Blog | Yuchen Li

Heidelberg Laureate Forum: the hidden paradise I had been looking for

Tue, 13 Oct 2026 00:00:00 +0000

The Heidelberg Laureate Forum is, without exception, the best academic event I have ever attended. Heidelberg itself is a sanctuary of scholarship with the charm of a classical German town (it even escaped the bombs of the Second World War) — so much so that I can’t bring myself to post about it on rednote. If too many people find out, 20,000 will apply next year, the acceptance rate will drop below one percent, and it won’t be fun anymore. So I’m only sharing it with friends here — consider this my recommendation.

The HLF is a forum for young researchers in mathematics and computer science. Each year about 200 of them attend — undergraduates, graduate students, postdocs and early-career scientists — half from computer science, half from mathematics. Their research is wonderfully diverse, and not that many of them work on AI, which made me marvel that there are still mathematicians and computer scientists who don’t do AI, haha (probably because my own AI concentration is far too high).

Of course, AI came up again and again this year, because people feel a strong sense of urgency about it. Some researchers have joined the AI journey (like this year’s Fields Medallist Jacob, who joined OpenAI and founded Maisi), while others are still holding out.

More than 20 laureates of the Fields Medal, the Nobel Prize, the Turing Award, the Abel Prize and the ACM Prize in Computing were there: I finally met people who had only existed in my textbooks, the pioneers and founding figures of computer science and mathematics. You can have long heart-to-heart conversations with them, or say hello and take photos at all kinds of events. My strongest impression from talking to them: they are real, living people with no airs at all, and they gave me a lot to think about philosophically (since I couldn’t follow the mathematics).

Still, the happiest part was being social with a crowd of young people — singing, dancing, having fun together. During research hours we may all be nerds, but once the party starts everyone unleashes a second personality: high energy, full of charisma, playing to their heart’s content.

The young researchers came from more than 60 countries — very much a global village. The organiser, the HLFF, planned every part of the week with great care, balancing science and socialising beautifully, and took us to Heidelberg’s sights, its castle and estates, and all sorts of good food. In a way, this is the hidden paradise I had been searching for all along.

The idea of looking for a “hidden paradise” first came to me in 2020, at a research poster competition at KAUST. Despite my clumsy spoken English, I felt that kind of atmosphere for the first time: a global village, where people could talk about anything under the sun and be curious about endless horizons and countless other people. You could talk tech, or fall back on anime and rap. Later, DBA took this kind of social life to its peak; then it went into a steep decline and stopped abruptly, and I gradually turned into a research introvert.

Some will ask: what is the point of going to something like this?

What is the point, anyway?

Pausing your research for a week, going to an unfamiliar country to be a “social terrorist” (Chinese internet slang for someone fearlessly outgoing), meeting a crowd of different people and stepping into different identities — it is like living your life one more time.

The existence of a paradise like this is the point in itself.

Oh, and all of this was funded by the HLFF. I also received the Abbe Grant issued by the Carl-Zeiss-Stiftung. Of course, those are the least important part. If I have the chance and the time, I’ll be back!

Perceptio: letting a vision-language model sketch space before it answers

Tue, 06 Oct 2026 00:00:00 +0000

Paper: arXiv 2603.18795 · Hugging Face · Long-form write-up on Zhihu (in Chinese)

The problem

Large vision-language models are very good at what is in an image and surprisingly weak at where. Ask which of two objects is closer to the camera, or to segment “the second mug from the left”, and the model has to infer the geometry implicitly. It never writes down a spatial interpretation that it, or we, could check.

The idea: perceive first, then answer

Perceptio puts perception into the token stream. Before it answers, the model generates two kinds of spatial tokens inside the same autoregressive sequence:

Segmentation tokens based on SAM2, which ground the object being talked about;
Depth tokens from a VQ-VAE codebook that we distil from a strong monocular depth teacher, so a dense depth map becomes a short, discrete sequence.

The answer is then conditioned on the model’s own spatial “sketch”. You can think of it as a chain of thought made of masks and depth codes instead of words, and, because the tokens decode back into a mask and a depth map, it is one you can look at.

