Non-autoregressive models advance document-level machine translation with impressive speedup. Delve into the opportunities and challenges of NAT on extended sequences: aclanthology.org/2023.findings-… #EMNLP2023

GEMINI: Controlling The Sentence-Level Summary Style in Abstractive Text Summarization #EMNLP2023 aclanthology.org/2023.emnlp-mai…

Thanks for sharing our work. Our primary contribution is the establishment of a framework that enhances the reliability of LLMs as it: 1) generalizes out-of-distribution data, 2) elucidates how LLMs benefit from discriminative models, and 3) minimizes hallucinations.

Excited to announce that Fast-DetectGPT made it to #ICLR2024 🎉 WestlakeNLP In the rebuttal phase, got an unfair 1 amidst 8866. Huge shoutout to Eric for the public support. Appreciate your sense of justice! ⚖️🔍 Paper: openreview.net/forum?id=Bpcgc…

I am seeking a research intern to collaborate on the development and application of large language models within real-world industry scenarios at ByteDance, Singapore. If you are interested, please drop me an email.

GWLS: a general framework to learn from ANY weak supervision! It outperforms existing methods on *11* weak supervision settings, e.g., partial label, multiple instance learning, label prop., multiclass multi-label... 📸Paper: arxiv.org/abs/2402.01922

[1/5] 🧵 Thrilled to unveil our latest research on "Causal Analysis of CoT in LLMs"! We delve into the intricate dynamics between Chain of Thought reasoning and answer generation in LLMs, revealing some unexpected insights. 🤖💭 📄Read the full paper: arxiv.org/abs/2402.16048

[2/5] Despite the potential of CoT to enhance task performance in LLMs, our findings show a surprising number of instances where correct answers follow incorrect CoTs and vice versa. This discrepancy raises fundamental questions about LLMs' reasoning capabilities.

[3/5] Employing causal analysis, we dissect the cause-effect relationships between CoTs/instructions and answers in LLMs. Our analysis exposes the Structural Causal Model (SCM) LLMs mimic, highlighting significant differences from human reasoning processes.

[4/5] We further explore how ICL, SFT and RLHF significantly influence the causal structures in LLMs. Our investigation sheds light on how these techniques impact the reasoning process, offering critical insights for future advancements.

[5/5] Our research contributes to the broader discourse on the role of CoT in LLM reasoning, offering new angles on the extent to which LLMs replicate human-like reasoning steps. Please refer to our paper for detailed results!

I just published “No Training Needed, Fast-DetectGPT Boosts Text Detection Speed by 340 Times” link.medium.com/sozGRpDd9Hb

I'm attending #ICLR2024 in Vienna from May 7-11. Our posters are at Halle B #256 and #116 on May 8. Looking forward to meeting old and new friends!🥰

Exciting news! 🎉 Our online demo for Fast-DetectGPT is now live! 🚀 Experience lightning-fast text detection in action. Give it a try here: [region-9.autodl.pro:21504] Let us know what you think! #FastDetectGPT #AI #TextDetection.

⛄️Excited to share our work on causal analysis of LLMs at COLING 2025!💖Hongbo Linyi Yang Cunxiang Wang "How Likely Do LLMs with CoT Mimic Human Reasoning?" Paper: arxiv.org/pdf/2402.16048

Welcome to try our system. The feedback system will not replace any human reviewers. The agent will not write reviews or make automated edits to reviews. Rather, it will serve as an assistant, providing optional feedback that reviewers can incorporate or disregard. #LLM #ICLR #AI

LLMs often rely on correlations, not causation. ❤️‍🔥 Our causal analyses show that RLVR-trained LRMs move closer to true causal reasoning — but distilled LRMs and LLMs do not⁉️ 🧠 Paper: "Correlation or Causation?" 📘 [arxiv.org/pdf/2509.17380](arxiv.org/pdf/2509.17380)

⛄️ Excited to share our EMNLP 2025 paper: Direct Value Optimization (DVO) 🌲 💡 Instead of pairwise DPO-style tuning, DVO learns directly from value signals in MCTS search data, enabling efficient RL training for reasoning LLMs ⚡️ 📘 arxiv.org/pdf/2502.13723