AI Paper Atlas

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

How does GPT-2 compute greater-than?: Interpreting mathematical abilities in a pre-trained language model

Michael Hanna, Ollie Liu, Alexandre Variengien

NeurIPS 2023 · Large language models › Language, safety and interpretability · approach: Reasoning mechanisms and failure modes · 14 citations (OpenAlex)

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
corrupt circuit nodes while the rest of the model sees normal input (ablation)GPT-2 small · Year-span predictionprobability difference: 81.7% -> -36.6%Hurt
remove linearly extractable YY information with LEACE from attention head output (ablation)GPT-2 small · Year-span predictionprobability difference: 81.7% -> 64.7%Hurt
ablate PCA dimensions of structured number representations (ablation)GPT-2 small · Year-span predictionprobability differenceNo clear effect
patch MLP 8 direct path vs indirect paths (ablation)GPT-2 small · Year-span predictionperformance drop: 14% direct vs 39% indirectMixed
apply year-span circuit to started/ended and price-range promptsyear-span circuit · related greater-than promptsloss recovery: 98.8% and 88.9%Improved
add MLP 7 and two extra attention heads to the circuit for number-sequence prompt (ablation)year-span circuit · 1599, 1607, 1633, 1679, 17YY, 17performance recovery: 67.8% -> 90.3%Improved

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