#neurips — Public Fediverse posts

Can language models monitor and steer their own internal activations? A neuroscience-inspired neurofeedback paradigm finds yes, but only within a low-dimensional metacognitive space: semantically interpretable directions are accessible, raw-variance directions aren't. The prerequisite for spoofing activation-based oversight already partially exists.

https://benjaminhan.net/posts/20260526-metacognitive-monitoring-control-activations/?utm_source=mastodon&utm_medium=social

#Paper #Metacognition #LLMs #AISafety #Neuroscience #NeurIPS #AI

Do current LLMs know when to say "I don't know"? AbstentionBench (NeurIPS '25) tests 20 frontier models across 20 unanswerable-question datasets. Reasoning fine-tuning degrades abstention recall by ~24% — RLVR has no "abstain" action, so there's no gradient toward "I don't know." Models hedge in CoT and commit anyway in the final answer.

https://benjaminhan.net/posts/20260523-abstentionbench-unanswerable-questions/?utm_source=mastodon&utm_medium=social

#Paper #AI #LLMs #Metacognition #Benchmark #Reasoning #NeurIPS

Do current LLMs know when to say "I don't know"? AbstentionBench (NeurIPS '25) tests 20 frontier models across 20 unanswerable-question datasets. Reasoning fine-tuning degrades abstention recall by ~24% — RLVR has no "abstain" action, so there's no gradient toward "I don't know." Models hedge in CoT and commit anyway in the final answer.

https://benjaminhan.net/posts/20260523-abstentionbench-unanswerable-questions/?utm_source=mastodon&utm_medium=social

#Paper #AI #LLMs #Metacognition #Benchmark #Reasoning #NeurIPS

Do current LLMs know when to say "I don't know"? AbstentionBench (NeurIPS '25) tests 20 frontier models across 20 unanswerable-question datasets. Reasoning fine-tuning degrades abstention recall by ~24% — RLVR has no "abstain" action, so there's no gradient toward "I don't know." Models hedge in CoT and commit anyway in the final answer.

https://benjaminhan.net/posts/20260523-abstentionbench-unanswerable-questions/?utm_source=mastodon&utm_medium=social

#Paper #AI #LLMs #Metacognition #Benchmark #Reasoning #NeurIPS

Do current LLMs know when to say "I don't know"? AbstentionBench (NeurIPS '25) tests 20 frontier models across 20 unanswerable-question datasets. Reasoning fine-tuning degrades abstention recall by ~24% — RLVR has no "abstain" action, so there's no gradient toward "I don't know." Models hedge in CoT and commit anyway in the final answer.

https://benjaminhan.net/posts/20260523-abstentionbench-unanswerable-questions/?utm_source=mastodon&utm_medium=social

#Paper #AI #LLMs #Metacognition #Benchmark #Reasoning #NeurIPS

Do current LLMs know when to say "I don't know"? AbstentionBench (NeurIPS '25) tests 20 frontier models across 20 unanswerable-question datasets. Reasoning fine-tuning degrades abstention recall by ~24% — RLVR has no "abstain" action, so there's no gradient toward "I don't know." Models hedge in CoT and commit anyway in the final answer.

https://benjaminhan.net/posts/20260523-abstentionbench-unanswerable-questions/?utm_source=mastodon&utm_medium=social

#Paper #AI #LLMs #Metacognition #Benchmark #Reasoning #NeurIPS

經濟學人: 中國包攬全球 逾半頂尖 AI 研究員 後年人才規模或達美國兩倍
英國《經濟學人》聯同智庫 MacroPolo 針對 2025 年 NeurIPS （神經資訊處理系統大會）逾 […]
#人工智能 #AI 人才 #DeepSeek #NeurIPS
https://unwire.hk/2026/05/18/china-ai-talent-neurips-2025-surpass-us/ai/?utm_source=rss&utm_medium=rss&utm_campaign=china-ai-talent-neurips-2025-surpass-us

