@Robro612

PhD Student @jhuclsp | Prev: @mixedbreadai, @JinaAI_ Interested in IR & NLP

 Baltimore, MD
cs.jhu.edu/~rjha5/
Joined June 2015
  • Tweets698
  • Following538
  • Followers446
41Media
New 📄: we replicate XTR, a multi-vector retrieval method that makes ColBERT faster by avoiding its expensive step of gathering full document embeddings XTR is not a free lunch over ColBERT, but its training objective is useful for modern efficient engines like PLAID and WARP 👇🏼
2
27
5
66
12,385
People in the encoder community don't care about text capabilities right? @ruyimarone
3
5
228
Super valuable-information-dense blogpost on a super promising method. Congratulations!
Paulo Moura@paulomouraj21h
We surfed the dense and late-interaction waves… now it's time to SlideOn sparse retrieval 🐕 🚀 Meet SlideOn and mSlideOn, fully open SOTA sparse retrievers at 174M and 339M parameters, trained by @LightOnIO R&D Models: huggingface.co/collections/l… Blog: huggingface.co/blog/lightona…
2
3
8
342
Strongly agree! Tip-of-the-tongue (ToT) queries deserve more attention in the era of modern and agentic retrieval. Even today's advanced retrieval systems still struggle with many of these queries. As @Robro612 and @bclavie pointed out, ToT retrieval has to bring together many key capabilities like query decomposition, parallel search, user intent understanding, personalization, and more. While ToT queries may seem rare, prior research suggests that known-item searches, including ToT queries, account for a substantial portion of real-world information needs. Studies also show that searchers experiencing ToT states report greater frustration than those facing other memory-related search failures. This makes ToT retrieval an important problem, spanning everything from finding forgotten games, music, and books to retrieving emails and enterprise documents. @841io , @UnderdogGeek , Jaime, @maik_froebe, Yifan, @DarrenHo233, myself, and the broader community (@HamedZamani, @ekanou, Samarth) have been pushing this direction through, TREC ToT (2023–2025) and NTCIR ToT (2026). Great to see @ExaAILabs , Oblique Bench (@dianetc_), @Julian_a42f9a @HamedZamani and others starting to incorporate ToT queries into their retrieval benchmarks. Excited to see this problem getting more attention!
Ben Clavié
@bclavieOct 16
an increasingly important question is going to be: is your retriever good at tip-of-the-tongue queries? Not because you'll often encounter them in the wild, but because they capture how deeply your model understands intent. @dianetc_ was ahead of the curve here
3
3
16
1,523
Congratulations! So glad to have these indices in a nice ecosystem to build on / evaluate with.
Silvio Martinico@SilvioMartinicoOct 16
Just received the award for the #SISAP2026 Indexing Challenge here in Brno! 🏆🇨🇿 Very proud of the tusKANNy team (@fmnardini, @cosimorulli1, @rventurini_). We took home both 1st and 2nd place in Task 3 (Learned Sparse Retrieval)! 🧵👇 #VectorSearch #InformationRetrieval
1
16
394
I hope that these findings from BGE-m3 (2024) and NeoMME (2026) that show how jointly training multiple representations lifts all boats will encourage more people to jointly rather than parallel/sequential train late interaction models
2
4
35
1,054
When one model can produce multiple "views" (sparse, dense, multi-vector) of the same Q/D, the efficiency economics of TusKANNy's RerankIndex become even more favorable! This mixed-representation pipeline is exactly what their ECIR 2025 paper advocates. xitter.cf/SilvioMartinico/status…
Silvio Martinico@SilvioMartinicoOct 13
The Vector Search landscape is full of complex methods. At tusKANNy, we like to simplify things through abstractions. Meet RerankIndex: a general framework for two-stage search, now available in kANNolo 0.12 (Python & Rust) 🧵👇 #VectorSearch #InformationRetrieval
1
3
131
Rohan Jha retweeted
Surya
@kickingkeysOct 13
Gemini 4 Argon is the first model where It felt like an image, video and language model all at once. Everything in this thread is made with code. @GoogleAIStudio More examples + the full code for all 534 paintings below ⬇️
84
197
67
2,471
171,928
The Vector Search landscape is full of complex methods. At tusKANNy, we like to simplify things through abstractions. Meet RerankIndex: a general framework for two-stage search, now available in kANNolo 0.12 (Python & Rust) 🧵👇 #VectorSearch #InformationRetrieval
5
8
1
29
1,116
Rohan Jha retweeted
Aayush Mishra
@aamixshOct 12
Looped LMs promise dynamic computation by allowing easy tokens to exit early, yet Ouro computes all loops anyway. We fixed it: 1. Ragged KV: −25% inference FLOPs & memory on Ouro, same accuracy, no retraining. 2. Global exit objective: true difficulty-aware exits 🧵
2
10
19
765
Rohan Jha retweeted
Sumit@_reachsumitOct 9
Compact and Efficient Indexes for Learned Sparse Retrieval @fmnardini et al. cut the memory of learned sparse retrieval indexes by using medoid documents as block representatives and compressing the forward index with bit-packing and 4-bit quantization 📝arxiv.org/abs/2610.12300

