Pinned Tweet
Weβre excited to introduce TiRex β a pre-trained time series forecasting model based on an xLSTM architecture.
Andreas Auer retweeted
π Excited to share our new paper on scaling laws for xLSTMs vs. Transformers.
Key result: xLSTM models Pareto-dominate Transformers in cross-entropy loss.
- At fixed FLOP budgets β xLSTMs perform better
- At fixed validation loss β xLSTMs need fewer FLOPs
π§΅ Details in thread
Andreas Auer retweeted
Ever wondered how linear RNNs like #mLSTM (#xLSTM) or #Mamba can be extended to multiple dimensions?
Check out "pLSTM: parallelizable Linear Source Transition Mark networks". #pLSTM works on sequences, images, (directed acyclic) graphs.
Paper link: arxiv.org/abs/2506.11997
We also introduce Contiguous Patch Masking, a masking strategy for training that facilitates TiRexβs ability to generate coherent quantile estimates in long-horizon settings, without sacrificing short-term accuracy.
1
2
410
Check out the paper and the model here:
π Paper: arxiv.org/abs/2505.23719
π€ Huggingface: huggingface.co/NX-AI/TiRex
π§βπ» Github: github.com/NX-AI/tirex
Joint work with Patrick Podest, @ido87 , Sebastian BΓΆck, @gklambauer , @HochreiterSepp , and the great team of @nx_ai_com .
1
12
411
Andreas Auer retweeted
Hey there π
Want to do a Phd with me? Apply: it-u.at/en/study-program/docβ¦
Andreas Auer retweeted
Introducing Vision-LSTM - making xLSTM read images π§ It works ... pretty, pretty well ππ But convince yourself :) We are happy to share code already!
π: arxiv.org/abs/2406.04303
π₯οΈ: nx-ai.github.io/vision-lstm/
All credits to my stellar PhD @benediktalkin
Andreas Auer retweeted
Really excited to announce Energy-based Hopfield Boosting. π
Hopfield Boosting advances the state-of-the-art in OOD detection. π
Our new energy function allows a model to focus on hard training examples close to the decision boundary between in-distribution and OOD samples. π§΅
Andreas Auer retweeted
The #xLSTM is finally live! What an exciting day!
How far do we get in language modeling with the LSTM compared to State-of-the-Art LLMs?
I would say pretty, pretty far!
How? We extend the LSTM with Exponential Gating and parallelizable Matrix Memory!
arxiv.org/pdf/2405.04517
Andreas Auer retweeted
I am so excited that xLSTM is out. LSTM is close to my heart - for more than 30 years now. With xLSTM we close the gap to existing state-of-the-art LLMs. With NXAI we have started to build our own European LLMs. I am very proud of my team. arxiv.org/abs/2405.04517
Andreas Auer retweeted
xLSTM is out -- putting LSTM networks on steroids to become a more than serious LLM competitor. How? Via exponential gating and enhanced (cell state) memory capacities. Does it work? Oh, yeah ππ
arxiv.org/abs/2405.04517
Andreas Auer retweeted
Interested in a semantic memory for reinforcement learning?
I was recently invited to a podcast talking about our #NeurIPS2023 paper: Semantic HELM (arxiv.org/abs/2306.09312).
In case you are interested, you can stream the episode here:
open.spotify.com/episode/4n2β¦
Andreas Auer retweeted
Our latest paper challenges the status quo in Unsupervised Learning for Combinatorial Optimization (CO) problems.π‘
π Unveiling the limitations of popular Mean Field methods, we showcase the power of Autoregressive Approximations of solving CO Problems on graphs. π§΅
Andreas Auer retweeted
Time series come with tradition at @HochreiterSepp's institute - @AndAuer puts a new perspective on conformal predictions for time series data - great work to appear @NeurIPSConf !
Interested in reliable prediction intervals for time series? π
Excited to announce our latest work, which proposes HopCPT, was accepted at #Neurips2023 π. HopCPT tackles the challenges of time series data for Conformal Prediction with the help of Modern Hopfield Networks.
1/4
Andreas Auer retweeted
Super paper on reliable prediction for time series at NeurIPS. Conformal Predictions adapted to time series yields excellent results.
Interested in reliable prediction intervals for time series? π
Excited to announce our latest work, which proposes HopCPT, was accepted at #Neurips2023 π. HopCPT tackles the challenges of time series data for Conformal Prediction with the help of Modern Hopfield Networks.
1/4
Interested in reliable prediction intervals for time series? π
Excited to announce our latest work, which proposes HopCPT, was accepted at #Neurips2023 π. HopCPT tackles the challenges of time series data for Conformal Prediction with the help of Modern Hopfield Networks.
1/4
Additionally, the updatable Hopfield Memory allows HopCPT to mitigate shifts in the error distribution π. The empirical evaluation on 6 real-world datasets shows that HopCPT establishes a new state of the art in Conformal Prediction for time series. π
3/4
1
6
983
Check out the paper and code here:
πΎ github.com/ml-jku/HopCPT
π arxiv.org/abs/2303.12783
Not yet convinced to read the paper? Then check out our blog post:
π° ml-jku.github.io/HopCPT
HopCPT is joint work @martingauch , @ido87 and @HochreiterSepp π
Many thanks to them
4/4
