經濟學人: 中國包攬全球 逾半頂尖 AI 研究員 後年人才規模或達美國兩倍
英國《經濟學人》聯同智庫 MacroPolo 針對 2025 年 NeurIPS （神經資訊處理系統大會）逾 […]
#人工智能 #AI 人才 #DeepSeek #NeurIPS
https://unwire.hk/2026/05/18/china-ai-talent-neurips-2025-surpass-us/ai/?utm_source=rss&utm_medium=rss&utm_campaign=china-ai-talent-neurips-2025-surpass-us

經濟學人: 中國包攬全球 逾半頂尖 AI 研究員 後年人才規模或達美國兩倍
英國《經濟學人》聯同智庫 MacroPolo 針對 2025 年 NeurIPS （神經資訊處理系統大會）逾 […]
#人工智能 #AI 人才 #DeepSeek #NeurIPS
https://unwire.hk/2026/05/18/china-ai-talent-neurips-2025-surpass-us/ai/?utm_source=rss&utm_medium=rss&utm_campaign=china-ai-talent-neurips-2025-surpass-us

AVeriTeC (NeurIPS 2023): 4,568 real-world fact-checked claims, web-retrieved evidence, four-way labels, temporal-leak-free split.

Two structural gaps: gold answers are frozen but the retrieval surface isn't (two systems a year apart hit different Google), and the not-enough-evidence class rewards weak retrievers — predicting NEI when retrieval fails matches gold by coincidence.

https://benjaminhan.net/posts/20260507-averitec/?utm_source=mastodon&utm_medium=social

#Paper #Benchmark #FactVerification #NeurIPS #AI

AVeriTeC (NeurIPS 2023): 4,568 real-world fact-checked claims, web-retrieved evidence, four-way labels, temporal-leak-free split.

Two structural gaps: gold answers are frozen but the retrieval surface isn't (two systems a year apart hit different Google), and the not-enough-evidence class rewards weak retrievers — predicting NEI when retrieval fails matches gold by coincidence.

https://benjaminhan.net/posts/20260507-averitec/?utm_source=mastodon&utm_medium=social

#Paper #Benchmark #FactVerification #NeurIPS #AI

AVeriTeC (NeurIPS 2023): 4,568 real-world fact-checked claims, web-retrieved evidence, four-way labels, temporal-leak-free split.

Two structural gaps: gold answers are frozen but the retrieval surface isn't (two systems a year apart hit different Google), and the not-enough-evidence class rewards weak retrievers — predicting NEI when retrieval fails matches gold by coincidence.

https://benjaminhan.net/posts/20260507-averitec/?utm_source=mastodon&utm_medium=social

#Paper #Benchmark #FactVerification #NeurIPS #AI

Муравьи против трансформеров: старый алгоритм 1992 года, который вернулся

Начну с признания: истории формата «природа оптимизирует лучше людей» меня обычно раздражают, слишком уж часто это все притянуто за уши. Но с муравьями история действительно странная, и мне ее захотелось проверить. Короткая справка по нашему герою. Аргентинский муравей Linepithema humile в миллиметр длиной, с глазами у него все плохо, а в мозге около 250 000 нейронов (у нас, напомню, 86 млрд). Карты местности он не помнит. В 1989 году четверо бельгийских биологов поставили этим муравьям простой эксперимент — гнездо, еда, два мостика, где один длиннее другого в два раза. Через несколько минут вся колония сошлась на короткой ветке в 100% прогонов. И все это без координатора, без плана и без голосования. Через три года этот эксперимент превратится в Ant Colony Optimization — алгоритм, который я сегодня натравлю на классический TSP-бенч и получу 0,10% отставания от оптимума. А в 2023, через 34 года после наблюдений в Брюсселе, тот же алгоритм вернулся на NeurIPS в качестве бэкбона для графовых нейросетей. Что же, приступим.

https://habr.com/ru/companies/selectel/articles/1031574/

#ACO #задача_коммивояжера #DeepACO #комбинаторная_оптимизация #NeurIPS #графовые_нейросети #стигмергия #selectel #ai #ml

MetaSPO meta-learns a task-agnostic system prompt via a bilevel loop: outer tunes system prompt across tasks, inner tunes per-task user prompts. Generalizes to 14 unseen tasks across 5 domains.

