Ningyu Xu @ ICML 2026

Our paper is now out in PNAS!💡 Are LLMs developing human-like concepts that are central to human cognition? If so, how are such concepts represented, organized, and related to behavior? doi.org/10.1073/pnas.2… 1/N

> replicate J-space on GLM 5.2 > train a reward model and run RL to reduce hallucinations > show me how this model makes cancer predictions Using our platform Silico is like having a team of AI researchers ready to run experiments like these. Private beta is open now. 🧵 (1/6)

Is your LM secretly an SAE? Most circuit-finding interpretability methods use learned features rather than raw activations, based on the belief that neurons do not cleanly decompose computation. In our new work, we show MLP neurons actually do support sparse, faithful circuits!

An investigation into representations formed by LLMs during inference tasks concludes that human-like conceptual representations can emerge purely from language prediction without real-world grounding. In PNAS: ow.ly/tCjT50XotMN

Finally, thanks to the reviewers and the PNAS editors for their constructive and thoughtful feedback, and for helping shape the final version of the paper into what it is today. 11/N = 11

📄 Paper: doi.org/10.1073/pnas.2… 🔨 Code and data: github.com/ningyuxu/llm_c… Work by Ningyu Xu @xny_ele, Qi Zhang @QiZhang_FDU, Chao Du, Qiang Luo, Xipeng Qiu @xpqiu, Xuanjing Huang @HXuanjing, and Menghan Zhang @HansonMeng84157 10/N

Our paper is now out in PNAS!💡 Are LLMs developing human-like concepts that are central to human cognition? If so, how are such concepts represented, organized, and related to behavior? doi.org/10.1073/pnas.2… 1/N

🥳 Introducing SpeechGPT 2.0-preview: A GPT-4o-level, real-time spoken dialogue system! (Only Chinese for now) 🎆 Highlights: ～⚡️ Real-time speech-to-speech dialogue with latency under 200ms ～😊 Rich in emotion and diverse in style, with strong speech style generalization ～🦁 Strong role-playing capabilities 🤖️ Try it out: Online system: sp2.open-moss.com Github: github.com/OpenMOSS/Speec… More demos: open-moss.com/en/speechgpt2-…

@GenBench Our work builds upon our earlier discoveries, showing that cross-linguistic syntactic difference is captured by LLMs within their representation space and significantly impacts their cross-lingual transfer performance. aclanthology.org/2022.emnlp-mai…

Had an incredible week at #EMNLP2023! I gave my 1st in-person presentation at the @GenBench workshop (on aligning conceptual correspondence for cross-lingual generalization) and had so many interesting conversations! You can check our paper at arxiv.org/abs/2310.12794….

@weGotlieb @tpimentelms @clara__meister @kmahowald @ryandcotterell Congratulations!!

Thank you to #EMNLP2023 chairs for the 😱 two 😱 outstanding paper awards! I am so grateful to have worked on these projects with wonderful colleagues — @tpimentelms (who is the first author on one of the papers!), @clara__meister, @kmahowald and @ryandcotterell