English(EN) Borrowed Geometry: Computational Reuse of Frozen Text-Pretrained Transformer Weights Across Modalities

Gemma 4 31B权重通过精细的可训练接口展示跨模态迁移

作者 PulseAugur 编辑部 · [2 个来源] · 2026-05-01 01:23

研究人员已经证明，来自Gemma 4 31B文本预训练模型的冻结权重可以有效地跨不同模态进行重用，包括机器人和联想回忆任务。通过采用一个精细的、可训练的接口，这些未修改的权重在机器人操作基准测试中取得了最先进的结果，并在强化学习中与Decision Transformer的性能相匹配，而可训练参数却显著减少。该研究还确定了对文本任务和跨模态应用都至关重要的特定Transformer头，这表明模型内部存在更深层次的计算重用机制。 AI

影响展示了使用冻结文本模型进行高效跨模态迁移学习的潜力，减少了新任务的训练需求。

排序理由学术论文，详细介绍了一种跨模态重用冻结Transformer权重的创新方法。

在 arXiv cs.CL 阅读 →

AI 生成摘要 · Google Gemini · 来自 2 个来源。我们如何撰写摘要 →

报道来源 [2]

arXiv cs.LG TIER_1 English(EN) · Abay Bektursun · 2026-05-04 04:00

Borrowed Geometry: Computational Reuse of Frozen Text-Pretrained Transformer Weights Across Modalities

arXiv:2605.00333v1 Announce Type: new Abstract: Frozen Gemma 4 31B weights pretrained exclusively on text tokens, unmodified, transfer across modality boundaries through a thin trainable interface. (1) OGBench scene-play-singletask-task1-v0: $+4.33$pt over published GCIQL at $n=3…
arXiv cs.CL TIER_1 English(EN) · Abay Bektursun · 2026-05-01 01:23

Borrowed Geometry: Computational Reuse of Frozen Text-Pretrained Transformer Weights Across Modalities

Frozen Gemma 4 31B weights pretrained exclusively on text tokens, unmodified, transfer across modality boundaries through a thin trainable interface. (1) OGBench scene-play-singletask-task1-v0: $+4.33$pt over published GCIQL at $n=3$ with std 0.74 -- a published-SOTA win on a rob…

报道来源 [2]

Borrowed Geometry: Computational Reuse of Frozen Text-Pretrained Transformer Weights Across Modalities

Borrowed Geometry: Computational Reuse of Frozen Text-Pretrained Transformer Weights Across Modalities

相关实体

相关话题