The decomposition is the contribution. Once prompts split into task-agnostic (system) vs task-specific (user), meta-learning follows. System-prompt optimization transfers; user-prompt doesn't.

https://benjaminhan.net/posts/20260430-metaspo-system-prompt-meta-learning/?utm_source=mastodon&utm_medium=social

#LLMs #AI #PromptEngineering #KAIST #NeurIPS

MetaSPO meta-learns a task-agnostic system prompt via a bilevel loop: outer tunes system prompt across tasks, inner tunes per-task user prompts. Generalizes to 14 unseen tasks across 5 domains.

The decomposition is the contribution. Once prompts split into task-agnostic (system) vs task-specific (user), meta-learning follows. System-prompt optimization transfers; user-prompt doesn't.

https://benjaminhan.net/posts/20260430-metaspo-system-prompt-meta-learning/?utm_source=mastodon&utm_medium=social

#LLMs #AI #PromptEngineering #KAIST #NeurIPS

MetaSPO meta-learns a task-agnostic system prompt via a bilevel loop: outer tunes system prompt across tasks, inner tunes per-task user prompts. Generalizes to 14 unseen tasks across 5 domains.

The decomposition is the contribution. Once prompts split into task-agnostic (system) vs task-specific (user), meta-learning follows. System-prompt optimization transfers; user-prompt doesn't.

https://benjaminhan.net/posts/20260430-metaspo-system-prompt-meta-learning/?utm_source=mastodon&utm_medium=social

#LLMs #AI #PromptEngineering #KAIST #NeurIPS

MetaSPO meta-learns a task-agnostic system prompt via a bilevel loop: outer tunes system prompt across tasks, inner tunes per-task user prompts. Generalizes to 14 unseen tasks across 5 domains.

The decomposition is the contribution. Once prompts split into task-agnostic (system) vs task-specific (user), meta-learning follows. System-prompt optimization transfers; user-prompt doesn't.

https://benjaminhan.net/posts/20260430-metaspo-system-prompt-meta-learning/?utm_source=mastodon&utm_medium=social

#LLMs #AI #PromptEngineering #KAIST #NeurIPS

MetaSPO meta-learns a task-agnostic system prompt via a bilevel loop: outer tunes system prompt across tasks, inner tunes per-task user prompts. Generalizes to 14 unseen tasks across 5 domains.

The decomposition is the contribution. Once prompts split into task-agnostic (system) vs task-specific (user), meta-learning follows. System-prompt optimization transfers; user-prompt doesn't.

https://benjaminhan.net/posts/20260430-metaspo-system-prompt-meta-learning/?utm_source=mastodon&utm_medium=social

#LLMs #AI #PromptEngineering #KAIST #NeurIPS

Nature: Boycott of major AI conference exposes a growing US–China divide . “A key Chinese research organization is set to boycott the prestigious NeurIPS conference, which is run by a US-based non-profit organization, after a row over a policy that initially seemed to exclude many Chinese researchers.”

Nature: Boycott of major AI conference exposes a growing US–China divide . “A key Chinese research organization is set to boycott the prestigious NeurIPS conference, which is run by a US-based non-profit organization, after a row over a policy that initially seemed to exclude many Chinese researchers.”

Nature: Boycott of major AI conference exposes a growing US–China divide . “A key Chinese research organization is set to boycott the prestigious NeurIPS conference, which is run by a US-based non-profit organization, after a row over a policy that initially seemed to exclude many Chinese researchers.”

Nature: Boycott of major AI conference exposes a growing US–China divide . “A key Chinese research organization is set to boycott the prestigious NeurIPS conference, which is run by a US-based non-profit organization, after a row over a policy that initially seemed to exclude many Chinese researchers.”

Nature: Boycott of major AI conference exposes a growing US–China divide . “A key Chinese research organization is set to boycott the prestigious NeurIPS conference, which is run by a US-based non-profit organization, after a row over a policy that initially seemed to exclude many Chinese researchers.”

