Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning

Consistency of Compositional Generalization Across Multiple Levels

AAAI 2025 · Label-efficient and robust learning › Zero-/few-shot and meta-learning · approach: Normalization and unsupervised meta-learning · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes a meta-learning framework (multilevel optimization with per-complexity meta-weight-nets) that learns compositions from simple to complex to improve consistency of compositional generalization across phrase-phrase, phrase-word and word-word levels, and releases the GQA-CCG dataset.

Problem

Existing VQA and video grounding models generalize to complex novel compositions but do not maintain consistency across the simpler phrase-word and word-word compositions derived from them.

Approach

Split training data into validation sets by compositional complexity (longest phrase length), attach a meta-weight-net per set, and optimize them sequentially in a multilevel optimization loop so the model fits simple compositions before complex ones.

Key results

• CFR+MLO reaches 74.23% overall accuracy and 49.27% consistency on GQA-CCG, improving over CFR's 46.46% consistency.
• MAC+MLO improves consistency from 30.82% to 34.10% on GQA-CCG, a 3.28-point gain.
• Large VLMs (7B to 37B) reach at most about 40.58% consistency (emu2, 37B), indicating consistency remains hard for them.

Measured effects reported in the paper

Surprising findings

• Learning validation sets from complex to simple gives almost no improvement over the baseline, with a 0.14-point drop in phrase-phrase accuracy for MAC.
• Pretrained LVLMs with 7B to 37B parameters trail small task-specific models such as CFR+MLO on consistency despite far more training data.

Limitations

• Compositional complexity is approximated only by the length of the longest phrase in a question.
• Improvement on IID GQA is marginal (e.g., 0.4 points for MAC).
• GQA-CCG relies on GPT-3.5 generation with manual review and is limited to VQA questions.

Experimental setup

Datasets: GQA GQA-CCG Charades-CG

Benchmarks: GQA-CCG GQA Charades-CG

Metrics: accuracy consistency score R1@0.5 R5@0.5

Baselines: MAC LCGN MMN VL-T5 CFR MS-2D-TAN LXMERT OpenFlamingo BLIP-2 Otter LLaVA-1.5 CogVLM

Builds on: meta-weight-net implicit function theorem bi-level optimization benepar

Closest papers by meaning

• Measuring Compositional Generalization: A Comprehensive Method on Realistic Data (ICLR 2020 · similarity 0.91)
• Compositional Generalization in Unsupervised Compositional Representation Learning: A Study on Disentanglement and Emergent Language (NeurIPS 2022 · similarity 0.88)
• Meta Compositional Referring Expression Segmentation (CVPR 2023 · similarity 0.87)
• When does compositional structure yield compositional generalization? A kernel theory. (ICLR 2025 · similarity 0.87)
• Scalable Evaluation and Neural Models for Compositional Generalization (NeurIPS 2025 · similarity 0.86)
• Compositional Generalization for Multi-Label Text Classification: A Data-Augmentation Approach (AAAI 2024 · similarity 0.86)