AI Research @ Arm All opinions on this account are mine and not a reflection of my affiliated organizations.
San Jose, CA
Joined September 2016
- Tweets96
- Following380
- Followers77
- Likes341
Payal π©π½βπ¬βπ©π½βπ» retweeted
Excited to announce our work "Attention Drift" has been accepted to NeurIPS 2026 π
I'm so happy to see our method's widespread adoptation. Now supported by all major inference engines, vLLM, Tokenspeed, SGLang and TRT-LLM. Also community trained speculative decoding models for Minimax M3 (@inferact and @AMD) and Kimi K2.6-K3 trained (@lightseekorg) being served in production as EAGLE 3.1
Shout out to all the great team mebers @themishtimoi @yigitcancomlek @kaan_oktay @hongyangzh and Stephen Xia.
And special thanks to @fal for helping for making this research possible π
Payal π©π½βπ¬βπ©π½βπ» retweeted
As TorchSpec team, we've been working on accelerating Kimi-K3 since its release. Now happy to share the SOTA speculative decoding models (DFlash2, EAGLE-3, DSpark) for K3.
P.s. we've observed high acceptance != faster, each architecture has their own drawbacks! π§΅
Releasing the @Kimi_Moonshot K3 Draft Collection β 3 draft models (EAGLE-3, DFlash2, DSpark) trained with TorchSpec and @vllm_project on @NVIDIAAI GB200. π
We also shared the data recipes. More on the blog β lightseek.org/blog/kimi-k3-dβ¦
Payal π©π½βπ¬βπ©π½βπ» retweeted
π§΅We discovered a new phenomenon in speculative decoding models that we call attention drift.
As the drafter generates tokens, its attention moves from the "sink" onto its own recently-generated tokens.
Fixing the underlying issue recovered up to 2Γ acceptance length, but why?
Missing #ICLR2026 but our time-series interpretability paper is in good hands, Prof. @zhuqical is presenting tomorrow!
TL;DR: linear composition of symbolic + latent reps β temporal attributions without hurting accuracy.
π Thu 10:30aβ1p BRT
π P3-#1320
iclr.cc/virtual/2026/poster/β¦
The best start to a Friday with my favorite (and first) collaborator aka friend, is getting feedback from the youngest minds in Computer Engineering ;)
#joysofresearch β at Evanston, IL
Excited to be at #NeurIPS2025 in San Diego till December 8. I am presenting two spotlight papers on practical ML for time-series! More details on the papers soon, but if you want to catch up/hunt for fun mixers/find good coffee/run along the harbor, hit me up!
It was a pleasure speaking at the IEEE @RTC_Conf on robust ML for real-world time-series! You can find slides and video link β¬οΈ
Well I have always had trouble getting on-board with papers on how general-purpose LLMs can be used for zero-shot time-series forecasting, time-series classification or really anything requiring non-trivial analyses of numerical data!
Still find this resource insightful -- beren.io/2023-02-04-Integer-β¦ [1/2]
We demonstrate that, under closed-vocabulary conditions, our approach using LLMs (LLaMA 2 and LLaMA 3) achieves an average word error rate (WER) of 0.49, and as low as 0.39. [2/n]
π Paper : aclanthology.org/2025.acl-shβ¦
π» Codebase & processed dataset: github.com/payalmohapatra/Siβ¦
We invite you to explore our work and would love to hear your thoughts on incorporating linguistic biosignals as a modality for LLMs. [n/n]
Collaborating with @Boeing since 2021 (my maiden PhD project). From modeling an ambi(tious)guous target, fatigue, to ergonomic assessment with commercial wearables (coming soon...)--this has been the most fun interdisciplinary playground to grow and learn!
#wearablesandAI β at North Charleston, SC
This resonates well with me, a fellow interdisciplinary researcher, who has been the recipient of one too many dogmatic encounters with CS researchers claiming to be fundamental-scientists.
Payal π©π½βπ¬βπ©π½βπ» retweeted
Aiming to improve occupational health among manufacturing workers, Ping Guo and Qi Zhu led a team of researchers in the development of a wearable multimodal sensor system that leverages ML to enable near-real-time fatigue prediction on the factory floor. spr.ly/6010SPRWQ
Our work on fatigue monitoring using a lightweight ML framework and on-body sensor network, getting some limelight!
A system that measures heart rate, skin temperature, and locomotive signs continuously monitors manufacturing workers for signs of fatigue, helping prevent mistakes, injuries, and the development of chronic conditions. In PNAS Nexus: ow.ly/vS5l50TPPpZ
ALT In this image, Vasudev Aravind, co-first author of the article, is shown performing the Task Harnessing. He is equipped with six wearable body sensors, a weighted vest, and wrist weights. During data collection, Vasudev completes five repetitions of the task segment, which involves applying zip ties to predefined zones, which simulates the conditions of a work shift.