Squeeze Evolve: Unified Multi-Model Orchestration for Verifier-Free Evolution
Monishwaran Maheswaran, Leon Lakhani, Zhongzhu Zhou, Shijia Yang, Junxiong Wang, Coleman Hooper, Yuezhou Hu, Rishabh Tiwari, Jue Wang, Harman Singh, Qingyang Wu, Yuqing Jian, Ce Zhang, Kurt Keutzer, Tri Dao, Xiaoxia Wu, Ben Athiwaratkun, James Zou, Chenfeng Xu
cs.AI, cs.CL
2026-04-09
By routing strong models to init and hard cases and cheaper models to the rest, Squeeze Evolve cuts reasoning cost ~3×, the first verifier-free evolution to match AlphaEvolve.
One branch of test-time scaling is evolutionary inference: have a model sample repeatedly, recombine the better answers, and sample again, leaning on population-level selection to push toward better solutions across loops. Single-model evolution has a known flaw. With no external judge to correct it, the population collapses into a few solution modes after the early loops, and pass@K actually degrades.
The standard fix is to bolt on a verifier (verifier-based evolution), as DeepMind's AlphaEvolve does, scoring candidates with an external reward. But verifiers can be prohibitively expensive. The paper notes that in some scientific-discovery tasks a single verification burns over 120 million CPU-hours, and many domains have no usable verifier at all. At the other extreme, running the whole loop on a single top-tier closed model wastes money, since proprietary output tokens cost roughly 4× to 25× what comparable open-weight models charge.
Squeeze Evolve targets both at once: no verifier, no all-frontier-model-all-the-time, and accuracy held or improved.
The framework first unifies various test-time scaling methods into one evolutionary loop, where the population evolves through select-then-recombine each round. The unification is not the point; the point is making "which model handles what" an explicit decision, in three parts.
Initialization uses the strong model. Empirically, initial population quality largely caps the final result. On HMMT 2025, strong-init with weak-aggregation reaches 89% while the reverse manages only 85%; on AIME 2025 the gap is 23 percentage points (88% vs 65%). So the expensive model is spent only on generating the ancestor population in loop 0.
A model-intrinsic signal serves as fitness to score candidate groups. Two options. Group Confidence (GC) takes the top-K token log-probabilities, where higher means the model is more self-assured, costing essentially nothing for self-scoring and one prefill pass for cross-model scoring; when the API exposes no logprobs (Gemini, for instance), it falls back to Group Diversity, the count of distinct final answers in a group. Both are lightweight, which is the appeal and the main weakness (see Limitations).
Fitness routes each group to a different model tier. Three tiers: groups that already agree go straight to majority vote with no LLM; confident or low-diversity "easy" groups go to the cheap Model 1; the remaining hard groups use the expensive Model 2. The threshold is an adaptive per-problem percentile rather than a global constant. The routing works because of a counterintuitive finding: when the candidate set is strong, a weak model can aggregate well. On AIME 2025, aggregation accuracy is 0% with zero correct seeds and 100% with all four correct. Seed quality, not aggregator strength, decides success, so strong-candidate groups can be delegated to the cheap model safely.
Two system pieces round it out. Latency-matched serving sizes GPU pools so Model 1 and Model 2 finish each loop in roughly the same wall-clock time, so neither bottlenecks the other. A custom Confidence Engine accumulates the confidence statistic on-GPU during vLLM's prefill, cutting per-request transfer from about 13 MB to about 100 bytes and lowering scoring latency 4–10×.
Across reasoning, code, vision, and discovery tasks, overall cost drops 1.3–3.3×, fixed-budget throughput rises up to about 10×, and accuracy holds or improves in most cases. Representative numbers:
| Task | Single-model baseline | Squeeze Evolve | Savings |
| AIME 2025 (Qwen3-30B) | 89.2% / $0.94 | 90.7% / $0.66 | 1.4× |
| HMMT 2025 (GPT-OSS-120B) | 89.7% / $0.41 | 92.0% / $0.25 | 1.6× |
| GPQA-Diamond | 74.0% / $0.57 | 75.9% / $0.32 | 1.8× |
| ARC-AGI-V2 (Gemini 3.1 Pro) | 93.3% / $28.85 | 97.5% / $7.74 | 3.7× |
| MMMU-Pro (Kimi-2.5) | 78.58% / $1.04 | 79.06% / $0.46 | 2.3× |
Two results carry the most weight. On ARC-AGI-V2, 97.5% at $7.74 sets a new cost-capability frontier and beats code-execution methods like Imbue (95.1% at $8.71). On the circle-packing discovery task (n=26), Squeeze Evolve (GPT-OSS-120B + 20B) scores 2.635896, slightly above the verifier-based AlphaEvolve (2.635862) and just below the ensemble method ShinkaEvolve (2.635982). This is the headline the paper returns to: the first verifier-free evolutionary method to match, and on some discovery tasks exceed, verifier-based ones.
The multimodal setting has a tidy corollary. A text-only cheap model (Qwen3.5-35B) that never sees the image actually edges out the much pricier vision model Kimi-2.5 Thinking on MMMU-Pro (79.1% vs 78.6%) at 2.7× savings, because loop 0 already digested the image into the candidates and later aggregation no longer needs it.
Routing adds only 2.4–4.3% to end-to-end latency.
The most direct takeaway for practitioners: evolutionary inference no longer requires building a verifier first. In domains where verification is costly or simply absent (scientific discovery, open-ended optimization), this path was largely closed, and there is now a baseline that matches verifier-based methods without an external judge. The division of labor, strong model for initialization and hard groups, cheap model for the rest, is a reusable heuristic not tied to this framework and worth borrowing when building your own pipeline. Mixing open and closed models becomes a first-class capability, with compute allocated by problem difficulty rather than spent uniformly.
Keep the contribution in perspective. This is an engineering and methods paper that pushes the cost-capability frontier of evolutionary inference, not one that crushes all baselines on raw accuracy. The accuracy gains on standard reasoning benchmarks are small; the real win is in cost and throughput.
The authors concede several points. The fitness signals (confidence and answer diversity) are lightweight but inherently noisy proxies; population size, group size, loop count, and routing threshold are fixed per task rather than adaptive; the method operates on full trajectories without decomposing reasoning into intermediate steps for selective regeneration; and when model-intrinsic confidence reliably separates correct from incorrect populations, plus any convergence guarantees, remain open theoretical questions.
A few concerns stand out after reading. Accuracy gains on standard reasoning benchmarks (AIME, HMMT, GPQA) are only 1–2 points, with a few slight regressions (LiveCodeBench -0.3, closed-source GPQA -1.4, BabyVision -1.67), so "no accuracy loss" does not always hold at the per-task level; the win is on average cost. The "first verifier-free to match verifier-based" banner rests almost entirely on circle packing, where 2.635896 vs AlphaEvolve's 2.635862 differs in the fifth decimal, closer to a tie than a win. Cost numbers are sensitive to the chosen model pair and pricing, and swapping pairs may not preserve the ratios. No single pair dominates all benchmarks, so deployment still requires picking pairs per task.