You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
{{ message }}
Repository navigation
Paper mentioned GNN differs from the one in code #9
In the paper Section 3.5 In the training stage, we set the graph neural network as a two-layer graph attention network, is mentioned but in the code GeneralConv layer is used, Can you please clarify on that or am I missing something?
I had a question about how task nodes are handled. In Eq. 1 of paper, task nodes have their own update and aggregate from their neighbouring queries (q ∈ N(t)). In the released code, edge_index is built only from query and LLM ids , and the task embedding is concatenated onto the query features in FeatureAlign before the convolutions, so there don't appear to be task-query edges or a task node update.
I assume the task is constant per query, so a task node with only outgoing edges would broadcast the same vector to all its queries. But I wanted to check whether this was an simplification, or whether I'm missing where the Eq. 1 update happens.
Activity
varunlmxd commentedon Sep 4, 2026
I had a question about how task nodes are handled. In Eq. 1 of paper, task nodes have their own update and aggregate from their neighbouring queries (q ∈ N(t)). In the released code, edge_index is built only from query and LLM ids , and the task embedding is concatenated onto the query features in FeatureAlign before the convolutions, so there don't appear to be task-query edges or a task node update.
I assume the task is constant per query, so a task node with only outgoing edges would broadcast the same vector to all its queries. But I wanted to check whether this was an simplification, or whether I'm missing where the Eq. 1 update happens.