China's largest tech federation has urged researchers to boycott the NeurIPS AI conference after it stopped accepting papers from US-sanctioned entities including Huawei. The China Computer Federation called the ban a violation of core academic values. NeurIPS, often the world's premier AI conference, drew tens of thousands of researchers last year. https://www.scmp.com/tech/article/3348006/ai-rift-widens-china-urges-boycott-top-us-conference-over-sanctions-ban #China #Tech #AI #NeurIPS

#AIResearch Is Getting Harder to Separate From #Geopolitics

A policy change announced by #NeurIPS ,the world’s leading #AI #research conference, drew widespread backlash from #Chinese researchers this week and then was quickly reversed.
#china #artificialintelligence #politics

https://www.wired.com/story/made-in-china-ai-research-is-starting-to-split-along-geopolitical-lines/

#NeurIPS paper by Merlijn Krale, Eline Bovy, Maris Galesloot, Thiago Simão, and Nils Jansen: On Evaluating Policies for Robust Partially Observable Markov Decision Processes #POMDPs

https://repository.ubn.ru.nl/bitstream/handle/2066/324616/324616.pdf

#TIL about #GPTZero that has built a hallucination check tool which found 100 hallucinated references in #NeurIPS 2025 accepted papers https://gptzero.me/news/neurips/

... and NeurIPS organizers just brushed it off! (sorry for the substack link) https://stevensalzberg.substack.com/p/an-embarrassing-scandal-for-ai-research

#AI #AIslop

Google’s new PaperBanana system strings together five AI agents to automatically draft scientific diagrams—right down to Matplotlib‑style plots. It works well, but still drops missing icons, a reminder that AI‑generated figures need human polish. See how this open‑source‑friendly approach performed at NeurIPS and what it means for academic publishing. #PaperBanana #AIAgents #DiagramGeneration #NeurIPS

🔗 https://aidailypost.com/news/googles-paperbanana-uses-five-ai-agents-auto-generate-diagrams

Công trình đạt giải Best Paper tại NeurIPS bởi Kevin Wang (Princeton) giới thiệu mạng thần kinh 1000 lớp ứng dụng trong học tăng cường tự giám sát. Kỹ thuật mới cải thiện khả năng học biểu diễn sâu mà không cần dữ liệu nhãn, mở đường cho hệ thống AI hiệu quả và tự chủ hơn. #NeurIPS #AI #MachineLearning #HọcTăngCường #TríTuệNhânTạo #SelfSupervisedLearning #DeepLearning

https://www.reddit.com/r/singularity/comments/1q593by/neurips_best_paper_1000_layer_networks_for/

🧠 Un paper interessantissimo dal #NeurIPS 2025 Best Paper Awards.

💡 Spiega perché i diffusion models, pur essendo enormi e sovraparametrizzati, non memorizzano subito i dati di training.
👉 I dettagli: https://www.linkedin.com/posts/alessiopomaro_neurips-ai-genai-activity-7412027678138023936-wxYV

___
✉️ 𝗦𝗲 𝘃𝘂𝗼𝗶 𝗿𝗶𝗺𝗮𝗻𝗲𝗿𝗲 𝗮𝗴𝗴𝗶𝗼𝗿𝗻𝗮𝘁𝗼/𝗮 𝘀𝘂 𝗾𝘂𝗲𝘀𝘁𝗲 𝘁𝗲𝗺𝗮𝘁𝗶𝗰𝗵𝗲, 𝗶𝘀𝗰𝗿𝗶𝘃𝗶𝘁𝗶 𝗮𝗹𝗹𝗮 𝗺𝗶𝗮 𝗻𝗲𝘄𝘀𝗹𝗲𝘁𝘁𝗲𝗿: https://bit.ly/newsletter-alessiopomaro

#AI #GenAI #GenerativeAI #IntelligenzaArtificiale #LLM 

Apfeltalk: Der Tag am 22. November 2025
Heute stehen mehrere bedeutende Entwicklungen im Apple-Ökosystem im Fokus, insbesondere im Bereich Künstliche Intelligenz und neue Produktankündigungen. Zudem gibt es Neuigkeiten zu Apples Prozessorarchitektur und den bevorstehenden Black Friday-Angeboten.

