You don't need long context
Frequent context compaction preserved task accuracy with little or no loss across the long-input and long-horizon tasks tested.
The authors report that later tokens can cost up to seven times as much to generate as earlier tokens on frontier open-weight models. On long-horizon software engineering tasks, performance did not collapse even when models took at most one action per context. They use that result to outline “semantic state space models,” questioning whether long contexts are necessary for the tasks studied.
HF Daily Papers' note
The authors report that later tokens can cost up to seven times as much to generate as earlier tokens on frontier open-weight models. On long-horizon software engineering tasks, performance did not collapse even when models took at most one action per context. They use that result to outline “semantic state space models,” questioning whether long contexts are necessary for the tasks studied.
HF Daily Papers' note
score 5