Raghuveer T

PhD Student at Duke University; Intern at @Meta FAIR.

[Bhuwan Dhingra @bhuwandhingra]

New preprint from @ghazalkhn and @raghavlite!

The dominant paradigm for multimodal document retrieval has been to represent pdf pages as image inputs to a VLM. We show that this approach is suboptimal for text-rich scientific documents.

With the same VLM, keeping text as tokens and interleaving with figure images leads to better performance than page-level images -- despite not being trained on such sequences!

Text embeddings alone also handle figure-based queries quite well -- by using captions and in-text references.

Overall, text-based embeddings are still quite effective for multimodal retrieval! Lots more interesting findings in the paper: https://t.co/OL2U9adNTn

[Shuyan Zhou @shuyanzh36]

It was really fun to meet new people and discuss agent environments. Thanks to the workshop organizers for putting together such a great event! Here is the slide deck from the talk: https://t.co/uyNh7l0vIn

[Dhanajit Brahma ✈️ ACL 2026 San Diego @DhanajitB]

👋 Hello #SanDiego!
We are excited to present our latest work bridging Causal Inference and GenAI at #NeurIPS2025.
For tackling unobserved confounders, come visit on Dec 5, 2025 at 11 AM PST.
📍Poster 2516
Coupling Generative Modeling and an Autoencoder with the Causal Bridge🧵👇

[DukeNLP @duke_nlp]

Our entire group has come to San Diego this week and we have a series of presentations at #NeurIPS! Come find us and chat about contrastive learning and LLM agents.
@bhuwandhingra @raghavlite @syz0x1 @JunlinWang3 @RoyXie_ @DCirci

[Raghuveer T @raghavlite]

📢 New preprint announcement!!

- Most image encoders are trained independently before being integrated into a VLM, resulting to generic, query-agnostic image representations that constrain downstream VLM performance.

- 🚀 We introduce the Text-Guided Semantic Image Encoder (TIE), which produces query/image-conditioned image representations, enabling VLMs to operate on task-relevant image features.

- Across nine image-to-text benchmarks, TIE-based VLMs achieve +1.5–1.3 average improvements, with gains of up to +6 points on DocVQA and InfoVQA.

- Notably, despite using only half the image tokens, and therefore offering faster inference, TIE-based VLMs outperform their standard counterparts.

Paper: https://t.co/MS1ZPg1A4a

[Raj Shah @rajsanjayshah]

Excited to be in Montreal this week for COLM 2025!
I have two papers accepted this year:
1️⃣ The Unlearning Mirage, a dynamic framework for evaluating LLM unlearning.
2️⃣ With @Harsh_N_Lalai10, using 20 Questions as a creative framework to systematically evaluate geographical bias in LLMs.
Check out the individual threads!

[Raghuveer T @raghavlite]

Breaking the Batch Barrier (B3) was accepted as a Spotlight presentation at NeurIPS 2025. Do check it out!!

[Bhuwan Dhingra @bhuwandhingra]

📢 New Preprint from @raghavlite on Multimodal Contrastive Learning: Breaking the Batch Barrier (B3) 📢

TL;DR: Smart batch mining based on community detection achieves state of the art on the MMEB benchmark.

Preprint: https://t.co/BU85rYwDBp
Code: https://t.co/bo2xLRBG7Q

[Bhuwan Dhingra @bhuwandhingra]

📢 New Preprint from @raghavlite on Multimodal Contrastive Learning: Breaking the Batch Barrier (B3) 📢

TL;DR: Smart batch mining based on community detection achieves state of the art on the MMEB benchmark.

Preprint: https://t.co/BU85rYwDBp
Code: https://t.co/bo2xLRBG7Q

[Raghuveer T @raghavlite]

I will be presenting ✨GenEOL: Harnessing the Generative Power of LLMs for Training-Free
Sentence Embeddings✨at #NAACL2025!

We propose a simple and effective training-free inference time method to achive high quality embeddings

- Beats other trainig-free methods by upto 4 points on STS benchmark.
- Notable gains over training-free, unsupervised methods on the MTEB benchmark
- Robust to prompt perturbations

Link: https://t.co/Txq10epNBR

📅 4/30, Wed, 2:00-3:30 Hall 3

Happy to chat more about embeddings and my current research!

[Vered Shwartz @VeredShwartz]

I just published Tips for Writing NLP Papers https://t.co/u9zbHdsMC5

I wrote it for my students so I don't have to sound like a broken record (and edit papers for the same issues over and over again 😃). But some of you might find it useful too.

[Shridhar @JupyterAI]

For very long, OntoNotes has served as the most important benchmark for coreference resolution.
We are happy to present "LongtoNotes" that extends OntoNotes to up to 8X longer documents.
Paper : https://t.co/HZlvTP8ZzK

[Raghuveer T @raghavlite]

@buzzganesh @coling2020 @mageed @laksvs Awesome. Keep rocking