AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

TEMPERA: Test-Time Prompt Editing via Reinforcement Learning

Tianjun Zhang, Xuezhi Wang, Denny Zhou, Dale Schuurmans, Joseph E. Gonzalez

ICLR 2023 · Large language models › Language, safety and interpretability · approach: In-context learning analysis

Read the paper

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 changedWhereMeasuredEffect
add test-time editing (query-dependent policy) vs query-agnostic policy (ablation)TEMPERA · AG Newsaccuracy: +3.0 (84.3 vs 81.3)Improved
add test-time editing (query-dependent policy) vs query-agnostic policy (ablation)TEMPERA · SST-2accuracy: -0.1 (91.9 vs 92.0)No clear effect
add instruction and verbalizer edits to exemplar-only editing (ablation)TEMPERA · AG Newsaccuracy: +3.3 (85.5 vs 82.2)Improved
add verbalizer edits (on top of instruction edits) (ablation)TEMPERA · AG Newsaccuracy: +1.2 (85.5 vs 84.3)Improved
add instruction edits (on top of exemplar and verbalizer edits) (ablation)TEMPERA · SST-2 / MRaccuracy: +0.5 SST-2 (92.4 vs 91.9); +0.2 MR (88.4 vs 88.2)Improved
increase in-context exemplars from 4 to 8 (ablation)TEMPERA · AG Newsaccuracy: -2.1 (82.2 vs 84.3)Hurt
increase in-context exemplars from 2 to 8 (ablation)TEMPERA · SST-2 / MRaccuracy: SST-2 +0.8 (92.4 vs 91.6); MR +0.5 (88.4 vs 87.9)Improved
increase example pool size from 8 to 32 (ablation)TEMPERA · AG Newsaccuracy: +0.6 (84.7 vs 84.1)Improved
TEMPERA vs RLPrompt (discrete prompt RL) on few-shot classificationTEMPERA · SST-2 / CRaccuracy: +1.8 SST-2; +3.9 CRImproved
TEMPERA on additional NLI/paraphrase tasks (RTE, QNLI, SNLI, MNLI, MRPC) vs finetuningTEMPERA · RTE / QNLI / SNLI / MNLI / MRPCaccuracy: RTE 60.3 vs 58.6; QNLI 57.4 vs 60.2; SNLI 56.4 vs 54.64; MNLI 45.2 vs 47.8; MRPC 74.0 vs 77.4Mixed
TEMPERA with 512 training examples per class vs finetuningTEMPERA · AG News (512/class)accuracy: 88.6 vs 89.5Hurt

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy