Contrastive Language Models

AI Digest

文章提出对比语言模型（CLMs），通过对比学习连接状态与动作，实现高效决策，性能接近Jev但延迟低，适用于多种任务。

This paper introduces CLMs, a contrastive learning model achieving fast decision-making with performance comparable to Jev but 9x lower latency across tasks.

Key points

• CLMs通过状态-动作对比学习实现零样本动作分类，延迟降低9倍 CLMs achieve zero-shot action classification via state-action contrastive learning with 9x lower latency
• CLM-8B在复杂编码基准测试中超越Jev，推理速度提升4-6倍 CLM-8B outperforms Jev on agentic benchmarks with 4-6x faster inference
• 模型架构分离状态与动作嵌入，支持独立缓存提升推理效率 Disaggregated state/action embeddings enable independent caching for efficiency
• 建立计算量-数据量-模型规模的幂律扩展规律 Establish power-law scaling laws between compute/data/model size
• 双向InfoNCE损失结合硬负样本优化提升训练效果 Bidirectional InfoNCE loss with hard negatives enhances training

Takeaway: CLMs通过对比学习重构决策流程，实现效率与性能的平衡突破。 / CLMs redefine decision-making through contrastive learning, achieving efficiency-performance balance.

Why it matters 提供新型模型架构与训练范式，对AI系统优化有实践启发价值。