AdaMem: Relevance-Guided Memory Token Allocation Boosts RAG Soft Compression by 3.2 Points
_reachsumit · x · 2026-09-22
A new paper (arXiv:2609.22100) introduces AdaMem, fixing an inefficiency in RAG soft compression: existing methods assign each retained passage an identical number of continuous memory embeddings regardless of its query-specific relevance.
Method: a shared query-conditioned compressor produces passage memories and relevance scores in a single pass; a deterministic allocation rule assigns more memory tokens to higher-scoring passages and can omit low-scoring ones entirely.
Results: across six open-domain QA benchmarks at matched budgets, AdaMem consistently beats the uniform-allocation baseline OSCAR: at standard 16x compression it improves sub-string match by up to 3.2 points (5.5% relative), with larger gains at aggressive 64x compression.
More from Research
- Boltzbit previews paper claiming BAST lets LLMs learn up to 1,000x faster than SOTA training — jmhernandez233 · 2026-09-22
- MiMo's GRS and GAR: scoring what makes an RL answer genuinely good, not just passing — tokenbender · 2026-09-22
- MiMo v2.6 details: grounded environment synthesis and multi-harness training for open source users — tokenbender · 2026-09-22
- MiMo leans on CodeMidas-style source-driven synthesis and agent-driven long-horizon tasks — tokenbender · 2026-09-22
- Xiaomi MiMo v2.6 ships with RL as the hero, scaling batch size, env diversity and grader compute — tokenbender · 2026-09-22
- Glasshouse v0.1 launches: an open memory benchmark with 2,847 questions over a 1.97M-token conversation — True_Mongoose_7073 · 2026-09-22