AI Paper Atlas

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

Learning to Parse and Reconstruct: Bidirectional Modeling of Question-to-Program Mapping

Zeying Duan, Youtian Du, Yuanlin Chang, Bowen Lin, Weijia Wu

AAAI 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
add question reconstruction auxiliary loss to program parsingTransformer program parser · CLEVR val (10K train)sequence-level accuracy: 0.093 -> 0.691Improved
add question reconstruction auxiliary lossTransformer program parser · GQA subset val (30K train)sequence-level accuracy: 0.517 -> 0.771Improved
add question reconstruction with large training set (700K)Transformer program parser · CLEVR val (700K train)sequence-level accuracy: 1.000 -> 1.000No clear effect
add question reconstruction auxiliary lossTransformer program parser · GQA subset val (700K train)sequence-level accuracy: 0.818 -> 0.833Improved
add question reconstruction auxiliary lossTransformer program parser · CLEVR val (30K train)convergence speed: BiPaR reaches about 0.9 accuracy within 30 epochs; Transformer only at end of trainingImproved
set loss weight lambda to 0.7 (vs smaller or larger) (ablation)BiPaR-Full · CLEVR val (10K, 30K, 50K train)sequence-level accuracyImproved
use decoder-only instead of encoder-decoder reconstruction module (ablation)BiPaR · CLEVR and GQA subsetsequence-level accuracy: CLEVR 10K: 0.676 -> 0.691No clear effect
add 50K unlabeled questions with pseudo programs to 10K labeledBiPaR-Full · CLEVR valsequence-level accuracy: 0.676 -> 0.932Improved
add 50K unlabeled questions with pseudo programs to 10K labeledBiPaR-Full · GQA subset valsequence-level accuracy: 0.579 -> 0.671Improved
increase unlabeled questions from 5K to 50K (ablation)BiPaR-Full · CLEVR valsequence-level accuracy: 0.876 -> 0.932Improved

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