Bottom line: A frozen 12B model combined with a verified solution store achieves 100% accuracy on verified problem families with zero token consumption and deterministic, bit-exact results.
Researchers have developed a procedure in which a large language model remains frozen and a persistently growing store of verified solutions is built alongside it. On 180 test instances across nine problem families, a 12B model across four different architectures from four vendors consistently achieves 180/180 correct answers without generation tokens.
The core concept inverts the previous optimization approach: instead of the model learning from scratch during the next training cycle, the existing model is frozen and a store of verified solutions is built alongside it. A problem instance is only added to the store after independent verification that does not access the answer. For verified problem families, each new request in the same category is then answered with exact retrieval — zero generation tokens, bit-exact, deterministic.
The evaluation covered 180 new instances across nine problem families and four model architectures (including Dense and Mixture-of-Experts from four different vendors). All achieved 180/180 correct answers. A negative control experiment showed: with empty memory, the system solves nothing — performance comes entirely from the store. For open reasoning tasks, 88/88 requests were machine-verified and accepted; the reasoning method transferred to new models with 77/80 successes.
Practical requirements are minimal: memory search takes 1.4 microseconds; a complete retrieval of verified solutions requires 6–23 ms with 36 mWh energy consumption. A critical observation: approximate similarity retrieval (similarity search) selects the wrong item in 94.3% of cases on a 4,500-item store, while exact addressing produces zero errors.
The store also functions as working context: the system manages a 6,000,000-token window on a single 46-GB GPU with flat memory architecture. By comparison: vLLM stops at 30,399 tokens, SGLang silently truncates at 32,000 tokens. On standard benchmarks, frontier models remain ahead on reasoning from scratch; for verified problems the relationship reverses: frontier APIs pay for a new generation pass on every query, while verified retrievals are cost-free and return identical bits.
Source: arxiv.org · Published 25 July 2026
Lumi AI News — AI-assisted curation in accordance with Article 50 EU AI Act. Paraphrase and classification through Lumi News Pipeline v1.7.3.