Squeeze Evolve: Unified Multi-Model Orchestration for Verifier-Free Evolution

1. Problem Statement

The paper addresses two critical bottlenecks in verifier-free evolutionary inference (improving LLM outputs through iterative selection and recombination without external ground-truth verification):

1. Diversity Collapse: Without external verification, repeated evolutionary loops tend to converge toward narrow solution modes. The population loses semantic diversity, causing the upper bound of performance (e.g., pass@K) to degrade significantly.
2. Cost Inefficiency: Existing methods often rely on a single, high-cost model for all evolutionary steps (initialization, generation, recombination). This wastes compute, as cheaper models are often sufficient for "easy" aggregation tasks, while expensive models are over-utilized.

The core challenge is: How can we achieve high-performance verifier-free evolution under tight budget constraints while preventing diversity collapse?

2. Methodology: SQUEEZE EVOLVE

The authors propose SQUEEZE EVOLVE, a unified multi-model orchestration framework that dynamically routes evolutionary operations to the most cost-effective model based on confidence-derived fitness signals.

A. Unified Evolutionary Framework

The authors formalize existing test-time scaling methods (e.g., Majority Voting, Self-Refinement, RSA, AlphaEvolve) into a single iterative operator $\Phi_f = \text{recomb}_f \circ \text{select}_f$.

• Initialization: A strong (expensive) model generates the initial population $P^{(0)}$.
• Evolutionary Loop:
  1. Selection: Candidates are grouped based on a fitness signal $f$.
  2. Routing: Each group is assigned to a specific recombination tier based on its difficulty.
  3. Recombination: The assigned model aggregates the group into a refined candidate.
  4. Update: The population is updated (either by replacement or accumulation).

B. Key Mechanisms

1. Confidence-Based Routing (The "Squeeze"):

   • The system uses Group Confidence (GC) or Group Diversity (D) as a proxy for fitness.
   • GC: Calculated from token log-probabilities (self-confidence) or a cross-model forward pass (cross-confidence). High confidence indicates the group has reached consensus (easy), while low confidence indicates conflict/uncertainty (hard).
   • Routing Logic:
     • High Confidence / Low Diversity: Routed to a lightweight non-LLM tier (e.g., majority vote) or a cheap model ($M_1$).
     • Low Confidence / High Diversity: Routed to an expensive, high-capability model ($M_2$) to resolve uncertainty and preserve diversity.
   • This ensures expensive models are reserved only for high-marginal-utility steps.

2. Multi-Model Orchestration Strategy:

   • Initialization: Always performed by the strongest model to ensure high-quality seeds (empirically shown to dominate final accuracy).
   • Aggregation: Weaker models can effectively aggregate candidates if the candidate set is already strong (i.e., contains correct seeds).
   • System Co-Design:
     • Latency-Matched Serving: GPU pools for $M_1$ and $M_2$ are sized to complete their respective workloads in roughly the same wall-clock time, preventing bottlenecks.
     • Custom Confidence Engine: A specialized vLLM path computes confidence scores directly on the GPU, returning only a scalar. This reduces transfer latency by 4–10× compared to standard log-prob extraction, making the routing overhead negligible.

3. Key Contributions

1. Unified Framework: Formalized diverse test-time scaling methods into a single evolutionary framework, identifying diversity collapse as the primary failure mode of verifier-free evolution.
2. Confidence-Based Routing: Introduced a lightweight mechanism to route recombination groups to models based on intrinsic confidence signals, eliminating the need for external verifiers or trained reward models.
3. Cost-Capability Frontier: Demonstrated that multi-model orchestration shifts the cost-performance frontier, achieving state-of-the-art (SoTA) results at significantly lower costs.
4. System Implementation: Developed a serving system with custom confidence engines and latency-matched GPU pools, achieving up to 10× throughput speedup under fixed budgets.

4. Experimental Results

The framework was evaluated across 8 benchmarks: Math (AIME 2025, HMMT 2025, GPQA-Diamond), Coding (LiveCodeBench V6), Vision (MMMU-Pro, BabyVision), Visual Reasoning (ARC-AGI-V2), and Scientific Discovery (Circle Packing).

• Cost Reduction:

  • Reduced API costs by 1.3× to 3.3× compared to single-model baselines while maintaining or exceeding accuracy.
  • On ARC-AGI-V2, achieved 97.5% accuracy at $7.74/task (vs. $17.60 for GPT-5.4-Pro), setting a new SoTA cost-capability frontier.
  • On Circle Packing, matched or exceeded verifier-based methods (like AlphaEvolve) without using external code execution or ground-truth feedback.

• Throughput & Efficiency:

  • Increased fixed-budget serving throughput by 4× to 10× (e.g., Qwen3-30B/235B pair) by offloading easy tasks to smaller models.
  • Routing overhead was minimal (2.4% – 4.3% of end-to-end latency).

• Multimodal Insights:

  • On vision tasks (MMMU-Pro), a text-only cheap model (Model 1) used for aggregation after the initial image-grounding step achieved 2.7× cost savings while matching or exceeding the performance of expensive vision-capable models. This suggests visual understanding is primarily needed at initialization.

• Diversity Preservation:

  • Unlike single-model evolution, which saw pass@K drop as diversity collapsed, SQUEEZE EVOLVE maintained high diversity and performance across loops by introducing models with different priors and reasoning styles.

5. Significance

• Economic Viability: Makes verifier-free evolutionary search economically practical for real-world deployment by drastically reducing token costs.
• Democratization: Enables the use of open-source models to achieve performance previously reserved for expensive proprietary models, especially in scientific discovery and reasoning tasks.
• Paradigm Shift: Moves the field from "scaling compute with one model" to "intelligent orchestration of heterogeneous models," proving that allocation efficiency is as critical as raw model capability.
• Verifier-Free Breakthrough: Demonstrates that verifier-free methods can match verifier-based approaches (like AlphaEvolve) in discovery tasks, removing the dependency on expensive or unavailable external reward signals.

In summary, SQUEEZE EVOLVE proves that by intelligently routing tasks between cheap and expensive models based on confidence, one can overcome the diversity and cost bottlenecks of evolutionary inference, achieving superior performance at a fraction of the cost.