README.md

metadata

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 Li¹, Thies Brockmoeller¹, Stefanie Jegelka^1,2, Sharut Gupta^2,◇, Nikolaos Karalias^2,◇

¹TUM ²MIT CSAIL ^◇Equal advisory role

Paper · Project page · Code

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.