Research map / Label-efficient and robust learning / Recognition and long-tail
Connecting Pre-trained Language Model and Downstream Task via Properties of Representation
Chenwei Wu, Holden Lee, Rong Ge
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| exclude top-10000 words instead of none from the bulk partition function (ablation) | anchor vector fit on GPT-2 XL · WikiText-2 | MSE of log bulk partition function: 0.7716 -> 0.0019 | Improved |
| exclude top-10000 words instead of none (ablation) | anchor vector fit on OPT-1.3B · WikiText-2 | MSE of log bulk partition function: 0.5234 -> 0.0006 | Improved |
| exclude top-10000 words instead of none (ablation) | anchor vector fit on GPT-2 Large · WikiText-2 | MSE of log bulk partition function: 0.7296 -> 0.0022 | Improved |
| exclude top-10000 words instead of none (ablation) | anchor vector fit on OPT-125M · WikiText-2 | MSE of log bulk partition function: 0.6688 -> 0.0006 | Improved |
| exclude top-10000 words instead of none (ablation) | anchor vector fit on GPT-2 Medium · WikiText-2 | MSE of log bulk partition function: 0.7647 -> 0.2396 | Improved |
| exclude top-10000 words instead of none (ablation) | anchor vector fit on OPT-350M · WikiText-2 | MSE of log bulk partition function: 2.2449 -> 0.4275 | Improved |
| exclude top-10000 words instead of none (ablation) | anchor vector fit on GPT-2 (base) · WikiText-2 | MSE of log bulk partition function: 625.8 -> 727.4 | Hurt |
| fit linear anchor to full (original) partition function instead of bulk partition function (ablation) | anchor vector hypothesis · WikiText-2 | linear approximation error | Hurt |
| assume uniform word embeddings on a sphere (Arora et al.) for partition stability (ablation) | large pre-trained LMs · GPT-2 and OPT embeddings | uniformity of embeddings | Hurt |
Closest papers by meaning
- On Linear Representations and Pretraining Data Frequency in Language Models (ICLR 2025 · similarity 0.88)
- Language models scale reliably with over-training and on downstream tasks (ICLR 2025 · similarity 0.87)
- Junk DNA Hypothesis: Pruning Small Pre-Trained Weights $\textit{Irreversibly}$ and $\textit{Monotonically}$ Impairs ``Difficult" Downstream Tasks in LLMs (ICML 2024 · similarity 0.87)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.87)
- Understanding the Transferability of Representations via Task-Relatedness (NeurIPS 2024 · similarity 0.87)
- Provable Multi-Task Representation Learning by Two-Layer ReLU Neural Networks (ICML 2024 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
