SparseRead puts a budget on what an agent reads, not just what it remembers
SparseRead puts a budget on what an agent reads, not just what it remembers
An agent does not necessarily need a whole webpage in its context to answer one question about it. SparseRead studies how to make that decision explicit before unnecessary text reaches the model. This is catch-up research coverage: the paper was first submitted August 23 and revised November 7, not released for the first time this week. Version history.
The authors describe a reading layer that tracks changing evidence needs, chooses how to expose source material and limits cumulative raw content. It retains a full-view fallback rather than assuming a short excerpt is always sufficient. Across six models and five scenarios, they report maximum reductions of 92.9% in total tokens and 89% in elapsed time; quality was preserved or improved in 97% of settings. The more specific web deep-search results were 31–43% fewer tokens and 23–37% less completion time. These are reported experimental results, not a universal saving. Paper.
Guth analysis: The useful distinction is between making a long context cheaper and avoiding unnecessary context altogether. A reading budget can make an agent's evidence choices inspectable, but the budget must not become a reason to skip inconvenient or contradictory material. A practical evaluation should track missed evidence, fallback frequency, answer correctness and end-to-end cost together. The largest percentage alone does not establish the best default for a production research agent.
Evidence limits: No independent replication was verified in this pass. The indexed repository appeared in search, but direct retrieval failed, so this draft makes no claim about a tested package, installation or current release.
Sources and citations
Revision history
Published .