Is the paper consistent with the code?

[Alokia]

In the paper, Collaboration Mode Determination $F_{\theta_t}$ is defined as $Q \to T$, but in the code it is $Q, {M} \to T$, and the LLM backbone is chosen only at the last step as written in the paper $F_{\theta_m} : (Q, T, {R}) \to {M}$, but in the code it is chosen at the first step $F_{\theta_m} : Q \to {M}$.

masrouter/MAR/MasRouter/mas_router.py

Lines 61 to 65 in 89f45ef

	llms_num, log_probs = self.llm_allocation(queries_embedding, selected_tasks_emb, llms_embedding)
	selected_llms_emb = llms_num @ llms_embedding / llms_num.sum(dim=1).unsqueeze(1)
	selected_llms:List[List[Dict[str,str]]] = [[llms[idx] for idx, num in enumerate(row) for _ in range(int(num))] for row in llms_num]
	selected_reason_idx, reason_log_probs = self.reasoning_selector(queries_embedding, selected_tasks_emb, selected_llms_emb, reasonings_embedding)