Grading the Unannotated: Insights from the WorkRB Skill Extraction Challenge at RecSys-HR 2026
Warre Veys, Luis Gasco, Jens-Joris Decorte
#arXiv #cs.AI
Warre Veys, Luis Gasco, Jens-Joris Decorte
#arXiv #cs.AI
Grading the Unannotated: Insights from the WorkRB Skill Extraction Challenge at RecSys-HR 2026
Skill extraction benchmarks score models against a few human-annotated positives and treat every other skill in the vocabulary as equally wrong. At the scale of ESCO (13,939 skills) this binary view can reward models that rank the annotated positive high for the wrong reason, and hides semantically…
arxiv.org
November 3, 2026 at 8:56 PM
Testing out v2 for social media recsys where GraphSAGE previously has done well. Zero-shot is impressive but the results themselves have been a bit of a mixed bag so far. Not confident that I know how to best make use of it though!
October 30, 2026 at 1:52 AM
Testing out v2 for social media recsys where GraphSAGE previously has done well. Zero-shot is impressive but the results themselves have been a bit of a mixed bag so far. Not confident that I know how to best make use of it though!
I still vividly remember one of the first I saw do this well, at RecSys 2010 by Xavier Amatriain I think. Color-coded the core equation, and broke each piece out to explain what it did.
October 26, 2026 at 7:49 PM
I still vividly remember one of the first I saw do this well, at RecSys 2010 by Xavier Amatriain I think. Color-coded the core equation, and broke each piece out to explain what it did.
Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI
tech_blogs_facebook
#MachineLearning
tech_blogs_facebook
#MachineLearning
Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI
Last year, we introduced GEM, the industry’s largest recommendation system (RecSys) foundation model. GEM established Meta’s first LLM-inspired paradigm for ads recommendation, driving major …
engineering.fb.com
October 24, 2026 at 11:07 AM
Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI
tech_blogs_facebook
#MachineLearning
tech_blogs_facebook
#MachineLearning
RecSys this year for example
October 21, 2026 at 11:47 AM
RecSys this year for example
RecSys Challenge 2026 参加レポート：コンペ参加から論文発表まで | Komekami #zenn
https://zenn.dev/dmmdata/articles/recsys-challenge-2026-report
https://zenn.dev/dmmdata/articles/recsys-challenge-2026-report
RecSys Challenge 2026 参加レポート：コンペ参加から論文発表まで
zenn.dev
October 20, 2026 at 5:24 AM
RecSys Challenge 2026 参加レポート：コンペ参加から論文発表まで | Komekami #zenn
https://zenn.dev/dmmdata/articles/recsys-challenge-2026-report
https://zenn.dev/dmmdata/articles/recsys-challenge-2026-report
Экосистема Сбера формирует 13 млрд персональных рекомендаций в месяц для более 110 млн человек. По данным конференции AI Driver & RecSys, благодаря этим алгоритмам совершается каждая третья продажа в экосистеме.
October 21, 2026 at 5:23 AM
Экосистема Сбера формирует 13 млрд персональных рекомендаций в месяц для более 110 млн человек. По данным конференции AI Driver & RecSys, благодаря этим алгоритмам совершается каждая третья продажа в экосистеме.
Predictable Scaling (PS) lets engineers forecast full-scale recommendation model performance from small-scale proxies, using composable scaling functions and curve fitting with calibrated uncertainty. Even simple…
#OpenSourceAI #MachineLearning #RecSys #MLOps
https://arxiv.org/abs/2610.23453
#OpenSourceAI #MachineLearning #RecSys #MLOps
https://arxiv.org/abs/2610.23453
October 19, 2026 at 4:01 AM
Predictable Scaling (PS) lets engineers forecast full-scale recommendation model performance from small-scale proxies, using composable scaling functions and curve fitting with calibrated uncertainty. Even simple…
#OpenSourceAI #MachineLearning #RecSys #MLOps
https://arxiv.org/abs/2610.23453
#OpenSourceAI #MachineLearning #RecSys #MLOps
https://arxiv.org/abs/2610.23453
i mean we kind of have the recipe down tbh. the bigger issue for me is mostly that recsys for things like shortform is interleaved there, and i dont know how you would replace that
October 16, 2026 at 1:55 PM
i mean we kind of have the recipe down tbh. the bigger issue for me is mostly that recsys for things like shortform is interleaved there, and i dont know how you would replace that
É louco que muita gente que trabalha com o lado técnico de ML, recsys, etc (sou dessa área desde 2011) vem debatendo isso faz décadas (inclusive com pesquisadores das humanidades, e tal)
Aí surge palhaço querendo negar o problema e falar merda sem saber nem o lado técnico nem o lado político.
Aí surge palhaço querendo negar o problema e falar merda sem saber nem o lado técnico nem o lado político.
October 15, 2026 at 1:54 PM
É louco que muita gente que trabalha com o lado técnico de ML, recsys, etc (sou dessa área desde 2011) vem debatendo isso faz décadas (inclusive com pesquisadores das humanidades, e tal)
Aí surge palhaço querendo negar o problema e falar merda sem saber nem o lado técnico nem o lado político.
