Xavier Thomas (Rohan) retweeted
Do VLMs truly reason or memorize data? We show
❗even transformations like rotation & scaling still throw them off
❗finetuning fails to generalize across domains
tinyurl.com/colmbu26
📅 Stop by our poster on Wed @ 4:30 PM and
chat with @xavierohan #COLM2026 @COLM_conf
Xavier Thomas (Rohan) retweeted
Thank you @GoodfireAI for supporting my research group's research on interpretability of multi-modal LLMs. Excited to use Silico! 🚀
Xavier Thomas (Rohan) retweeted
Arsenal are the Premier League Champions. Every Arsenal fan right now 💪
Xavier Thomas (Rohan) retweeted
for the folks who know ball
Xavier Thomas (Rohan) retweeted
🔥 [New Paper] Semantic Richness or Geometric Reasoning?
❗️GPT-5.2, Gemini-2.5-Pro struggle on basic geometric tasks when semantic context is sparse.
❗️Performance is tied to data familiarity.
Conclusion: We are far from true visual invariance.
More here: xthomasbu.github.io/visual_i…
Paper: arxiv.org/pdf/2603.14056
Website (we have results on real-world real-time tasks on hardware and standard simulated benchmarks): keyed-drifting.github.io/
Code/models: coming next week!
Xavier Thomas (Rohan) retweeted
[1/2] The visual world is diverse—so why are the generative models stuck in a "one-patch-size-fits-all" architecture?
Meet Dynamic-DiT: A simple yet powerful technique that adaptively selects patch sizes based on how the latent manifold evolves at every denoising step.
Xavier Thomas (Rohan) retweeted
Thrilled to share that our work investigating recurrence in trained ViTs has been accepted at #ICLR2026! 🎉
It turns out DINOv2 may be secretly recurrent.🦖
A 2-block recurrent transformer can match its activations layer-by-layer and recover 96% of its ImageNet performance.
🧵
Are ViTs secretly RNNs? #ICLR2026
Our 2-block recurrent transformer recovers 96% of DINOv2’s IN-1k accuracy & reproduces its activations 1-to-1, motivating the Block-Recurrent Hypothesis: arxiv.org/abs/2512.19941
w/ @thomas_fel_ @RichieHakim @ABrondetta Demba Ba @t_andy_keller
Xavier Thomas (Rohan) retweeted
Grateful to be interviewed by @RSIPvision on my research journey so far rsipvision.com/ComputerVisio…
Xavier Thomas (Rohan) retweeted
[1/4] 🧠 The Great Modality Test: Are MLLMs Blind, Deaf, or Both?
Humans are perception pros: We instinctively know what information to trust.
Blindfolded, you describe the chirping birds.
Headphones on, you focus entirely on the silent visuals.
We robustly use the available and reliable modality.
Can SOTA MLLMs do the same?
We put the best Multimodal LLMs to the ultimate test.
Xavier Thomas (Rohan) retweeted
📢📢 Can we automate how humans instantly perceive if a generated video captures a human action done wrong?
Details: xthomasbu.github.io/video-ge…
Paper: huggingface.co/papers/2512.0…
Our key idea is to build a manifold of correctly performed actions from several real-world videos. Next, we measure how far off a generated video is from this manifold to get a robust measurement of action correctness.
We perform extensive subjective, quantitative, and quantitative evaluation of different actions, generative models, and MLLMs.
Work done by @xavierohan Youngsun, Ananya, and Audrey
Xavier Thomas (Rohan) retweeted
Thrilled to receive BU's Career Development Professorship award Thanks, BU for this very kind acknowledgement! bu.edu/articles/2025/bu-juni…
Xavier Thomas (Rohan) retweeted
🕳️🐇Into the Rabbit Hull – Part I (Part II tomorrow)
An interpretability deep dive into DINOv2, one of vision’s most important foundation models.
And today is Part I, buckle up, we're exploring some of its most charming features.
Xavier Thomas (Rohan) retweeted
At ICML for the next 2 days to present multiple works, if you're into interpretability, complexity, or just wanna know how cool @KempnerInst is, hit me up 👋


























