File size: 1,933 Bytes
448d6a5 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 | """Baselines and ablations for Claim 2 / Claim 3.
- Generate, GenerateCoT: vanilla base models.
- Static pipeline: a fixed hand-crafted operator workflow (stand-in for the paper's
fixed agentic pipelines AI-Scientist / AI-Researcher / VIRSCI, whose exact code we
do not run; the point of the paper is *fixed* vs *learned/adaptive* flows).
- Single-operator: Generate -> <one operator> (Table 5 ablation).
- Shuffle operator: random flow, no controller (Appendix E baseline).
- MindFlow deploy: deterministic top-p rollout of the trained supernet controller.
"""
from __future__ import annotations
import numpy as np
from . import operators as ops
from .flow import execute_flow
STATIC_PIPELINE = ["Critical", "Constraint", "Convergent"]
def gen_idea(topic, related, model=None, seed=0):
return ops.op_generate(topic, related, model=model, seed=seed)
def gen_cot_idea(topic, related, model=None, seed=0):
return ops.op_generate_cot(topic, related, model=model, seed=seed)
def flow_idea(seq, topic, related, model=None, seed=0):
idea, _ = execute_flow(seq, topic, related, model=model, seed=seed)
return idea
def static_pipeline_idea(topic, related, model=None, seed=0):
return flow_idea(STATIC_PIPELINE, topic, related, model=model, seed=seed)
def single_operator_idea(op_name, topic, related, model=None, seed=0):
return flow_idea([op_name], topic, related, model=model, seed=seed)
def shuffle_flow(rng, min_ops=1, max_ops=3):
k = rng.integers(min_ops, max_ops + 1)
return list(rng.choice(ops.REFINE_OPS, size=k, replace=False))
def shuffle_idea(topic, related, rng, model=None, seed=0):
return flow_idea(shuffle_flow(rng), topic, related, model=model, seed=seed)
def mindflow_deploy_idea(net, topic, related, model=None, seed=0, threshold=0.6):
seq, _ = net.rollout_flow(topic, threshold=threshold)
return flow_idea(seq, topic, related, model=model, seed=seed), seq
|