2309.07311.pdf
arxiv.org · 7,459 words · saved by 1 readers
N/A
S UDDEN D ROPS IN THE L OSS : S YNTAX ACQUISITION , P HASE T RANSITIONS , AND S IMPLICITY B IAS IN MLM S Angelica Chen1 Ravid Shwartz-Ziv1 Kyunghyun Cho1,2,3 Matthew L. Leavitt4 Naomi Saphra5 {angelica.chen, ravid.shwartz.ziv, kyunghyun.cho}@nyu.edu matthew@datologyai.com nsaphra@fas.harvard.edu 1 2 3…
saved by
related reading
- [2511.08579] Training Language Models to Explain Their Own Computationsarxiv.org
- 2403.09611.pdfarxiv.org
- A Survey on Sparse Autoencoders: Interpreting the Internal Mechanisms of Large Language Modelsarxiv.org
- [2210.13382] Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Taskarxiv.org
- 2409.02908arxiv.org
- 2502.11089arxiv.org
- Self-Adapting Language Modelsarxiv.org
- Sparse Autoencoders Find Highly Interpretable Features in Language Modelsarxiv.org
- 2308.03958arxiv.org
- Position: It's Time to Optimize for Self-Consistencytime-for-consistency.github.io
- 2310.01405.pdfarxiv.org
- 1810.04805arxiv.org