Making depth tokens trainable

Generating long runs of depth codes is unstable if you only use the usual next-token loss. Two things helped:

Composite depth-token objectives: a marker, a token and a count loss, used together;
Soft merging, which makes depth reconstruction from the predicted tokens differentiable, so the model is also supervised on the depth map it implies.

The model is built on InternVL and co-trained on a mix of tasks, so the same perception tokens serve segmentation, spatial question answering and general VQA.

Results

Compared with the strongest prior baselines reported in the paper:

Benchmark	Gain
Referring segmentation, RefCOCO / RefCOCO+ / RefCOCOg (cIoU)	+0.8 / +1.4 / +1.1
HardBLINK spatial understanding (accuracy)	+10.3%
MMBench (accuracy)	+1.0%

The HardBLINK jump is the result I care most about: it is the benchmark where “just look harder” does not work and explicit geometry does.

What’s next

I think explicit, inspectable intermediate perception is a good direction for spatial reasoning in multimodal models, and for world models more broadly. That is the line I want to keep pushing after my PhD.

This work was done during my internship at Amazon (Prime Video) with Amanmeet Garg, Shalini Chaudhuri, Rui Zhao and Garin Kessler. Thanks to all of them.

Everything about Yuchen Li, verified: a bilingual index of press, papers and social posts

Wed, 30 Sep 2026 00:00:00 +0000

Open the full interactive index → /press/

There are a lot of people called Yuchen Li — and a Chinese novel whose two protagonists’ names happen to combine into my Chinese name. So when I collected everything written about me across Chinese and English media, I applied one rule: a third-party item is only listed when its own text names Yuchen Li (or my Chinese name) together with an identity anchor — MBZUAI, KAUST, the SUSTech–KAUST 3+2 programme, SUSTech class of 2017, the 13th Heidelberg Laureate Forum, Dubai Business Associates, one of my handles (@Dr_Yuchen, @yuchenlichuck, “Dubai Prince” on rednote) or one of my papers. Each entry on the index shows the exact sentence that ties it to me.

What’s in it

Media (in Chinese) — 36Kr and Tencent News on 5Y Capital’s 72-hour AI Survival Challenge, plus my own Bilibili episode An AI PhD’s 72 hours; the DeepMENA podcast episode on studying in the Gulf (Ximalaya / Xiaoyuzhou / Apple Podcasts); NextCapital’s transcript of my talk on the Middle East 2030 vision; the Shenzhen Foundation for International Exchange and Cooperation’s two Dubai Business Associates articles; the Zhihu deep-dive on Perceptio; and SUSTech’s official National Day / Mid-Autumn feature quoting me from the Chinese Embassy reception in Abu Dhabi.
Papers, fellowships & workshops — the 13th Heidelberg Laureate Forum (HLF 2026) Young Researcher selection & Carl-Zeiss-Stiftung Abbe Grant (Top 3%), Perceptio (arXiv 2026), UCanDance (Sony AI & MBZUAI 2026), 3DCoMPaT++ (TPAMI 2025), Exploring Scaling Laws of PointNets (3DV 2025 Nectar Track Spotlight), PointNeXt (NeurIPS 2022), 3D CoMPaT (ECCV 2022 oral), Prototypical Random Walks (AAAI-21 workshop), and the CVPR 2023 C3DV workshop I helped organise.
Campus & school — SUSTech department and college posts from 2017–2021, and Hefei No. 8 High School’s physics-olympiad and class-of-2017 posts.
Social — my 18 most-read rednote (Xiaohongshu) notes (“Dubai Prince”, 6.1K followers), my top Zhihu answers and articles (as Nanxiaohuan), and recent posts on X, LinkedIn and Instagram.
Talks on stage — the 36Kr WAVES 2025 “Let’s WAVES” talk-show night (36Kr recap + Waves video), my NextCapital salon talk on video, and Link Island’s recharge-programme recap.

How to read it

Filter by kind (news, video, podcast, social, papers, campus), by language, or search any title, outlet or piece of evidence. Switch the interface between English and Chinese with the toggle in the top bar. Items marked verified come from third parties; items marked own account were published by me.

The index is hand-compiled and was last updated on 24 October 2026 (with Google Scholar citation counts refreshed automatically). WeChat links were captured through Sogou WeChat search and can expire — if one stops opening, search the title inside WeChat.

Browse the index → /press/