Learned Context Planning
A method in which a dedicated trained planner filters key evidence passages from long documents before the answer model performs inference, aiming to reduce contextual noise and lower inference cost.
Timeline (last 90 days)
Under an 18k character budget, anchored hybrid retrieval achieved 36.18% accuracy, BM25 achieved 35.98%, while the best direct planner-guided method reached only 34.19%.
On a test subset of 152 completely untouched questions, anchored hybrid retrieval achieved 42.11%, while the planner-guided method reached only 36.84%.
The study concluded that, under its experimental setup, learned contextual planning is only a weak correlation signal rather than a substitute for strong retrieval.
The planner was trained via supervised fine-tuning (SFT) on outcome-selection trajectories from 140 training questions and 28 development questions.
The 503-question analysis set itself includes training questions, so the results are partly transductive, which gives the planner a certain advantage.
At a 6k budget, the best planner leads by only 0.40 percentage points; by a 9k budget, the planner actually falls behind.
The authors conducted packing-order and score-flatness analyses but did not identify a reliable mechanism by which the planner works.
All Facts (7)
Under an 18k character budget, anchored hybrid retrieval achieved 36.18% accuracy, BM25 achieved 35.98%, while the best direct planner-guided method reached only 34.19%.
50%UnverifiedOn a test subset of 152 completely untouched questions, anchored hybrid retrieval achieved 42.11%, while the planner-guided method reached only 36.84%.
50%UnverifiedThe study concluded that, under its experimental setup, learned contextual planning is only a weak correlation signal rather than a substitute for strong retrieval.
50%UnverifiedThe planner was trained via supervised fine-tuning (SFT) on outcome-selection trajectories from 140 training questions and 28 development questions.
50%UnverifiedThe 503-question analysis set itself includes training questions, so the results are partly transductive, which gives the planner a certain advantage.
50%UnverifiedAt a 6k budget, the best planner leads by only 0.40 percentage points; by a 9k budget, the planner actually falls behind.
50%UnverifiedThe authors conducted packing-order and score-flatness analyses but did not identify a reliable mechanism by which the planner works.
50%