JPMorgan Chase & Co (United States)

Multi-task deployment often requires merging independently fine-tuned models into one, but existing merging methods compress all tasks into a single shared low-rank space, so tasks whose subspaces are misaligned lose task-relevant directions. GRAM uses only the fine-tuned weights, first partitioning tasks by the principal-angle geometry of their output subspaces and then running a standard SVD-based merge inside each group. On GLUE with Flan-T5-Base it reaches 83.8% average accuracy, above the 81.9% of global merging and the 81.7% of the strongest baseline TSV-Merge, at the cost of one extra deployed model.

Multi-task deployment often requires merging independently fine-tuned models into one, but existing merging methods compress all tasks into a single shared low-rank space, so tasks whose subspaces are misaligned lose task-relevant directions. GRAM uses only the fine-tuned weights, first partitioning tasks by the principal-angle geometry of their output subspaces and then running a standard SVD-based merge inside each group. On GLUE with Flan-T5-Base it reaches 83.8% average accuracy, above the 81.9% of global merging and the 81.7% of the strongest baseline TSV-Merge, at the cost of one extra deployed model.

Personalized assistants must recover sparse evidence from long conversation histories, while queries vary widely in complexity. APDMem organizes history into four layers, from thematic summaries and personalized key facts down to turn-level evidence notes and raw messages, and a controller reads high-level summaries first, drilling into finer evidence only when needed. On LongMemEval it achieves strong long-context memory reasoning while accessing only 8% of the total conversations.

Personalized assistants must recover sparse evidence from long conversation histories, while queries vary widely in complexity. APDMem organizes history into four layers, from thematic summaries and personalized key facts down to turn-level evidence notes and raw messages, and a controller reads high-level summaries first, drilling into finer evidence only when needed. On LongMemEval it achieves strong long-context memory reasoning while accessing only 8% of the total conversations.