#compositional generalization

7 stories · 1 in 7 days · 1 in 30 days

CoGe-GCD 1 DiT 1 FID 1 GaTaP 1 GCD 1 LiveCodeBench 1 Llama3-8B 1 Transformer 1

Stories, products and papers tagged "compositional generalization", newest first.

Nov 11

Papers Builders10:44

Chinese content研究揭示 Transformer 组合泛化评估中的等价性与模块覆盖问题

这项研究解释了为什么不同论文测组合泛化结论互相矛盾：很多基准里的模块序列其实功能等价，模型是在走捷径。做基准设计或评估模型泛化能力的开发者应该看看，结论会直接影响你怎么切训练测试集。

#Transformer #compositional generalization #Llama3-8B #LiveCodeBench

arXiv cs.LG@Purva Pruthi et al. (4)Source

Oct 8

Papers BuildersSelected10:16

ORCA Improves Compositional Alignment in Text-to-Image Diffusion Models

The reframing matters: compositional failures come from misaligned rather than missing information, meaning the fix is a cheap auxiliary loss rather than a bigger text encoder. Teams training diffusion transformers can adopt ORCA to halve training cost while boosting attribute binding and multi-object accuracy. Compared with REPA, the current popular alignment baseline, ORCA adds a prompt-dependent visual subspace selection signal at no inference overhead.

#image generation #diffusion models #DiT #compositional generalization

arXiv cs.LG@Arshia Hemmat et al. (7)Source

Oct 6

Papers09:36

Chinese contentGaTaP：通过门控目标传播实现持续学习中的组合泛化

一篇持续学习论文，提出 GaTaP 算法，用门控机制既防遗忘又能 few-shot 学新任务组合，实验覆盖 MLP 和 CNN。

#continual learning #GaTaP #few-shot learning #compositional generalization

arXiv cs.LG@Abdel Mfougouon Njupoun et al. (4)Source

Sep 10

Papers BuildersSelected73°09:26

CoGe-GCD Model for Category Discovery

CoGe-GCD can be integrated into existing GCD frameworks without modifying heads or losses, offering practical improvements for researchers working with category discovery.

#CoGe-GCD #GCD #compositional generalization #computer vision

arXiv cs.LG@Luyao Tang et al. (6)Source

Aug 21

Papers10:01

Chinese content限制证据可见性提升语言模型组合泛化能力

研究人员发现，在共享语言模型中，限制模块间的证据可见性，能让模型在组合任务上表现更好，比全局可见的模型平均高出20个百分点，而且学到的接口更可复用。

#compositional generalization #language models

arXiv cs.LG@Narcis MarincatSource

Jul 21

PapersSelected01:10

Chinese contentRLM Harness使Transformer实现组合泛化

这篇研究很酷，RLM harness让Transformer在短任务上训练就能自动泛化到长任务，甚至跨领域。不用改模型结构，靠框架就能实现。