Aman Singh Thakur

MLE Amazon | LLM Research | UMass MSCS

[Aman Singh Thakur @iamsingh96aman]

WARP: Wasserstein-Aligned RAG for Population Opinions accepted at EMNLP2026 Industry Track 🎉
@emnlpmeeting

Ask a Traditional RAG system "what do people think about X?" and every sentence it gives back is grounded in a real review.

It can still be wrong. If 40% of guests complained about noise and none of them made the top-k, the answer reads like consensus.

WARP fixes this at re-rank time under 330ms latency 📷 A thread - [1/n] 📷📄

📄 https://t.co/577wBMbOSZ

[Aditya Agrawal @bringyourdata]

Nice surprise: Jack Ding cited our position paper in his piece on why AI needs humans it can't replace.

His take on the opinion problem in AI search is worth a read.

https://t.co/2kh5PTN4mK

[Aditya Agrawal @bringyourdata]

Our opinion-aware RAG work just got cited in the OWASP AI security verification standard draft 🎉

They flag our exact point: correctly-cited answers can still misrepresent the opinion landscape.

Source: https://t.co/QYIHDuwK9l

[Aditya Agrawal @bringyourdata]

1/6 Excited to share that WARP: Wasserstein-Aligned RAG for Population Opinions is headed to EMNLP 2026! Led by @iamsingh96aman, with Alwarappan Nakkiran and Alex Karlsson.

🧵 on what we built:

[Aman Singh Thakur @iamsingh96aman]

@emnlpmeeting Work possible due to -

Aditya Agrawal, Sr. Applied Scientist, Amazon
Alwarappan Nakkiran, Applied Science Manager, Amazon
Alex Karlsson, Sr. Data Engineer, Amazon

@AmazonScience @amazon

[Aman Singh Thakur @iamsingh96aman]

WARP: Wasserstein-Aligned RAG for Population Opinions accepted at EMNLP2026 Industry Track 🎉
@emnlpmeeting

Ask a Traditional RAG system "what do people think about X?" and every sentence it gives back is grounded in a real review.

It can still be wrong. If 40% of guests complained about noise and none of them made the top-k, the answer reads like consensus.

WARP fixes this at re-rank time under 330ms latency 📷 A thread - [1/n] 📷📄

📄 https://t.co/577wBMbOSZ

[Aman Singh Thakur @iamsingh96aman]

@emnlpmeeting @Yelp WARP: Wasserstein-Aligned RAG for Population Opinions

No fine-tuning, no retriever changes, no extra inference calls. <330ms p99 re-ranking, more opinionated results!

To appear at #EMNLP2026 Industry Track (oral) 🎉 - [4/n]

📄 https://t.co/PxhZ8r3lQL

[Aman Singh Thakur @iamsingh96aman]

Across 3 review domains, 35K docs, 156 queries, 26 entities, WARP results:

→ ≥43% reduction in distributional error vs. top-k. 43% Seller Forums · 79% Yelp · 69% OpinRank
→ 88% on Yelp with pool expansion.

Just entity selection boosts Entity match rate on
@Yelp : 40.1% → 99.2% - [4/n]

[Aman Singh Thakur @iamsingh96aman]

WARP: Wasserstein-Aligned RAG for Population Opinions

No fine-tuning, no retriever changes, no extra inference calls. <330ms p99 re-ranking, more opinionated results!! [5/n]

To appear at #EMNLP2026 Industry Track (oral) 🎉 📷

https://t.co/577wBMbOSZ

[Aman Singh Thakur @iamsingh96aman]

Across 3 review domains, 35K docs, 156 queries, 26 entities, WARP results:

→ ≥43% reduction in distributional error vs. top-k. 43% Seller Forums · 79% Yelp · 69% OpinRank
→ 88% on Yelp with pool expansion.

Just entity selection boosts Entity match rate on
@Yelp : 40.1% → 99.2% [4/n]

[Matt @mattamys]

Did anyone else catch this in F1's Cooldown Room?
During Max's crash replay, they had an ai video transition from the trackside camera to Max's onboard camera - what do you think of the effect? 🏁

[Aman Singh Thakur @iamsingh96aman]

🙌

[Natural Language Processing Papers @HEI]

Training Leaves Traces: Centered Residual Signatures for Language Model Lineage Verification

Aman Singh Thakur, Rayan Khoury
https://t.co/ShpOGo3NK5 [𝚌𝚜.𝙲𝙻 𝚌𝚜.𝙻𝙶]

[Aman Singh Thakur @iamsingh96aman]

We establish a new passive, datafree, fast model provenance signal.

Caveats: residual architectures only, white-box access needed, and the score says two checkpoints are related, not which came first.

Paper: https://t.co/auKaslxec6

Joint work with Rayan Khoury. If you care about open weight provenance, reach out! [6/n]

[Aman Singh Thakur @iamsingh96aman]

🚨 New Preprint on intrinsic model fingerprinting 🚨

Training Leaves Traces: Centered Residual Signatures for Language Modeling Lineage Verification

In this new preprint, we ask - Can the weights alone tell you whether two model checkpoints share ancestry?

Open weight models get fine-tuned, quantized, pruned, merged, renamed. The only link to the parent is a model card. Hashes break, watermarks must be inserted before release.

We built a data-free fingerprint. A thread [1/n]

https://t.co/auKaslxec6

[Aman Singh Thakur @iamsingh96aman]

Stress tests on real 7B models: shuffle hidden units and our score moves less than 10^-7.

Weight cosine loses 93% of its signal. 76x faster too. Quantization, LoRA merges, fine-tuning, pruning to 70%: all inherited.

Distilled students score ~0. Mimicry is not ancestry.
https://t.co/4T50hHWs3i
[5/n]