PhD Student @jhuclsp | Prev: @mixedbreadai, @JinaAI_ Interested in IR & NLP
Baltimore, MD
Joined June 2015
- Tweets732
- Following550
- Followers453
- Likes8.8K
Pinned Tweet
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 👇🏼
Rohan Jha retweeted
Introducing e-means, an algorithm beating k-means in 10x fewer epochs.
With order of magnitude wall-clock speedups, e-means is released in FAISS and needs no hyperparameter tuning beyond k. It is robust to various scales and datasets. Still skeptical ? Just try e-means [1/10]🧵
Rohan Jha retweeted
No one told me recreating the @cartesia blog post would be this fun (and this exciting!)
Thinking of blog posting some of the recreation / improvement process, but for now my favorite findings are below. (1/4)
Very excited by the new e-means paper from @swann_bessa et al.! Better convergence at a fraction of the training time than conventional Lloyd kmeans.
The blogpost has lots of beautiful visualizations/animations of the studies in the paper!
I first heard about this technique in VecFlow-Chamfer (quite similar intuitions to TACHIOM but for unified-memory GPUs). It has the benefit of not requiring a tf-idf inductive bias (visual docs) but is definitely worse than TAC when you have token ids.
dl.acm.org/doi/10.1145/37867…
4
130
Rohan Jha retweeted
Is there a way to scale a reasoning model’s judgment for searching a massive corpus?
We introduce the Amortized Reasoning Index (ARI), a new kind of retriever that updates its weights during search, using gradients derived at inference time from the ranking preferences of a reasoning model.
Every session, current agents spend lots of reasoning deciding which search results matter, and then throw all of that judgment away upon compaction. ARI is our attempt to “amortize” that judgment, so that reasoning spent on past queries keeps improving search on future ones.
With no relevance labels, ARI takes a 400M cross-encoder that starts out weak and grows its Recall@100 on held-out queries by up to 7×, ending above Qwen3-Embedding-8B on various oblique/reasoning-intensive search tasks.
Read our blogpost: dianetc.com/musings/ari/
Preliminary results and discussions are below!
Rohan Jha retweeted
Need different model weights? Just change them 🙃
Our new paper VFold does exactly that to enable low-overhead cross-layer value cache compression: half the value cache, >98% of full-cache performance 🧵
📄 arxiv.org/abs/2610.12338
Rohan Jha retweeted
An awesome final project by @zhengping_jiang as he wraps up the PhD. Since 2018 when I first met him there have been so many good papers. This one is especially topical to a lot of labs and businesses (he's on the market!).
Long-horizon agents can take hundreds of actions before receiving a single outcome reward. Training them raises two questions:
➡️ Which actions deserve credit?
➡️ How does training fit with existing compaction and memory strategies?
Introducing FOCUS: active critic learning over managed context states, with modular, customizable context protocols and integrations with Tinker and slime.
Docs: zipjiang.github.io/FOCUS (1/4)
Rohan Jha retweeted
Long-horizon agents can take hundreds of actions before receiving a single outcome reward. Training them raises two questions:
➡️ Which actions deserve credit?
➡️ How does training fit with existing compaction and memory strategies?
Introducing FOCUS: active critic learning over managed context states, with modular, customizable context protocols and integrations with Tinker and slime.
Docs: zipjiang.github.io/FOCUS (1/4)
Rohan Jha retweeted
I think LLM inference might be in the middle of a paradigm shift. Read my blog if you want my full thoughts. TLDR is that I think we are moving from batch parallelism to sequence parallelism
Rohan Jha retweeted
Foundation models are trained in stages: pretraining, midtraining, and post-training. But is this separation fundamental?
We introduce unified training: a single algorithm for the full pipeline. This gives
⚙️ Simpler infrastructure
🔬 A framework for better science
(1/n)
Highest compliment one can give to research-applicable software. I feel the same about PyLate, Sentence Transformers, and TuskANNy's libraries.
btw, tabpack doesn't just give you good out-of-the-box performance on benchmarks, it also speeds up research. We had a lot of MLP models to run for our paper's scaling study:
teapawt.org/puhsuuu/status/2113696…
We wouldn't even have considered running a grid this big before tabpack
Rohan Jha retweeted
you can just sparsify and compress thing
Rohan Jha retweeted
Applying to PhD programs is hard! Take advantage of this opportunity, especially if you may not otherwise have access to application support:
forms.gle/W2pXmman47CBoNZp9
Rohan Jha retweeted
Worried about the memory footprint of your HNSW index? You can now compress its edges without sacrificing search speed (and it works in production at eBay!).
Check it out: the compressed layout is already available in kANNolo!
@FBenoch, @fmnardini, @cosimorulli1, @rventurini_
Lossless Compression of Proximity Graph Indexes for Efficient Approximate Nearest Neighbors Search
Compresses HNSW edge lists by reordering node IDs offline and using fast SIMD integer codecs, cutting memory.
📝 arxiv.org/abs/2610.37176
👨🏽💻 github.com/tusKANNy/papers_r…
Rohan Jha retweeted
Lossless Compression of Proximity Graph Indexes for Efficient Approximate Nearest Neighbors Search
Compresses HNSW edge lists by reordering node IDs offline and using fast SIMD integer codecs, cutting memory.
📝 arxiv.org/abs/2610.37176
👨🏽💻 github.com/tusKANNy/papers_r…
This is an example of one of the biggest perspective shifts @ben_vandurme gave me when joining the lab - lots of different cost models make sense / must suffice when you expand you concept of retrieval tasks beyond textual web search
y'all obviously don't read enough vision papers to see the atrocities of "cross-encoders at scale." apparently anything goes when your test collection is only 1k videos...
1
5
509