Indie researcher needs arXiv endorsement for TensionLM (new attention mechanism)
Posted by Stock_Palpitation442@reddit | LocalLLaMA | View on Reddit | 3 comments
Hey everyone,
I'm an independent researcher who just submitted "TensionLM: Sigmoid Tension as Constraint Relaxation for Language Modelling" to arXiv (cs.AI + cs.LG).
- 117M model with full public code + weights
- Sigmoid tension attention (replaces softmax) + TS-native auxiliary losses
- Logic → Language → Math curriculum gives 96× better first-contact math PPL
Would really appreciate an endorsement so the paper can go live.
Endorsement code: **RLU86M**
Paper + code: https://github.com/BoggersTheFish/bozo
Model: https://huggingface.co/BoggersTheFish/TensionLM-117M-Curriculum
Thanks so much!
NoFaithlessness951@reddit
I just see someone who talked to Claude a little too much
Stock_Palpitation442@reddit (OP)
Its a different kind of theory - kind of hard to talk to anyone else about it when no one i have contact with is in the space, the models are on hugging face and you can see the results - or dont but i would much prefer any interaction with actual people than these damn bots😭
NoFaithlessness951@reddit
Find some online communities with people interested in the same things you are, discord etc.