely aligns existing latent capabilities, arguing that exploration remains confined within the pre-trained model's low-rank bias manifold. In this work, we challenge this accessibility boundary hypothesis by demonstrating that the latent reasoning space can be fundamentally expanded through targeted geometric interventions. We propose Manifold-Reshaping Policy Optimization (MRPO), a geometric framework designed to fundamentally restructure the inference space of LLMs. MRPO operates in two stages: first, we employ Spectral Orthogonal Exploration (SOE) to eject the policy initialization into the null space of the bias manifold; second, we integrate an Effective Rank regularization term into the policy optimization objective. This approach incentivizes the discovery and maintenance of high-dimensional reasoning trajectories against the entropy-reducing tendency of standard RL. Empirically, our 4B-parameter method achieves state-of-the-art performance on mathematical tasks, significantly outperforming larger models (e.g., Qwen3-32B) and expanding the capability boundary beyond standard GRPO. Our code is available at https://anonymous.4open.science/r/MRPO-D57B/
Stay in Character, Stay Safe: Dual-Cycle Adversarial Self-Evolution for Safety Role-Playing Agents
LLM-based role-playing has rapidly improved in fidelity, yet stronger adherence to persona constraints commonly increases vulnerability to jailbreak attacks, especially for risky or negative personas. Most prior work mitigates this issue with training-time solutions (e.g., data curation or alignment-oriented regularization). However, these approaches are costly to maintain as personas and attack strategies evolve, can degrade in-character behavior, and are typically infeasible for frontier closed-weight LLMs. We propose a training-free Dual-Cycle Adversarial Self-Evolution framework with two coupled cycles. A Persona-Targeted Attacker Cycle synthesizes progressively stronger jailbreak prompts, while a Role-Playing Defender Cycle distills observed failures into a hierarchical knowledge base of (i) global safety rules, (ii) persona-grounded constraints, and (iii) safe in-character exemplars. At inference time, the Defender retrieves and composes structured knowledge from this hierarchy to guide generation, producing responses that remain faithful to the target persona while satisfying safety constraints. Extensive experiments across multiple proprietary LLMs show consistent gains over strong baselines on both role fidelity and jailbreak resistance, and robust generalization to unseen personas and attack prompts.
Student Guides Teacher: Weak-to-Strong Inference via Spectral Orthogonal Exploration
While Large Language Models (LLMs) demonstrate near-human capabilities, they often suffer from "Reasoning Collapse" in complex mathematical proving and long-horizon planning. Models tend to degenerate into low-rank Bias Manifold, where stochastic sampling merely produces lexical variations of erroneous logic rather than semantic exploration. This geometric collapse renders the model "blind" to high-value solutions that lie within its Null Space. To address this, we propose Spectral Orthogonal Exploration (SOE), a geometric framework operating on a counter-intuitive "Student Guides Teacher" paradigm. Specifically, we utilize a weak auxiliary agent not for imitation, but as an orthogonal probe. By explicitly navigating the Teacher's Null Space, SOE serves as a geometric bridge, effectively ejecting the model from local optima to explore diverse, high-value solution spaces. Experiments on mathematical benchmarks demonstrate that, relative to baseline methods, our approach improves average accuracy by 62.4% and increases average sampling efficiency by 113.7%, indicating a promising path toward overcoming performance plateaus in advanced reasoning tasks.