Multiverse Computing's pruning method beats a baseline by nearly 23 MMLU points

Why it matters

If the results hold across independent evaluations, model builders gain a cheaper way to search aggressive pruning configurations, while Multiverse Computing gains a bridge from quantum optimization to a large, immediate AI market.

Roman Orus, the quantum-information physicist who co-founded Multiverse Computing, and a team of researchers have recast one of the bluntest ways to shrink a large language model as a problem from statistical physics. In a technical account published on Hugging Face on September 21st, the researchers described a method for choosing whole transformer blocks to delete by modeling the choices as an Ising glass.

The work gives Multiverse Computing a direct route from its founders' physics backgrounds into a pressing AI infrastructure problem: large models cost too much to run in many edge, private-cloud and offline settings. Orus is an Ikerbasque research professor at the Donostia International Physics Center and a visiting professor at the Barcelona Supercomputing Center. Multiverse Computing's founding group also includes CEO Enrique Lizaso, CTO Samuel Mugel and CMO Alfonso Rubio.

Antonio Tiene is the lead contributor named on the Hugging Face article. The block-pruning project poses a practical question: when a model has dozens of transformer blocks, which combination can disappear with the least damage?

Treating deletion choices as interacting particles

Depth pruning makes a model shorter by removing complete transformer blocks. The resulting inference speedup is predictable because each deleted block eliminates a stage of computation. Depth pruning can also sit alongside quantization, low-rank compression and knowledge distillation.

Most block-removal systems assign each block an importance score and cut the lowest-ranked candidates. That approach becomes unreliable when the effect of deleting one block depends on which other blocks are deleted with it. A block that appears disposable in isolation can become important after several neighboring or functionally related blocks have disappeared.

Multiverse Computing represents each block with a binary variable: zero keeps the block and one removes it. A second-order Taylor expansion of model loss produces an approximate Hessian matrix. Its diagonal entries estimate the importance of individual blocks, while the off-diagonal entries represent interactions between pairs of removal decisions.

The resulting constrained binary optimization problem asks for exactly M blocks out of N that minimize a calculated energy. In physics terms, the same system is an Ising glass with all-to-all interactions and a fixed number of upward-pointing spins.

That translation matters because Multiverse Computing can calculate the Hessian once using forward and backward passes over a calibration dataset. Researchers can then score proposed block combinations through relatively cheap energy calculations instead of building and benchmarking a pruned model for every candidate. The same Hessian can also be reused at several compression levels.

The search space still grows quickly. The researchers said checking roughly 29 billion ways to remove eight blocks from Llama-3.3-70B-Instruct took about two days on one GPU. For larger searches, they used heuristic solvers, including an open-source tabu solver, to generate promising low-energy configurations within seconds.

The nearly 23-point result needs a denominator

The headline result comes from removing 40 of Llama-3.3-70B-Instruct's 80 blocks without retraining. In Table 2 of the paper, Multiverse Computing reported an MMLU score of 76.9 for its constrained binary optimization method, compared with 54.0 for the block-influence baseline. The original, uncompressed model scored 82.2.

The nearly 23-point figure therefore measures the advantage over another pruning method at the same extreme compression level. It does not represent a 23-point improvement over the original model. The compressed model remained 5.3 points behind the uncompressed version on MMLU.

The pattern across the experiments is also narrower than the headline number suggests. At lighter compression levels, the paper found the methods broadly comparable. The advantage widened as the researchers removed more blocks, when interactions between deletion choices became harder to ignore. On Qwen3-14B with 12 of 40 blocks removed, the method led the comparison on MMLU by about 10 points.

The paper also applies the technique to NVIDIA's Nemotron-3-Nano-30B-A3B-FP8, which interleaves Mamba2, attention and mixture-of-experts layers. The method does not require every block to have the same structure, giving Multiverse Computing a way to search models whose internal components vary.

A promising search method still needs a human referee

The energy calculation is a proxy for model quality rather than a perfect predictor. In one Llama-3.1-8B-Instruct experiment, the lowest-energy configuration was outperformed after retraining by the 17th excited state, a higher-energy candidate that removed an early block. The finding weakens the idea that a solver can simply return one minimum-energy answer and finish the job.

The paper, first submitted on January 29th and revised on September 11th, says candidate solutions were selected manually and that an automated selection procedure would be desirable. It also reports that calibration-dataset quality affected results, with Wikitext producing weaker results than OpenHermes-2.5 or LMSYS-Chat-1M, and that the method could trail alternatives on benchmarks other than MMLU. The Hugging Face article is the latest explanation of research that Multiverse Computing has been developing through 2026, rather than the first release of the underlying work.

Multiverse Computing has published the experimental implementation on GitHub, including the Hessian-construction, energy-computation, compression and benchmarking pipeline. The repository labels the method patent pending and restricts intended use to research and non-commercial projects, which limits how directly another AI vendor can turn the code into a product.

The commercial path runs through CompactifAI, Multiverse Computing's model-compression platform. Multiverse Computing introduced CompactifAI in 2023 around tensor-network compression and has since added pruning, quantization and model-recovery techniques. In June 2025, Multiverse Computing announced a $215M funding round led by Bullhound Capital, with HP Tech Ventures, SETT, Forgepoint Capital International, CDP Venture Capital, Santander Climate VC, Quantonation, Toshiba and Capital Riesgo de Euskadi-Grupo SPRI participating.

For Orus and Multiverse Computing, the research turns a familiar startup problem into a technical advantage. Quantum computing remains a long-horizon market, while AI developers already need smaller models that preserve useful performance. Ising optimization gives Multiverse Computing's physics-heavy team a credible way to compete for that work using tools it already knows how to build.