Aí surge palhaço querendo negar o problema e falar merda sem saber nem o lado técnico nem o lado político.
Youtube Recommendation Architecture. #BigData #Analytics #DataScience #AI #MachineLearning #RecSys #IoT #IIoT #GoLang #Python #RStats #TensorFlow #JavaScript #ReactJS #CloudComputing #Serverless #DataScientist #Linux #Programming #Coding #100DaysofCode
bit.ly/3pB8ZJm
bit.ly/3pB8ZJm
October 15, 2026 at 6:42 AM
Youtube Recommendation Architecture. #BigData #Analytics #DataScience #AI #MachineLearning #RecSys #IoT #IIoT #GoLang #Python #RStats #TensorFlow #JavaScript #ReactJS #CloudComputing #Serverless #DataScientist #Linux #Programming #Coding #100DaysofCode
bit.ly/3pB8ZJm
bit.ly/3pB8ZJm
Wondering if something similar could be done for "who to follow" algorithms, basically based on your follows, and activity the recsys would extract signal and predict similar between users and rank them. I've read (a little) that this is challenging to do on bsky (because decentralization?).
October 14, 2026 at 4:25 AM
Wondering if something similar could be done for "who to follow" algorithms, basically based on your follows, and activity the recsys would extract signal and predict similar between users and rank them. I've read (a little) that this is challenging to do on bsky (because decentralization?).
Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI | Engineering at Meta
Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI
Last year, we introduced GEM, the industry’s largest recommendation system (RecSys) foundation model. GEM established Meta’s first LLM-inspired paradigm for ads recommendation, driving major …
engineering.fb.com
October 13, 2026 at 4:14 PM
Meta’s Generative Ads Model (GEM): Scaling the Central Model and Compute Engine for Frontier Ads AI | Engineering at Meta
One thing i dont like is that recsys is global, and very clearly attackable by bots. Would be worth thinking about
October 11, 2026 at 2:35 PM
One thing i dont like is that recsys is global, and very clearly attackable by bots. Would be worth thinking about
NVIDIA recsys-examples now provides an end-to-end HSTU inference workflow with Dynamo-Triton.
PyTorch AOTI compiles HSTU models to native C++, while FlexKV stores attention states to boost efficiency. On an RTX PRO 6000 Blackwell GPU, this yields up to 5.93x lower latency.
https://bit.ly/4AT5Vws
PyTorch AOTI compiles HSTU models to native C++, while FlexKV stores attention states to boost efficiency. On an RTX PRO 6000 Blackwell GPU, this yields up to 5.93x lower latency.
https://bit.ly/4AT5Vws
Deploying an HSTU Generative Recommender with NVIDIA Dynamo-Triton | NVIDIA Technical Blog
Generative recommender (GR) systems are emerging as a powerful new approach for large-scale personalization. Instead of treating recommendation as a set of isolated retrieval, ranking…
developer.nvidia.com
October 8, 2026 at 9:00 PM
NVIDIA recsys-examples now provides an end-to-end HSTU inference workflow with Dynamo-Triton.
PyTorch AOTI compiles HSTU models to native C++, while FlexKV stores attention states to boost efficiency. On an RTX PRO 6000 Blackwell GPU, this yields up to 5.93x lower latency.
https://bit.ly/4AT5Vws
PyTorch AOTI compiles HSTU models to native C++, while FlexKV stores attention states to boost efficiency. On an RTX PRO 6000 Blackwell GPU, this yields up to 5.93x lower latency.
https://bit.ly/4AT5Vws
The TORS paper and RecSys slides on participatory design for recommender systems: md.ekstrandom.net/pubs/tors-co...
Recommending With, Not For
TORS paper on the need for participatory approaches to designing and evaluating recommender systems and other information access systems.
md.ekstrandom.net
October 5, 2026 at 6:30 PM
The TORS paper and RecSys slides on participatory design for recommender systems: md.ekstrandom.net/pubs/tors-co...
@inquirer.com if you'd like to talk with local recsys researchers about future directions for this work, hmu 😊.
October 5, 2026 at 6:32 PM
@inquirer.com if you'd like to talk with local recsys researchers about future directions for this work, hmu 😊.
This is a really interesting example of expert-mediated/expert-in-the-loop #recsys, developed in collaboration with those experts. Also relevant to one of the Qs we got on our #RecSys2026 participatory design talk! www.lenfestinstitute.org/solutions-re...
How The Philadelphia Inquirer built Scrape, an AI tool to surface hyperlocal news leads
The Philadelphia Inquirer's AI tool, Scrape, combs through hyperlocal news leads, enabling editors to report on a broader geographic area.
www.lenfestinstitute.org
October 5, 2026 at 6:30 PM
This is a really interesting example of expert-mediated/expert-in-the-loop #recsys, developed in collaboration with those experts. Also relevant to one of the Qs we got on our #RecSys2026 participatory design talk! www.lenfestinstitute.org/solutions-re...
