| pretty_name: "On the Fragility of Tasks under Composition: Implicit Budgeting in Language Models" | |
| viewer: false | |
| extra_gated_prompt: "These runs contain GPQA questions in plain text. Please do not repost them anywhere public." | |
| # On the Fragility of Tasks under Composition: Implicit Budgeting in Language Models | |
| **Chenpan Li**<sup>1</sup>, **Thies Brockmoeller**<sup>1</sup>, **Stefanie Jegelka**<sup>1,2</sup>, | |
| **Sharut Gupta**<sup>2,◇</sup>, **Nikolaos Karalias**<sup>2,◇</sup> | |
| <sup>1</sup>TUM <sup>2</sup>MIT CSAIL <sup>◇</sup>Equal advisory role | |
| [Paper](https://chenpanli.github.io/fragility/) · [Project page](https://chenpanli.github.io/fragility/) · [Code](https://github.com/chenpanli/fragility) | |
| Every dataset, batch, model run and result behind the paper. Fetch it with | |
| `bash download_runs.sh` from the code repository, then `bash reproduce_results.sh` | |
| rebuilds every figure and table without a GPU. lm-eval's raw `samples_*.jsonl` are not | |
| included; no analysis reads them. | |