←
david yan184 posts
david yan avatar✓

david yan ✓

@dzyan01

phd student @PrincetonVL, research intern @elorianai. formerly: bse @princetoncs

◷ Joined March 2025▧ 32 Photos and videos
david yan avatar
david yan@dzyan01·Nov 9

On BrickVision, existing models vary widely in performance! 6/7

Post attachment
david yan avatar
david yan@dzyan01·Nov 9

Because many existing VQA benchmarks are solvable with simple textual references, we introduce BrickVision, a challenging benchmark designed to measure a model's capability to reason with undescribable visual references. 5/7

Post attachment
david yan avatar
david yan@dzyan01·Nov 9

To explain these results, we demonstrate that grounded reasoning provides advantages on tasks that require undescribable references, but not when objects are describable in language. 4/7

Post attachment
david yan avatar
david yan@dzyan01·Nov 9

We also find that, when compared to text reasoning, grounded reasoning is not particularly helpful for general VQA, but improves counting performance! 3/7

Post attachment
david yan avatar
david yan@dzyan01·Nov 9

We introduce a new technique to measure grounding usage in existing reasoning models and find that many open models don't actually rely on their generated coordinates when reasoning. 2/7

Post attachment
david yan avatar
david yan@dzyan01·Nov 9

Grounded reasoning is a popular thinking strategy for VLMs. But do models even use their grounding? (spoiler: not always!) In our @ElorianAI paper, we dig into the grounded reasoning recipe and release BrickVision, a benchmark that’s challenging to solve without grounding. 1/7

0:30
david yan avatar
david yan@dzyan01·Nov 6

super cool use of infinigen for zero-shot sim2real!

2:47
david yan avatar
david yan@dzyan01·Nov 6

"Personal reward" here refers to some abstract net reward relative to other options (e.g. working in industry). Person b is not necessarily unhappy or sad, but just could be happier elsewhere! In contrast, doing a PhD is locally optimal for person a.

david yan avatar
david yan@dzyan01·Nov 6

Random thought on "why do a (ml/ai) phd?" I believe AI has shortened one's optimization horizon, due to high future uncertainty. As such, you should be confident that you will get positive short-term value, instead of indexing on the eventual delayed reward from finishing.

Post attachment
david yan avatar
david yan@dzyan01·Nov 2

To reproduce the demo scene, just run python -m infinigen_examples.generate_nature --seed 2 --task coarse populate fine_terrain --output_folder out/mountain_2 --configs mountain noisy_video performance -p "execute_tasks.frame_range=[1,10]" land_process_pool.enabled=True 3/3

david yan avatar
david yan@dzyan01·Nov 2

Excited to announce a major performance update for Infinigen! We've overhauled various assets and systems, focusing on efficiency. Our `--performance` config reduces the average generation time of nature scenes by 8x without sacrificing any visual fidelity! 1/3

0:16
david yan avatar
david yan@dzyan01·Oct 29

my bitter lesson is finding out that caffeine significantly increases my focus and productivity

Kartik Garg avatar
⟳ david yan retweeted
Kartik Garg@_pacificax·Oct 27

Sprout Jackson

0:19