compositionality

论文提出开源 RAG 受控评测框架 LASAGNA，在八个基准和 5K-200K 文档语料上测试 450 种配置，发现单独有效的 RAG 技术组合后收益大幅缩水（预期 +47% 仅得 +17%），检索质量与答案忠实度的设计取舍互相冲突，且最优组合随语料规模变化。

This paper formalizes compositional structure using algorithmic information theory, defining it as shared structure that emerges when parts are considered together rather than separately, and inverts the framework to generate compositional strings and images. It also discusses implications for neural network depth and compositional data curricula for open-ended learning.

RoboCoach introduces a world-model-guided coaching framework (RIDI loop) that uses imagined failures to select targeted demonstrations and expert updates for compositional robot skills, boosting Franka success from 13.3% to 75.0% with only 150 extra subtask demos. Imagined and deployed success correlate strongly (rho=0.840) across simulation and real-robot platforms.

The paper introduces the Differentiable Fuzzy Inference Layer (DFIL), a novel prediction head for large language models that enables monotonic and compositional ordinal reasoning without requiring compositional training data.

This paper introduces a concept-centric benchmark to probe LLMs' understanding of geo-spatial concepts like direction, distance, and topology, testing abstraction, compositionality, and grounding across various model architectures and scales. Findings reveal clear limitations in current LLMs' conceptual understanding.

The authors use an iterated learning model to study how meaning frequency affects the emergence of compositionality, expressivity, and stability in evolved languages. They find that high-frequency whole meanings escape grammatical pressure, but frequency applied to sub-meaning parts leads to transmission failure.

This paper argues that concepts in LLMs should be treated as a design axis, mapping the design space along pipeline stage and grounding dimension, and proposes moving from recovering concepts post-hoc to designing explicit conceptual representations.

This paper proposes representing multimodal data (image, video, text) as bags of atomic propositions (e.g., 'person holding cup'), unified via a global semantic codebook, enabling interpretable, compositional, and cross-modal understanding. The framework is demonstrated on autonomous driving and open-world data.

This paper argues that recent claims that neural networks have solved Fodor and Pylyshyn's systematicity challenge are premature. The authors show that the meta-learning for compositionality model fails to generalize out-of-distribution and behaves unsystematically even on in-distribution problems, concluding the challenge remains unmet.

This paper investigates principles of concept representation in sentence encoders through the lens of compositional semantics, identifying four key principles: fine-tuning recalibrates latent geometry, semantic signal concentrates in the final layer, hard negatives improve discrimination but not ranking, and supervision effectiveness depends on composition type.

Dwarkesh Patel tweets about Sergey Levine's argument that emergent capabilities in LLMs arise from compositionality, not just from training data.

This paper reframes model collapse in LLMs as a cultural transmission phenomenon, showing that iterated learning theory predicts a non-monotonic trajectory of compositionality under self-training, confirmed across multiple languages and models.

This paper proposes a Polar Probe that linearly recovers semantic structures from LLM activations by representing entity relations through distance and direction in a learned subspace. Testing across arithmetic, visual scenes, family trees, metro maps, and social interactions shows the code emerges in middle layers, generalizes to new entities, and causally influences model predictions.