frontier models
Isolated Planners Avoided Multi-Task Collapse
GRASP separated guideline generation, strategy revision and trajectory assessment into context-isolated modules.
Summary
GRASP separated guideline generation, strategy revision and trajectory assessment into context-isolated modules.
The framework pre-compiles global planning guidance, explores local alternatives in separate context windows and assigns independent modules to judge candidate trajectories. Across calendar scheduling, ZebraLogic and SciBench Math, the authors report gains over direct planners; in interleaved dual-task tests, accuracy improved by as much as 16.7 percentage points and did not show the immediate degradation seen in the compared standard planners. These are results from the selected benchmarks and model comparisons, not a guarantee for open-ended strategic work.
Why it matters
GRASP separated guideline generation, strategy revision and trajectory assessment into context-isolated modules.
Limits and context
- Across calendar scheduling, ZebraLogic and SciBench Math, the authors report gains over direct planners; in interleaved dual-task tests, accuracy improved by as much as 16.7 percentage points and did not show the immediate degradation seen in the compared standard planners.
- These are results from the selected benchmarks and model comparisons, not a guarantee for open-ended strategic work.
Key claims
GRASP separated guideline generation, strategy revision and trajectory assessment into context-isolated modules.
Qualification: Across calendar scheduling, ZebraLogic and SciBench Math, the authors report gains over direct planners; in interleaved dual-task tests, accuracy improved by as much as 16.7 percentage points and did not show the immediate degradation seen in the compared standard planners.
Evidence: source-2026-09-26-016
Sources
- arXiv preprint 2609.30147arXiv · primary research
Corrections
No corrections have been recorded for this story.