Bye bye #RecSys2026! Thanks for a great week of research, presentations, networking, inspiration, and an overall great #recsys vibe.
Happy to be heading home and looking forward to a few less intense days, but also already looking forward to #RecSys2027.
Thanks @recsys.bsky.social!
Happy to be heading home and looking forward to a few less intense days, but also already looking forward to #RecSys2027.
Thanks @recsys.bsky.social!
October 3, 2026 at 11:09 PM
Bye bye #RecSys2026! Thanks for a great week of research, presentations, networking, inspiration, and an overall great #recsys vibe.
Happy to be heading home and looking forward to a few less intense days, but also already looking forward to #RecSys2027.
Thanks @recsys.bsky.social!
Happy to be heading home and looking forward to a few less intense days, but also already looking forward to #RecSys2027.
Thanks @recsys.bsky.social!
$NVDA NVIDIA: NVIDIA RECSYS-EXAMPLES SHIPS END-TO-END HSTU INFERENCE WITH DYNAMO-TRITON, UP TO 5.93X LOWER LATENCY AT BATCH 8
October 2, 2026 at 1:01 PM
$NVDA NVIDIA: NVIDIA RECSYS-EXAMPLES SHIPS END-TO-END HSTU INFERENCE WITH DYNAMO-TRITON, UP TO 5.93X LOWER LATENCY AT BATCH 8
The #RecSys2026 main conference closed today. Thank you, Thorsten Joachims, for the closing keynote, and congratulations to the Women in RecSys Journal Award winners.
Next year: #RecSys2027 in Honolulu, August 23 to 27. See you all then!
Next year: #RecSys2027 in Honolulu, August 23 to 27. See you all then!
October 2, 2026 at 12:37 AM
The #RecSys2026 main conference closed today. Thank you, Thorsten Joachims, for the closing keynote, and congratulations to the Women in RecSys Journal Award winners.
Next year: #RecSys2027 in Honolulu, August 23 to 27. See you all then!
Next year: #RecSys2027 in Honolulu, August 23 to 27. See you all then!
Day 2 of #RecSys2026 opened with the RecSys 20th Anniversary Presentation. The first RecSys was also held in Minnesota, in 2007.
Six paper sessions and the Day 2 posters followed.
Tomorrow is the last day of the main conference.
Program: recsys.acm.org/recsys26/rec...
Six paper sessions and the Day 2 posters followed.
Tomorrow is the last day of the main conference.
Program: recsys.acm.org/recsys26/rec...
October 1, 2026 at 7:54 AM
Day 2 of #RecSys2026 opened with the RecSys 20th Anniversary Presentation. The first RecSys was also held in Minnesota, in 2007.
Six paper sessions and the Day 2 posters followed.
Tomorrow is the last day of the main conference.
Program: recsys.acm.org/recsys26/rec...
Six paper sessions and the Day 2 posters followed.
Tomorrow is the last day of the main conference.
Program: recsys.acm.org/recsys26/rec...
Day 3 @recsys.bsky.social:
A highlight of RecSys is hearing directly from builders how YouTube , eBay, and DisneyPlus build under the hood.
Beyond typical two-tower setups, YouTube integrated an LLM-driven user embedding pipeline.
P.S It's already live in the US!
#RecSys2026
A highlight of RecSys is hearing directly from builders how YouTube , eBay, and DisneyPlus build under the hood.
Beyond typical two-tower setups, YouTube integrated an LLM-driven user embedding pipeline.
P.S It's already live in the US!
#RecSys2026
October 2, 2026 at 12:05 AM
Day 3 @recsys.bsky.social:
A highlight of RecSys is hearing directly from builders how YouTube , eBay, and DisneyPlus build under the hood.
Beyond typical two-tower setups, YouTube integrated an LLM-driven user embedding pipeline.
P.S It's already live in the US!
#RecSys2026
A highlight of RecSys is hearing directly from builders how YouTube , eBay, and DisneyPlus build under the hood.
Beyond typical two-tower setups, YouTube integrated an LLM-driven user embedding pipeline.
P.S It's already live in the US!
#RecSys2026
@recsys.bsky.social ACM Conference on Recommender Systems #RecSys2026 Main Conference Day 2: RecSys 20th Anniversary Presentation
September 30, 2026 at 3:21 PM
@recsys.bsky.social ACM Conference on Recommender Systems #RecSys2026 Main Conference Day 2: RecSys 20th Anniversary Presentation
thinking about that one PM i interviewed with who insisted i answer why, years ago, i hadnt done bayesian clustering for my recsys i build which was a cv line item that was under scrutiny.
like, relax?
like, relax?
October 1, 2026 at 2:49 AM
thinking about that one PM i interviewed with who insisted i answer why, years ago, i hadnt done bayesian clustering for my recsys i build which was a cv line item that was under scrutiny.
like, relax?
like, relax?