Aditya Agrawal

Sr. Applied Scientist @ Amazon | LLM & RAG Research | UMass MSCS

[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:

[Aditya Agrawal @bringyourdata]

Amazon Science covered WARP, our EMNLP 2026 paper. The core idea: calibrate retrieved evidence to the population's opinion distribution, so answers reflect what people actually think:

https://t.co/wm60JpSdny

[Aditya Agrawal @bringyourdata]

6/6 Paper: https://t.co/urwBAtmQtp

[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:

[Aditya Agrawal @bringyourdata]

5/6 Across 3 review domains (35K docs, 156 queries, 26 entities), WARP cuts distributional error by at least 43% at sub-second latency. The gains carry through to generation: a 5-judge LLM panel prefers WARP answers in 86% of decided comparisons at k <= 5.

[Aditya Agrawal @bringyourdata]

4/6 WARP recovers underrepresented opinions that cosine ranking buries, then selects documents whose sentiment-intensity distribution matches the population target using Wasserstein-1 distance, which respects the ordinal structure of opinions.

[Aditya Agrawal @bringyourdata]

3/6 Diversity re-rankers like MMR spread documents apart but aim at no target distribution. KL/JS calibration targets one, but treats opinion bins as unordered: mistaking strong positive for strong negative costs no more than a near miss.

[Aditya Agrawal @bringyourdata]

2/6 Ask a RAG system "what do people think about X?" and you get an answer that reads as consensus. Standard top-k retrieval ranks by query similarity, so minority views quietly disappear from the retrieved evidence.

[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]

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]

7/ We lay out a full research agenda in the paper. If you build or evaluate RAG, would love your thoughts.

with @iamsingh96aman

https://t.co/W8oHDTrpWc

[Aditya Agrawal @bringyourdata]

🎉 Our INLG 2026 position paper!

Ask RAG "what do people think about X?" — the answer mirrors the most similar docs, not the true opinion distribution. Majority amplified, minority erased.

This factual bias is structural. Fix: Opinion-Aware RAG 🧵

https://t.co/W8oHDTrpWc

[Aditya Agrawal @bringyourdata]

6/ Existence proof on seller forums + hotel reviews: 18–48% lower Wasserstein distance to true sentiment distributions, +26.8% sentiment diversity, +42.7% entity match. Humans preferred opinion-enriched answers 79.2% of the time.

[Aditya Agrawal @bringyourdata]

5/ Our fix: Opinion-Aware RAG (O-RAG). LLM-based opinion extraction + entity-linked opinion metadata, jointly optimized for coverage, fidelity, and fairness via Wasserstein distance.

[Aditya Agrawal @bringyourdata]

4/ This matters because RAG mediates what people read. A system that amplifies the majority view builds echo chambers — and makes opinions easier to manipulate.

[Aditya Agrawal @bringyourdata]

3/ We surveyed 34 major RAG benchmarks. Only one covers opinion synthesis. The bias runs through the datasets, the retrieval objectives, and the eval metrics.

[Aditya Agrawal @bringyourdata]

2/ We frame it through uncertainty. Factual queries should minimize posterior entropy. Opinion queries should preserve it — disagreement is the signal, and today's RAG discards it.

[Aditya Agrawal @bringyourdata]

@Bose I would be trying my best to avoid ordering products through the website. Horrible ordering experience.

[Aditya Agrawal @bringyourdata]

@riaghosh_19 @USEmbassyPH Hope they reconsider!

[Ria Ghosh @riaghosh_19]

I sincerely request you to reconsider her visa application on humanitarian grounds. Her appointment confirmation number is AA00BT9IS3.

[Ria Ghosh @riaghosh_19]

@USEmbassyPH My sister's visa was rejected today. She had all documents in order but was told to apply in India even though the backlog there is well-known. She was supposed to travel to the US for my wedding next month and I was going to see her again after 4 years
#USVisa