DSGC lifts full-chain retention from 0.03 to 0.90 on lexical encoders in arXiv:2608.20400 benchmark
arXiv:2608.20400 defines structurally indirect prerequisite eviction and demonstrates that a one-hop graph rule recovers near-perfect chain retention under constrained budgets. The benchmark and traces isolate retention failure as a distinct boundary from retrieval ranking. DSGC performance is reported across lexical and sentence encoders with explicit scaling and budget thresholds.
The paper isolates structurally indirect prerequisite eviction as a retention failure that occurs before retrieval queries are issued. Under fixed memory budgets, blocks that hold weak lexical overlap with the final query but supply necessary upstream facts are evicted. The authors supply per-seed trace diagnostics and a reproducible pipeline that records exact eviction order and chain completeness.
Main-suite results show the one-hop graph rule in DSGC restores 0.87 and 0.77 absolute gains respectively. Robustness checks map the regimes where the rule holds: budgets below 40 percent of required chain length and scaling factors above 3x cause measurable degradation. These thresholds are reported with seed-level variance rather than aggregate scores.
The work reframes agent memory evaluation around pre-retrieval retention boundaries instead of post-retrieval ranking metrics. Released artifacts allow direct measurement of eviction order against dependency graphs, enabling targeted comparison with future multi-hop or learned eviction policies.
Operational deployment will require budget schedulers that expose per-block dependency metadata before any eviction decision is executed.
DSGC: One-hop retention will fall below 0.70 at 3x scale factor when budget drops under 35 percent of chain length within the next benchmark release.
Sources (2)
- [1]Primary Source(https://arxiv.org/abs/2608.20400)
- [2]Supporting Source(https://arxiv.org/abs/2305.14325)