Develop query-conditioned context pruning for SmartSearch

Develop and evaluate aggressive query-conditioned context pruning methods that retain only words relevant to the specific input question within retrieved passages in SmartSearch, and assess their impact on token reduction and answer accuracy.

Background

The paper identifies a compilation bottleneck where ranking and truncation, rather than retrieval, limit performance. It suggests context compression as a promising direction to reduce token usage without sacrificing evidence quality, noting preliminary gains from format-level changes.

However, beyond basic format-level compression, the authors highlight that more aggressive, query-conditioned pruning—selectively retaining only words pertinent to the question—has not yet been explored, leaving a concrete opportunity for future research.

References

More aggressive query-conditioned pruning (retaining only words relevant to the specific question) remains unexplored.

— SmartSearch: How Ranking Beats Structure for Conversational Memory Retrieval  (2603.15599 - Derehag et al., 16 Mar 2026) in Section 6: Future Work (Context compression)

Reducing gallery-side storage under query-dependent activation remains open.

— AdaptiveEmbed: Sample-Adaptive Multi-Vector Representation for Multimodal Retrieval  (2608.25412 - Liu et al., 26 Aug 2026) in Section D, Discussion and Limitations

The placebo control (appendix I) finds much of the same deficit with off-topic solutions, leaving the benefit of relevance filtering unresolved.

— Selection, Recombination, or a Fresh Solve? A Candidate-Free Control for Single-Pass Test-Time Aggregation  (2608.18379 - Farmanfarmaian, 18 Aug 2026) in Appendix C, “What the regimes would imply if observable”

The replay also filters a candidate generated from the complete bank rather than regenerating from a pruned bank; a true pre-generation top-$k$ context may change the answer, and the direction of that difference is unknown.

— Whose record is this? Diagnosing and authorizing record use in personalized multimodal models  (2609.04801 - Mao et al., 4 Sep 2026) in Section 6, Limitations and Discussion