Squeeze Evolve Framework Overview

• Squeeze Evolve is a hybrid framework that combines resource squeezing with evolutionary search to optimize machine learning models effectively.
• It employs a multi-model orchestration strategy using intrinsic fitness signals, achieving up to 3.3× API cost reductions and 10× throughput gains.
• It evolves compact neural architectures, balancing model size, accuracy, and speed for applications in resource-constrained environments.

The Squeeze Evolve Framework refers to a family of machine learning and optimization approaches that integrate "squeezing" (resource or parameter reduction, model selection, or localized exploitation) with "evolution" (diversity maintenance, population-based search, or iterative improvement). These frameworks are characterized by their hybridization of efficiency-driven constraints and evolutionary search, typically in the context of model architectures, inference workflows, or complex system optimization. Three distinct and influential instantiations appear in the literature: multi-model evolutionary inference for verifier-free settings (Maheswaran et al., 9 Apr 2026), deep neural architecture compression via evolutionary synthesis (Shafiee et al., 2017), and metaheuristic parameter fitting via evolutionary SMC with local search acceleration (Beguerisse-Diaz et al., 2011). Each embodies the "Squeeze + Evolve" paradigm but differs in domain and technical construction.

1. Unified Multi-Model Orchestration for Verifier-Free Evolutionary Inference

The Squeeze Evolve framework in evolutionary inference targets verifier-free search, where external correction (oracle or reward model) is unavailable or prohibitively expensive. In this scenario, candidate solution populations are evolved interatively using only model-intrinsic signals—such as log-probabilities or cohort diversity—as fitness proxies (Maheswaran et al., 9 Apr 2026).

Core Principle: Marginal Utility Allocation

Squeeze Evolve employs a simple allocation policy: "allocate model capability where it has the highest marginal utility." This entails routing only the hardest candidate groups, identified by proxy metrics, to high-cost, high-capability models, while delegating the remaining (easier) groups to cheaper models or lightweight aggregators. Formally, for group $g$, the marginal utility of upgrading from base model $M_1$ to stronger model $M_2$ is

$$MU(g) = \frac{U(M_2, g) - U(M_1, g)}{C(M_2) - C(M_1)}$$

where $U$ quantifies group-level aggregation success or fitness gain, and $C(\cdot)$ gives per-use cost. A cost-capability trade-off is realized by maximizing the sum of expected utility minus regularized cost over all group queries.

Workflow and Routing Policy

The evolutionary process iterates over $T$ loops. At each loop:

• Populations $P^{(t)}$ are grouped via selection on an intrinsic fitness signal $f$.
• For each group $g$, a proxy fitness $M_1$0—e.g., group confidence or answer diversity—determines routing:
  • Groups with $M_1$1: sent to a non-LLM lite aggregator.
  • Next $M_1$2 (by $M_1$3): served by the cheap model $M_1$4.
  • Remaining: routed to $M_1$5.

Token- and group-level confidences are defined:

$M_1$6

$M_1$7

$M_1$8

Empirically, this policy achieves up to $M_1$9 API cost reduction and $M_2$0 throughput gains over uniform single-model evolution, while preserving or exceeding best-known accuracy on benchmarks such as AIME 2025, LiveCodeBench V6, GPQA-Diamond, Multimodal Vision, ARC-AGI-V2, and scientific discovery tasks.

Table: Squeeze Evolve Empirical Results

Benchmark	Cost Reduction	Throughput Gain	Accuracy Impact
Math & Coding	1.4–3.3×	up to 10×	Matches/exceeds strong model
Multimodal Vision	1.8–2.7×	–	No image after loop 0 for $M_2$1
ARC-AGI-V2	$M_2$2	–	97.5% (@\$7.74), SOTA
Discovery (Packing)	–	–	Matches AlphaEvolve, ShinkaEvolve

2. Evolutionary Deep Architecture Synthesis: SquishedNets

In highly resource-constrained scenarios (e.g., edge deployment), Squeeze Evolve describes a framework for synthesizing extremely compact neural architectures by sequentially combining macroarchitectural minimization ("squeeze") with evolutionary pruning ("evolve") (Shafiee et al., 2017).

Macroarchitecture "Squeeze"

The initial step adapts SqueezeNet v1.1 (originally 1000-class) to the 10-class ImageNet-10 by reducing the output layer, minimally affecting other configurations. This modification alone reduces parameter count by ≈40%.

Evolutionary Synthesis "Evolve"

An iterative evolutionary strategy encodes network weights (or filters) as binary genes in chromosomes. At each generation:

• Offspring are generated by stochastic synaptic pruning, driven by synaptic probability vectors updated via selection and resource constraints.
• Fitness is a tradeoff: $M_2$3 (accuracy vs. size).
• Environmental factor $M_2$4 ensures progressively smaller models.

The process yields Pareto-optimal networks: model sizes from 2.4MB to 0.95MB (5.17× smaller than SqueezeNet v1.1; 253× smaller than AlexNet), with accuracy from 81.2% to 77.0%, and speed from 156 to 256 img/sec on Jetson TX1—demonstrating a size-efficiency-accuracy tradeoff.

Table: SquishedNets Performance

| Model | Size (MB) | Speed (img