Research map / Label-efficient and robust learning / Recognition and long-tail
On the Predictive Power of Representation Dispersion in Language Models
Yanhong Li, Ming Li, Karen Livescu, Jiawei Zhou
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 |
|---|---|---|---|
| add spread-out auxiliary loss (lambda 0.1) | GPT-2 small, LR 1e-3 · WikiText single-domain, step 1000 | test perplexity: 111.3 -> 108.2 | Improved |
| add spread-out auxiliary loss (lambda 0.1) | GPT-2 small, LR 5e-4 · WikiText single-domain, step 500 | test perplexity: 166.2 -> 165.6 | No clear effect |
| add cross-domain spread-out auxiliary loss (lambda 0.01) | GPT-2 small, LR 7e-4 · WikiText+Code, Wiki test, step 1000 | test perplexity: 175.7 -> 150.2 | Improved |
| add cross-domain spread-out auxiliary loss (lambda 0.01) | GPT-2 small, LR 7e-4 · WikiText+Code, Code test, step 1000 | test perplexity: 22.8 -> 20.8 | Improved |
| add cross-domain spread-out auxiliary loss (lambda 0.001) | GPT-2 small, LR 1e-3 · WikiText+Code, Wiki test, step 500 | test perplexity: 295.6 -> 270.8 | Improved |
| add cross-domain spread-out auxiliary loss (lambda 0.02) | GPT-2 small, LR 5e-4 · WikiText+Code, Code test, step 1000 | test perplexity: 22.1 -> 20.5 | Improved |
| take key from attention sub-layer instead of FFN output | gpt2-medium · Wikitext-103 dispersion | dispersion (cosine distance): 0.21 -> 0.67 | Improved |
| take key from attention sub-layer instead of FFN output | gpt2-large · Wikitext-103 dispersion | dispersion (cosine distance): 0.68 -> 0.80 | Improved |
| increase number of sampled chunks from 10 to 100 | dispersion estimate · Wikitext-103 dispersion | dispersion (cosine distance): 0.83 -> 0.82 | No clear effect |
| use dispersion to predict accuracy instead of perplexity | Llama-3.2-1B-Instruct · MMLU slices | mean absolute error (%): 1.39 | Improved |
| increase fraction of correct answers in slice from 0% to 100% | Llama-3.2-3B · MMLU controlled slice | dispersion: 0.0923 -> 0.1525 | Improved |
| increase fraction of correct answers in slice from 0% to 100% | Llama-3.2-3B · MMLU controlled slice | perplexity: 4.751 -> 4.740 | No clear effect |
Closest papers by meaning
- Repurposing Language Models into Embedding Models: Finding the Compute-Optimal Recipe (NeurIPS 2024 · similarity 0.88)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.88)
- Analyzing Similarity Metrics for Data Selection for Language Model Pretraining (NeurIPS 2025 · similarity 0.87)
- DEPT: Decoupled Embeddings for Pre-training Language Models (ICLR 2025 · similarity 0.87)
- Retrieval-based Disentangled Representation Learning with Natural Language Supervision (ICLR 2024 · similarity 0.87)
- Test-Time Training on Nearest Neighbors for Large Language Models (ICLR 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.
