Multi-Turn Reasoning When Context Arrives in Pieces: Scalable Sharding and Memory-Augmented RL

In three linesLLMs lose up to 65% accuracy when task-critical information arrives across conversation turns. Compact rolling memory replaces growing history attention. A sharding pipeline converts QA datasets into multi-turn fragmented-information episodes without manual annotation. Training on sharded GSM8K improves multi-turn accuracy and generalizes zero-shot to harder math and out-of-domain long-context QA.

Summary generated by Claude — human-verified