AI Paper Atlas

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

Learning syntax without semantics: Disentangled tiny language models

Ezra Winston, J Zico Kolter

ICML 2026 · Large language models › Language, safety and interpretability · approach: Faithfulness of LLM explanations

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 changedWhereMeasuredEffect
train on relexicalized grammatical nonsense (SAMBAL) instead of original textGPT-BERT 30M · BLiMP (long)accuracy: 79.2 -> 79.3No clear effect
train on SAMBAL textGPT-BERT 30M · EWoKaccuracy: 54.3 -> 50.8Hurt
train on SAMBAL textGPT-BERT 30M · SyntaxGym (all suites, long)accuracy: 70.47 -> 66.0Hurt
train on SAMBAL textGPT-BERT 30M · SyntaxGym without reflexive suitesaccuracy: 78.6 -> 79.8No clear effect
train on SAMBAL textGPT-BERT 30M · grammar-plausibility conflict setaccuracy: 10.2% -> 100%Improved
train on SAMBAL textGPT-BERT 30M · swap probe semantic plausibilityaccuracy: 100.0 -> 55.0Hurt
train 5M model on SAMBAL text at 10% data5M-parameter model · BLiMPaccuracy: 58.0 -> 60.6Improved
SAMBAL at 14M/30M models with 50-100% datalarger models · BLiMPaccuracyNo clear effect
5M SAMBAL model vs 30M baseline at 10% data5M-parameter model · BLiMPaccuracy: 59.6 -> 60.6Improved
remove corpus-conditioned candidate sampling from SAMBAL pipeline (ablation)SAMBAL-short · BLiMPaccuracy: 71.2 -> 66.42Hurt
ablate gender / humanness features in relexicalization (rather than preserve) (ablation)SAMBAL-short · BLiMPaccuracy: 71.2 -> 70.47 (gender), 69.86 (humanness)Hurt
LoRA fine-tune SAMBAL model on Lord of the RingsSAMBAL (Long) · LoTR perplexityperplexity: 190.4 -> 36.3Improved

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