Yair Feldman

Yair Feldman, Yoav Artzi

A common strategy to reduce the computational costs of using long contexts in retrieval-augmented generation (RAG) with large language models (LLMs) is soft context compression, where the input sequence is transformed into a shorter continuous representation. We develop a lightweight and simple mean-pooling approach that consistently outperforms the widely used compression-tokens architecture, and study training the same compressor to output multiple compression ratios. We conduct extensive experiments across in-domain and out-of-domain QA datasets, as well as across model families, scales, and compression ratios. Overall, our simple mean-pooling approach achieves the strongest performance, with a relatively small drop when training for multiple compression ratios. More broadly though, across architectures and training regimes the trade-offs are more nuanced, illustrating the complex landscape of compression methods.

Giovanni Monea, Yair Feldman, Shankar Padmanabhan et al.

The scalability of large language models for long-context reasoning is severely constrained by the linear growth of their Transformer key-value cache, which incurs significant memory and computational costs. We posit that as a model generates reasoning tokens, the informational value of past generated tokens diminishes, creating an opportunity for compression. In this work, we propose to periodically compress the generation KV cache with a learned, special-purpose token and evict compressed entries. We train the model to perform this compression via a modified joint distillation and reinforcement learning (RL) framework. Our training method minimizes overhead over the conventional RL process, as it leverages RL outputs for distillation. Empirically, our method achieves a superior memory-accuracy Pareto frontier compared to both the model without cache compression and training-free compression techniques.