CTWM
Core-Tail World Model; a rank-based memory controller that dynamically allocates context budgets via a single exponential parameter τ. It grants full representation to core states while summarizing tail states, achieving a balance between coverage and token efficiency in long-context language agent memory systems.
Timeline (last 90 days)
In CTWM, as τ approaches 0 the budget tends toward uniform allocation, and as τ increases resources become increasingly skewed toward top-ranked core states.
In Synthetic Graph World, CTWM reduces prompt tokens by 5.9% compared to the graph-memory baseline, lowers lower-half tail prediction error by 13.6%, and fully preserves state and transition coverage.
On ALFWorld, CTWM achieved consistent token savings under identical paired comparisons.
On LongMemEval, CTWM achieved a 24.48% token reduction while maintaining overall accuracy at the same level.
Heavy-tailed memory traces are not only a diagnostic signature of limited retrieval mechanisms but can also serve as a practically exploitable control signal.
As a newly released preprint, the study's conclusions still await validation in broader settings, such as whether a single exponential τ is sufficient for real-world tasks with dynamically shifting distributions.
All Facts (6)
In CTWM, as τ approaches 0 the budget tends toward uniform allocation, and as τ increases resources become increasingly skewed toward top-ranked core states.
50%UnverifiedIn Synthetic Graph World, CTWM reduces prompt tokens by 5.9% compared to the graph-memory baseline, lowers lower-half tail prediction error by 13.6%, and fully preserves state and transition coverage.
50%UnverifiedOn ALFWorld, CTWM achieved consistent token savings under identical paired comparisons.
50%UnverifiedOn LongMemEval, CTWM achieved a 24.48% token reduction while maintaining overall accuracy at the same level.
50%UnverifiedHeavy-tailed memory traces are not only a diagnostic signature of limited retrieval mechanisms but can also serve as a practically exploitable control signal.
50%UnverifiedAs a newly released preprint, the study's conclusions still await validation in broader settings, such as whether a single exponential τ is sufficient for real-world tasks with dynamically shifting distributions.
50%