Research map / Large language models / Language, safety and interpretability
Unveiling Downstream Performance Scaling of LLMs: A Clustering-Based Perspective
Chengyin Xu, Kaiyuan Chen, Xiao Li, Ke Shen, Chenggang Li
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 subset-to-full mapping (ablation) | COD · 70B dense, mean error | absolute prediction error (%): 2.24 -> 1.55 | Improved |
| cluster tasks by difficulty and extrapolate predictable subset instead of end-to-end extrapolation | End-to-end(exp) · 70B dense, mean error | absolute prediction error (%): 3.10 -> 1.55 | Improved |
| use COD on cross-architecture target (dense clusters to MoE) | Loss-intermediate · 32B MoE, mean error | absolute prediction error (%): 3.65 -> 3.11 | Improved |
| use COD on cross-architecture target (dense clusters to MoE) | Loss-intermediate · 32B MoE, max error | absolute prediction error (%): 7.24 -> 8.11 | Hurt |
| use Improved-MeanShift instead of plain KMeans clustering (ablation) | COD · mean final error | absolute prediction error (%): 3.06 -> 1.55 | Improved |
| use Improved-MeanShift instead of DBScan clustering (ablation) | COD · mean final error | absolute prediction error (%): 6.43 -> 1.55 | Improved |
| use Improved-MeanShift instead of Improved-KMeans (ablation) | COD · mean final error | absolute prediction error (%): 2.82 -> 1.55 | Improved |
| include random-guess term g in fitting formula (ablation) | COD · BBH final error | absolute prediction error (%): 11.65 -> 0.47 | Improved |
| include constant c in fitting formula (ablation) | COD · BBH final error | absolute prediction error (%): 4.10 -> 0.47 | Improved |
| use sigmoid-style formula instead of direct power law (ablation) | COD · BBH final error | absolute prediction error (%): 8.85 -> 0.47 | Improved |
| add thresholds on fitted a, b, c to filter predictable subsets (ablation) | COD · mean final error | absolute prediction error (%): 2.61 -> 1.55 | Improved |
| relax individual thresholds a or b (ablation) | COD · mean final error | absolute prediction error (%): 1.55 -> 1.54 | No clear effect |
Closest papers by meaning
- Fast and Accurate Probing of In-Training LLMs' Downstream Performances (NeurIPS 2026 · similarity 0.92)
- LLMs on the Line: Data Determines Loss-to-Loss Scaling Laws (ICML 2025 · similarity 0.92)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.91)
- $\texttt{Model-GLUE}$: Democratized LLM Scaling for A Large Model Zoo in the Wild (NeurIPS 2024 · similarity 0.91)
- Language models scale reliably with over-training and on downstream tasks (ICLR 2025 · similarity 0.90)
- Position: Enough of Scaling LLMs! Lets Focus on Downscaling (ICML 2025 · similarity 0.90)
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.