Apple
https://www.apfeltalk.de/magazin/news/apfeltalk-der-tag-am-22-november-2025/
#News #Apple #AppleTV #BlackFriday #IOS262 #iPhone #KI #M4Max #MacBook #neurips #Produkte2026 #Studien

Feeling rewarded... Honored to be among the Top Reviewers at #NeurIPS this year.

#AI #ML #UAntwerp #sqIRL #IDLab #imec

Looking to follow (and be followed!) #mlbio #bioml #biotech #bayes #bayesian #generativemodels #startup #startups #medicine #biology #representationlearning #techbio #immunology #neurips #icml #iclr #cambridge #nyc

https://winbuzzer.com/2025/11/08/google-research-unveils-nested-learning-a-new-ai-paradigm-to-overcome-catastrophic-forgetting-xcxwbn

AI Memory: Google Research Unveils Nested Learning, a New AI Paradigm to Overcome 'Catastrophic Forgetting'

#AI #MachineLearning #GoogleAI #DeepLearning #DeepMind #NeuralNetworks #ContinualLearning #Google #AIResearch #NeurIPS

https://winbuzzer.com/2025/11/08/google-research-unveils-nested-learning-a-new-ai-paradigm-to-overcome-catastrophic-forgetting-xcxwbn

AI Memory: Google Research Unveils Nested Learning, a New AI Paradigm to Overcome 'Catastrophic Forgetting'

#AI #MachineLearning #GoogleAI #DeepLearning #DeepMind #NeuralNetworks #ContinualLearning #Google #AIResearch #NeurIPS

https://winbuzzer.com/2025/11/08/google-research-unveils-nested-learning-a-new-ai-paradigm-to-overcome-catastrophic-forgetting-xcxwbn

AI Memory: Google Research Unveils Nested Learning, a New AI Paradigm to Overcome 'Catastrophic Forgetting'

#AI #MachineLearning #GoogleAI #DeepLearning #DeepMind #NeuralNetworks #ContinualLearning #Google #AIResearch #NeurIPS

https://winbuzzer.com/2025/11/08/google-research-unveils-nested-learning-a-new-ai-paradigm-to-overcome-catastrophic-forgetting-xcxwbn

AI Memory: Google Research Unveils Nested Learning, a New AI Paradigm to Overcome 'Catastrophic Forgetting'

#AI #MachineLearning #GoogleAI #DeepLearning #DeepMind #NeuralNetworks #ContinualLearning #Google #AIResearch #NeurIPS

https://winbuzzer.com/2025/11/08/google-research-unveils-nested-learning-a-new-ai-paradigm-to-overcome-catastrophic-forgetting-xcxwbn

AI Memory: Google Research Unveils Nested Learning, a New AI Paradigm to Overcome 'Catastrophic Forgetting'

#AI #MachineLearning #GoogleAI #DeepLearning #DeepMind #NeuralNetworks #ContinualLearning #Google #AIResearch #NeurIPS

I’ll be at #NeurIPS the coming days. Ping me if you wanna hang out, grab a drink, or talk #probabilistic #machinelearing.

Steffen Schotthöfer & Emanuele Zangrando from our lab are attending #NeurIPS next week in person and will present our work on #lowrank #training & #pruning of #NNs

Meet both at the poster session in HallJ#604 on Wed 30 Nov 9:30amPST

What is it? We developed a framework to perform stable and efficient training on low-rank manifolds, resulting in an order of magnitude less memory cost & training time! Tested successfully on #imagenet1k #transformers +other benchmarks

https://openreview.net/forum?id=IILJ0KWZMy9&referrer=%5Bthe%20profile%20of%20Francesco%20Tudisco%5D(%2Fprofile%3Fid%3D~Francesco_Tudisco1)

Presenting a poster in NewInML this afternoon at Sorbonne Center for Artificial Intelligence in Paris.