Compact and Efficient Indexes for Learned Sparse Retrieval

This paper investigates how to substantially reduce the memory footprint of learned sparse retrieval indexes without sacrificing the efficiency of state-of-the-art retrieval data structures....

arxiv.org
12
43
1,552
Rohan Jha retweeted
Andrew Drozdov
@mrdrozdovOct 8
Will be at COLM today and tomorrow. Reach out if you'd like to meet up.
🤖 Made with AI
1
1
1
6
822
Rohan Jha retweeted
Tony Wu
@tonywu_71Oct 7
Cross-model retrieval makes so much sense and should be the norm: it reduces latency and unlocks retrieval on edge devices with limited performance degradation. Overall super interesting insights from the pplx team, alongside a very strong multimodal model family 🔥
Antoine Chaffin
@antoine_chaffinOct 7
Replying to @antoine_chaffin
One of the key features of this release is that both models share the same embedding space, which allows cross-model retrieval, a very convenient setup for production We evaluated the performance of using the 0.6B model to query a 9B index and compare it to symmetric 9B and 0.6B querying Results show that querying the 9B index with the smaller model boosts the scores against using the 0.6B index, at no query-time cost Using the 9B model for querying is still stronger, but you can close around half of the gap with the 0.6B without increasing latency
2
3
21
894
Now the interesting part: the shared space We train an 18B teacher contrastively, then distill it into the 9B and 0.6B students with LEAF-style representation distillation to align the student output with the teacher's for every token Result: both models live in the same multi-vector space
1
1
20
534
Rohan Jha retweeted
JHU CLSP@jhuclspOct 6
Congratulations to @KaiserWhoLearns and collaborators for the Outstanding Paper Award at @colm_conf!! 🎉🎉
6
1
20
3,256
I’ll be presenting this paper today at COLM in the Imperial Ballroom 4:30 PM to 6:30 PM. Come by if you’re interested!
Julian Killingback@Julian_a42f9aApr 17
Late-interaction retrieval models are widely used for their strong performance, but their representations can be utilized beyond just retrieval. Our new paper demonstrates that these representations can effectively replace raw document text in RAG tasks.
4
26
818
The 🐐 of keeping benchmark scores high and scientific standards higher
Antoine Chaffin
@antoine_chaffinOct 6
Replying to @antoine_chaffin
If you need more proof that our enhancements are not benchmaxxing than keeping top-1 on a totally new leaderboard, you can look at the gap between public and private splits on the new index Some models collapse, our just does not flinch
1
2
12
856
Had a great time at ACL this weekend, but I’m even more excited to spend this week at #COLM2026! Let’s chat about uncertainty calibration, economical agent orchestration, and any other bottlenecks to widespread adoption. Catch me at our poster on Wednesday morning, DMs open!
2
4
27
1,264
Our CTO #DataScience Speaker Series welcomes 2025 Karen Spärck Jones (KSJ) Award recipient @andrewyates of @JohnsHopkins' HLTCOE to our Engineering office in NY today to talk with our #AI researchers about learned sparse retrieval bloom.bg/4jqvw9E #InformationRetrieval
Andrew Yates is a Senior Research Scientist with the Human Language Technology Center of Excellence at Johns Hopkins University.

ALT Andrew Yates is a Senior Research Scientist with the Human Language Technology Center of Excellence at Johns Hopkins University.

4
8
925
Single-mode relevance labels produce sparse, noisy supervision for training and eval, and can't be reshaped for different downstream uses. Good to see answer/evidence splits and rubric-based gains showing up in the same week (here and @cohere's RCP-nDCG@10 just today!)
Perplexity
@perplexity_aiSep 30
We built a new way to train contextual embedding models, which encode each chunk of a document with the whole document in view. pplx-embed-v2-context-9b-preview sets a new state of the art on ConTEB and @turbopuffer's new, privately held context-bench. perplexity.ai/hub/blog/conte…
1
8
687
Rohan Jha retweeted
Orion Weller@orionwellerSep 30
Life update: I joined @GoogleDeepMind earlier this year! Very happy to finally be releasing a non-flash model 😅 Lots of positive changes internally in the last couple months @GoogleDeepMind and I couldn’t be more optimistic about the upcoming releases we have in the pipeline 🚀
Logan Kilpatrick
@OfficialLoganKSep 30
Introducing Gemini 4 Argon, our new frontier model, rolling out to cyber defenders starting today, and more widely as soon as possible. I am really excited by the progress we have made here. Argon is priced at $2 in and $10 out during introductory pricing!
28
24
3
940
30,688
Load more