| May 2026 | Our new paper PW-OPSD is now on arXiv! We find that teacher-token reliability in on-policy self-distillation for reasoning is position-structured, and propose Position-Weighted On-Policy Self-Distillation (PW-OPSD) to up-weight reliable later tokens at no extra teacher cost. Check out our paper and code. |
| May 2026 | ROM is on arXiv. We frame overthinking in large reasoning models as a latent productive-to-redundant transition that surfaces in hidden states around first-correct-solution (FCS) boundaries, and propose ROM, a model-agnostic streaming framework that monitors a frozen LRM with a lightweight hidden-state detector and intervenes at well-formed reasoning boundaries. Our Counterfactual Self-Correction (CSC) augmentation preserves useful pre-FCS self-correction while labeling only post-FCS continuation as redundant. On Qwen3-8B and DeepSeek-R1-Distill-Qwen-32B across MATH500, GSM8K, AIME25, and MMLU-Pro, ROM improves the accuracy–length tradeoff (e.g., Qwen3-8B: 4262 → 3107 tokens with slightly higher accuracy), stacks with L1 for another ~21% token reduction at zero accuracy loss, and cuts wall-clock latency by 46.5%. Check out our project page, code, and dataset. |
| Apr 2026 | Our paper ReasoningBomb has been accepted to ACM CCS 2026! We propose an RL-based framework that crafts short, natural-language prompts to trap LRMs into pathologically long reasoning, with a constant-time surrogate reward enabling 4.39×10⁵× training speedup. Just 10% malicious traffic cuts benign throughput by 49.8% and monopolizes 64.3% of compute. Check out our paper, website, code, and dataset. |
| Feb 2026 | ReasoningBomb is now on arXiv. Check out our website, code, and dataset. |
| Sep 2024 | Passed my Qualifying Exam. |