Hi authors, thanks for releasing RMBench. I have two questions:
-
TMC is defined by the minimum number of “task-relevant observations,” but how are “task-relevant” and “encoded from an observation” operationally determined? Since the memory representation is unrestricted, it seems possible to compress an arbitrary history into a fixed-size state, so the number of retained frames may not reflect memory complexity.
-
On M(n) tasks, Mem-0 also introduces subtask decomposition, a planner, subtask supervision, and a termination classifier, while the baselines do not use decomposition. Does Table 1 therefore measure the gain of the complete hierarchical system rather than memory alone?
Could you clarify these points or provide a comparison where only the memory input is changed?
Hi authors, thanks for releasing RMBench. I have two questions:
TMC is defined by the minimum number of “task-relevant observations,” but how are “task-relevant” and “encoded from an observation” operationally determined? Since the memory representation is unrestricted, it seems possible to compress an arbitrary history into a fixed-size state, so the number of retained frames may not reflect memory complexity.
On M(n) tasks, Mem-0 also introduces subtask decomposition, a planner, subtask supervision, and a termination classifier, while the baselines do not use decomposition. Does Table 1 therefore measure the gain of the complete hierarchical system rather than memory alone?
Could you clarify these points or provide a comparison where only the memory input is changed?