AdaMem
A soft compression method for RAG that performs relevance-guided dynamic memory token allocation. Using a shared query-conditioned compressor, it simultaneously generates passage memory representations and relevance scores in a single forward pass, enabling efficient use of a limited memory budget.
Timeline (last 90 days)
AdaMem is a relevance-guided soft compression framework that maps learned passage relevance estimates to query-dependent dynamic allocation of a fixed memory token budget.
AdaMem uses a shared, query-conditioned compressor that simultaneously produces continuous passage memory representations and a relevance score for each passage in a single forward pass.
AdaMem partitions memory tokens via deterministic allocation rules, assigning more memory tokens to higher-scoring passages and optionally omitting passages with scores below a threshold.
AdaMem is evaluated on six open-domain question answering benchmarks and consistently outperforms OSCAR and other soft compression methods under the same memory budget.
At 16× compression, AdaMem's substring match improves by up to 3.2 percentage points (5.5%) over the uniform allocation baseline, with an average relative gain of 3.4%.
At 64× compression, AdaMem's average relative gain expands to 14.6%, with a maximum improvement of 9.8 percentage points (19.7%) on the PopQA dataset.
The higher the compression ratio and the tighter the budget, the greater the advantage of intelligent allocation over uniform allocation.
Compared to full-context inference, AdaMem can reduce inference latency by up to 4×, and can match the answer quality of the uncompressed scheme while achieving latency as low as 1/4 of the full-context baseline.
All Facts (8)
AdaMem is a relevance-guided soft compression framework that maps learned passage relevance estimates to query-dependent dynamic allocation of a fixed memory token budget.
50%UnverifiedAdaMem uses a shared, query-conditioned compressor that simultaneously produces continuous passage memory representations and a relevance score for each passage in a single forward pass.
50%UnverifiedAdaMem partitions memory tokens via deterministic allocation rules, assigning more memory tokens to higher-scoring passages and optionally omitting passages with scores below a threshold.
50%UnverifiedAdaMem is evaluated on six open-domain question answering benchmarks and consistently outperforms OSCAR and other soft compression methods under the same memory budget.
50%UnverifiedAt 16× compression, AdaMem's substring match improves by up to 3.2 percentage points (5.5%) over the uniform allocation baseline, with an average relative gain of 3.4%.
50%UnverifiedAt 64× compression, AdaMem's average relative gain expands to 14.6%, with a maximum improvement of 9.8 percentage points (19.7%) on the PopQA dataset.
50%UnverifiedThe higher the compression ratio and the tighter the budget, the greater the advantage of intelligent allocation over uniform allocation.
50%UnverifiedCompared to full-context inference, AdaMem can reduce inference latency by up to 4×, and can match the answer quality of the uncompressed scheme while achieving latency as low as 1/4 of the full-context baseline.
50%