@TaimingLu

Ph.D student at @Princeton | Formerly @JohnsHopkins ’25, @HopkinsDSAI @JHUCompSci @jhuclsp @CCVLatJHU | AI/ML/NLP/CV

 Princeton, NJ
taiminglu.com
Joined June 2024
  • Tweets20
  • Following642
  • Followers382
6Media
You want a strong small LLM. Would you start small — or inherit from something bigger? 📄 New paper: Small LLMs: Pruning vs. Training from Scratch We find that pruning is more than a better initialization: simply giving randomly initialized LLMs more training tokens is often not enough to catch up.
14
120
10
619
54,001
Taiming Lu retweeted
Zhuang Liu
@liuzhuang1234Jun 22
Can a small academic team build a strong text-to-image model using only public datasets? Introducing i1: a simple, fully open recipe for strong text-to-image models
11
78
9
440
71,521
Knowledge doesn't always flow downhill. We find that in LLM pretraining, a weaker teacher can improve a stronger student, and pushing the teacher further can actually hurt. New paper: Strong Teacher Not Needed? On Distillation in LLM Pretraining.
A weaker teacher can still help and an overly strong one can hurt — what matters for distillation in LLM pretraining is teacher–student compatibility, not raw teacher strength.

ALT A weaker teacher can still help and an overly strong one can hurt — what matters for distillation in LLM pretraining is teacher–student compatibility, not raw teacher strength.

7
57
6
348
48,633
more replies
Distillation improves generalization more readily than in-domain fit. Out-of-distribution perplexity and downstream accuracy improve more consistently than in-domain perplexity, where some configurations help OOD/downstream while doing nothing for in-domain.
As α grows, in-domain improvement collapses onto a single curve in teacher loss, but OOD and downstream split by architecture, meaning teacher train loss alone is a misleading proxy for teacher quality.

ALT As α grows, in-domain improvement collapses onto a single curve in teacher loss, but OOD and downstream split by architecture, meaning teacher train loss alone is a misleading proxy for teacher quality.

1
2
7
1,230
Teacher–student compatibility matters more than raw teacher strength. This changes how you pick a teacher: both for frontier training (where the best available teacher is often a prior generation) and for efficient small models, where "bigger teacher is better" isn't the right rule. Thanks @liuzhuang1234 for the support! arxiv: arxiv.org/abs/2605.23857 code: github.com/zlab-princeton/st…
See our paper for more details.

ALT See our paper for more details.

1
2
10
1,151
Stronger Normalization-Free Transformers – new paper. We introduce Derf (Dynamic erf), a simple point-wise layer that lets norm-free Transformers not only work, but actually outperform their normalized counterparts.
19
174
13
1,077
167,066
🤯 Think better visuals mean better world models? Think again. 💥 Surprise: Agents don’t need eye candy— they need wins. Meet World-in-World, the first open benchmark that ranks world models by closed-loop task success, not pixels. We uncover 3 shocks: 1️⃣ Visuals ≠ utility 2️⃣ Action data > bigger models 3️⃣ Scaling test-time compute = more success 🤗 huggingface.co/papers/2510.1… 🌍 world-in-world.github.io 📄 arxiv.org/abs/2510.18135 github.com/World-In-World/wo…
2
37
3
153
42,919
Excited to share our lab’s first open-source release: LLM-Distillation-JAX supports practical knowledge distillation configurations (distillation strength, temperature, top-k/top-p), built on MaxText designed for reproducible JAX/Flax training on both TPUs and GPUs
4
28
1
218
20,826
Meet the AI system that can envision an entire world from a single picture. @genex_world—developed by @jieneng_chen, @YuilleAlan, @TaiMingLu, @DanielKhashabi, & @tianminshu—imagines in-depth scenarios to make informed decisions. Learn more here: hub.jhu.edu/2024/12/19/a-gen…
6
12
1,843
Thrilled to introduce GenEx: Generating an Explorable World. ✨ ✨ GenEx takes a single image 🖼️ and create a 3D generative world 🌍 — you can dive in for interactive exploration, and so as embodied AI agent. Follow our X for more demos: xitter.cf/genex_world Paper on huggingface: huggingface.co/papers/2412.0… Tech details: genex.world (1/n)
2
32
2
103
10,150
Taiming Lu retweeted
GenEx@genex_world13 Dec 2024
Introducing GenEx: Turn any image into a 3D world adventure! 1️⃣ Create a fully explorable 360° world in 3D from just a single image! 2️⃣ Explore interactively or with GPT assistance. 3️⃣ Advance embodied AI with this imagined world! Check out our website: genex.world
2
11
6
37
10,589
Introducing Genex: Generative World Explorer. 🧠 Humans mentally explore unseen parts of the world, revising their beliefs with imagined observations. ✨ Genex replicates this human-like ability, advancing embodied AI in planning with partial observations. (1/6)
6
51
11
164
37,050
Taiming Lu retweeted
Muhan Gao@muhan_gao28 Jun 2024
🤖LLMs know more long-context information than they show! 🔍Probing reveals higher accuracy than generation output. #LLMs know but don't tell.🤐 The earlier relevant information is learned within the layers, the higher the final output accuracy! 📈 (arxiv.org/abs/2406.14673)
5
6
14
2,198
Load more