BWhat changed
A short project summary is easy to search, but may omit the number or exception needed to answer a later question. Xia and colleagues pair a primary abstraction with concrete memory content, then add cue anchors that offer several ways to find the same entry. Their retrieval policy can refine a query, explore connected entries or stop. The work tests whether this combination improves answer quality on conversational-memory benchmarks. Its contribution is the external abstraction-and-cue representation and the specified retrieval methods, distinct from the neural Memora model introduced within MIRAS.
Scientific scope
A short project summary is easy to search, but may omit the number or exception needed to answer a later question. Xia and colleagues pair a primary abstraction with concrete memory content, then add cue anchors that offer several ways to find the same entry. Their retrieval policy can refine a query, explore connected entries or stop. The work tests whether this combination improves answer quality on conversational-memory benchmarks. Its contribution is the external abstraction-and-cue representation and the specified retrieval methods, distinct from the neural Memora model introduced within MIRAS. The reported averages depend on model-based judging and specific benchmark settings. Some question categories and word-overlap metrics favor other configurations. The policy retriever uses additional model calls and is slower than semantic retrieval in the measured setup. The trained-policy comparison uses a separate held-out split and is not the configuration behind every headline result. Context reduction is not total cost reduction. Automatically consolidating two entries does not establish that two real people or scientific claims are identical.
Evidence
The linked original works support the scoped statement and contribution attribution. WRS has not independently replayed experiments or reconstructed a full mathematical proof. The reported averages depend on model-based judging and specific benchmark settings. Some question categories and word-overlap metrics favor other configurations. The policy retriever uses additional model calls and is slower than semantic retrieval in the measured setup. The trained-policy comparison uses a separate held-out split and is not the configuration behind every headline result. Context reduction is not total cost reduction. Automatically consolidating two entries does not establish that two real people or scientific claims are identical.
Contributors
- The named authors in the research record · SCIENCE: Source-listed human researchers and engineers for the exact method or resource. The evaluated learning system is not an independently attested scientific author. Earlier methods, funding and data-collection roles remain separate. Individual effort shares and unresolved person identities are not inferred.
Provisional WRS assessment, edition 2026-10-23.9. Independent expert calibration is pending. This grade does not certify correctness or assign ownership.
Full scientific record · Browse contribution grades · Request a correction