I am trying out the style of online poster advocated by @mikemorrison

Feedback appreciated :blobthanks:

📄 https://solalnathan.com/Baylearn_2022.pdf

#MastodonPoster #Poster #BetterPoster #NewInML #NeurIPS2022 #NeurIPS

A bittersweet retrospective by Tomas Mikolov on #word2vec after receiving the “Test Of Time” award for it at #NeurIPS:

https://www.facebook.com/1533402400/posts/yesterday-we-received-a-test-of-time-award-at-neurips-for-the-word2vec-paper-fro/10231641326366223/

Their seminal #paper was rejected four times 😥:

https://openreview.net/forum?id=idpCdOWtqXd60

Via HN: https://openreview.net/forum?id=idpCdOWtqXd60

#ML #MachineLearning

Explaining model *predictions* is all well and good – but what about model *uncertainty*?

Pleased to announce that our paper on information theoretic Shapley values will be presented at #NeurIPS2023.

Joint work with David Watson, Josh O’Hara, Richard Mudd, and Ido Guy.

https://arxiv.org/abs/2306.05724

#NeurIPS #xai #UncertaintyQuantification #machinelearning

If you go to New Orleans for #neurips, don't miss your chance to learn about New Orleans unique history and culture, with french, African and native American roots! It's also the place where all modern guitar music originated.

I can recommend the #Treme walking tour (and the HBO series, of course).

https://www.frenchquarterphantoms.com/tour-treme

I am going to share some live summaries from #EMNLP #NEURIPS #CONLL conferences

Hope I am not too spammy (will pass soon)

And tell me if I can make it more useful for people not participating in any way
#machinelearning #LLM #LLMs #NLP #NLProc #ML #CV #data

I was of three minds,
Like a paper
On which there are three reviews.

#NeurIPS #EMNLP #AAAI

It's stunning in sunny San Diego! 🌴
I'm so honored to be representing my co-authors and presenting our work for the #NeurIPS2025 Education track and then on Thursday, I'll be helping run the inaugural #NeurIPS #PositionPaper track!

Come find me:
- TUESDAY presenting #Benchmarks 101 for the Education Track: https://neurips.cc/virtual/2025/loc/san-diego/131571
- THURSDAY morning cheering on the Position Papers solo talks: https://neurips.cc/virtual/2025/loc/san-diego/session/122560
- THURSDAY afternoon moderating a Position Papers panel: https://neurips.cc/virtual/2025/loc/san-diego/session/122565

It's stunning in sunny San Diego! 🌴
I'm so honored to be representing my co-authors and presenting our work for the #NeurIPS2025 Education track and then on Thursday, I'll be helping run the inaugural #NeurIPS #PositionPaper track!

Come find me:
- TUESDAY presenting #Benchmarks 101 for the Education Track: https://neurips.cc/virtual/2025/loc/san-diego/131571
- THURSDAY morning cheering on the Position Papers solo talks: https://neurips.cc/virtual/2025/loc/san-diego/session/122560
- THURSDAY afternoon moderating a Position Papers panel: https://neurips.cc/virtual/2025/loc/san-diego/session/122565

It's stunning in sunny San Diego! 🌴
I'm so honored to be representing my co-authors and presenting our work for the #NeurIPS2025 Education track and then on Thursday, I'll be helping run the inaugural #NeurIPS #PositionPaper track!

Come find me:
- TUESDAY presenting #Benchmarks 101 for the Education Track: https://neurips.cc/virtual/2025/loc/san-diego/131571
- THURSDAY morning cheering on the Position Papers solo talks: https://neurips.cc/virtual/2025/loc/san-diego/session/122560
- THURSDAY afternoon moderating a Position Papers panel: https://neurips.cc/virtual/2025/loc/san-diego/session/122565

It's stunning in sunny San Diego! 🌴
I'm so honored to be representing my co-authors and presenting our work for the #NeurIPS2025 Education track and then on Thursday, I'll be helping run the inaugural #NeurIPS #PositionPaper track!

Come find me:
- TUESDAY presenting #Benchmarks 101 for the Education Track: https://neurips.cc/virtual/2025/loc/san-diego/131571
- THURSDAY morning cheering on the Position Papers solo talks: https://neurips.cc/virtual/2025/loc/san-diego/session/122560
- THURSDAY afternoon moderating a Position Papers panel: https://neurips.cc/virtual/2025/loc/san-diego/session/122565