Patrick Pynadath
@PatrickPyn35903
Followers
72
Following
92
Media
5
Statuses
32
Phd Student @purdue cs. working on making continuous gradients discrete
Joined October 2024
Continuous diffusion dominates images but fails on discrete data—despite learning continuous gradients that should enable coordinated updates. "CANDI: Hybrid Discrete-Continuous Diffusion Models" explains why and how why hybrid diffusion fixes it! (1/8)
4
12
61
How powerful are Diffusion LLMs? Can they solve problems that Auto-Regressive (AR) LLMs can’t solve? Check our new paper "On Powerful Ways to Generate: Autoregression, Diffusion, and Beyond" 🔗 https://t.co/aiGTbXMWFE In this work, we show while Diffusion LLMs are indeed more
16
77
372
We asked the same question: how can we combine the strengths of continuous and discrete approaches? Similar to CDCD, in our work, Purrception, we extend Variational FM to model VQ latents through continuous-discrete transport for image generation :D 👉 https://t.co/KIog9mLNWb
In diffusion LMs, discrete methods have all but displaced continuous ones (🥲). Interesting new trend: why not both? Use continuous methods to make discrete diffusion better. Diffusion duality: https://t.co/KPO56vDygp CADD: https://t.co/CNOIWcUIMo CCDD:
1
13
70
Amazing work by Olga ❤️ We extended Variational FM to Manifold domains for Protein and Material Design, and show how standard Riemannian FM connects to this variational counterpart. Special thanks to all great collaborators on this project. 🐳
Cool news: our extended Riemannian Gaussian VFM paper is out! 🔮 We define and study a variational objective for probability flows 🌀 on manifolds with closed-form geodesics. @FEijkelboom @a_ppln @CongLiu202212 @wellingmax @jwvdm @erikjbekkers 🔥 📜 https://t.co/PE6I6YcoTn
1
6
42
Overwhelmed by the number of Diffusion LLM papers? 🌊 Same here 😭 So I’m starting a Discrete Diffusion Reading Group (@diffusion_llms) with my favorite disciples @jdeschena and @zhihanyang_ ✨ We’ll cover everything—from theory to empirics, from language to molecules. Join
20
41
319
Awesome work by lab mate! if you are at EMNLP, make sure to check this out!
Can we accelerate test-time alignment? YES! 📃paper: Reward-Shifted Speculative Sampling Is An Efficient Test-Time Weak-to-Strong Aligner 🔗arXiv: https://t.co/hzDG2l9KZG 📌EMNLP 2025
0
0
2
Tired to go back to the original papers again and again? Our monograph: a systematic and fundamental recipe you can rely on! 📘 We’re excited to release 《The Principles of Diffusion Models》— with @DrYangSong, @gimdong58085414, @mittu1204, and @StefanoErmon. It traces the core
43
438
2K
The rehabilitation of continuous diffusion for discrete data continues! Check out CANDI by @PatrickPyn35903 @thjashin @ruqi_zhang Their insightful analysis explains why continuous methods have fallen behind, and why self-conditioning is so important. https://t.co/Bqn8Zd7hRz
In diffusion LMs, discrete methods have all but displaced continuous ones (🥲). Interesting new trend: why not both? Use continuous methods to make discrete diffusion better. Diffusion duality: https://t.co/KPO56vDygp CADD: https://t.co/CNOIWcUIMo CCDD:
3
22
144
Continuous diffusion dominates images but fails on discrete data—despite learning continuous gradients that should enable coordinated updates. "CANDI: Hybrid Discrete-Continuous Diffusion Models" explains why and how why hybrid diffusion fixes it! (1/8)
4
12
61
Shoutout to concurrent work CADD ( https://t.co/YZQbJnuW46) and CCDD ( https://t.co/2BDlFsyWmB) exploring hybrid diffusion! Our angle: we explain why continuous diffusion fails on discrete spaces and why hybrid methods work—providing theoretical grounding for this direction (8/8)
arxiv.org
Diffusion language models, especially masked discrete diffusion models, have achieved great success recently. While there are some theoretical and primary empirical results showing the advantages...
0
0
4
Interested? Check out these links! Project Page: https://t.co/9tTm81lhxX ArXiv: https://t.co/Xzc97EBZGR Github: https://t.co/jBPI6fddBO Grateful to @thjashin and @ruqi_zhang for guidance and help with this project! (7/8)
github.com
CANDI: Continuous and Discrete Diffusion. Contribute to patrickpynadath1/candi-diffusion development by creating an account on GitHub.
1
0
5
Benefits of continuous diffusion for discrete spaces: 1. Coordinated updates across positions—crucial at low NFE when many tokens update simultaneously 2. Trivial classifier guidance—just compute gradients from off-the-shelf classifiers and add to inference! (6/8)
1
0
2
Token identifiability predicts: Continuous diffusion: works on small vocabs, fails on large CANDI: effective regardless of vocab size We test on text8 vs. OWT with identical architectures. All three predictions confirmed—theory matches practice! (5/8)
1
0
2
Simple solution: decouple discrete corruption from noise level and coordinate both explicitly. CANDI's hybrid kernel draws from masked + continuous diffusion, letting models learn discrete conditional structure AND continuous score functions. Best of both worlds! (4/8)
1
0
2
Key finding: discrete corruption scales exponentially with |V|, but continuous degradation is independent of |V|. This creates "temporal dissonance": when the model learns conditional structure, it can't learn the score function, and vice versa. (3/8)
1
0
2
Token identifiability reveals two corruption mechanisms: Discrete Identity Corruption: is the wrong token closest? Continuous Rank Degradation: how many wrong tokens are closer? The first enables learning dependencies; the second, the score function. (2/9)
1
0
3
#ICLR2025 wrapped up on a high note. From insightful exchanges on LLMs, probabilistic inference, and optimization at the workshops to reconnecting with @PatrickPyn35903 from @PurdueCS. Grateful to my advisor @ZHANG_Ruda for the chance to be part of such a meaningful event.
0
5
18
Excited to present this at today’s Poster session! Quick update: poster number is 592. Whova event seems to be outdated, but ICLR website has correct info. Check out the project page if you want to read more! Link: https://t.co/Qc8nRMllxP Time: 10am-12pm, poster number 592
DAB is a controlled decoding algorithm using gradient-based discrete sampling. It achieves better fluency and constraint satisfaction—all with much less computational cost.
0
2
6
DAB is a controlled decoding algorithm using gradient-based discrete sampling. It achieves better fluency and constraint satisfaction—all with much less computational cost.
1
1
1
I won’t be attending #ICLR2025 this year, but my amazing students will be presenting several exciting works: 1️⃣ Inference-time safety in VLMs 2️⃣ Controlled decoding via discrete sampling 3️⃣ Gradient genetic algorithms for drug discovery 4️⃣ Single-step diffusion samplers Catch
1
4
32