News · 2026-07-18

Grafting a Verified Solution Cache Lets a Frozen Model Skip Fine-Tuning Entirely

A frozen small language model can score dramatically better on a hard math test just by loading in a saved cache of verified solutions, with no retraining and no weight changes at all. A paper posted July 15, 2026 reports that this 'grafting' pushed a frozen Gemma-4-12B model from 80.0% to 93.3% accuracy on the AIME 2025 math-competition benchmark, and the authors argue it points toward models that keep improving through a growing library of verified knowledge rather than through additional training runs.

Key facts

• A frozen Gemma-4-12B model rose from 80.0% to 93.3% on AIME 2025 (a hard math-competition test) after grafting in a library of verified solutions, with zero weight updates.
• Published July 15, 2026, in "Smarter and Cheaper at Once: Byte-Exact KV-Cache Grafting Turns a Frozen Small Model into a Verified-Knowledge Flywheel" (arXiv 2607.14431).
• The grafting is byte-exact: restoring a cached solution reproduces SHA-256-identical output logits under pinned deterministic settings, not an approximation of the original computation.
• The result lands inside a broader July 2026 cluster of on-policy-distillation papers exploring how to reuse a model's own verified outputs as a training or inference-time signal.

Fine-tuning a model is expensive and risky: it takes compute, it takes time, and it can quietly erode capabilities the model already had. This paper asks a different question entirely -- what if you never touch the weights, and instead hand the model a way to remember work it has already verified as correct?

The mechanism it exploits is the KV-cache, the internal scratchpad a transformer builds as it generates text. Normally, as a model works through a prompt, it stores the key and value vectors from its attention layers so it does not have to recompute them for every new token; that cache is usually thrown away the moment the conversation ends. The paper's contribution is showing that this scratchpad can instead be captured, saved as an exact artifact, and reloaded later into the same frozen model, so the model effectively resumes from the exact internal state it was in when it solved a problem correctly the first time.

That distinguishes this from ordinary distillation or fine-tuning, both of which change what a model's weights encode. Here, nothing about the underlying network is touched. A useful everyday comparison: it is less like teaching a student a concept so they remember it forever, and more like handing them a calculator that already has the exact keystrokes for a solved problem loaded into its memory -- press restore, and it picks up exactly where the verified work left off, byte-for-byte, not an approximate crib note.

The headline number is the AIME 2025 jump: a frozen Gemma-4-12B went from getting roughly four in five problems right to better than nine in ten, purely by grafting in the verified-solution cache. Because none of that gain came from retraining, the cost of getting there is just the cost of building and storing the cache library -- an argument for treating verified solutions as a reusable, growing resource rather than something a model has to relearn every time. That connects to the broader interest in test-time compute: rather than spending more compute per query on reasoning at inference, this approach spends it once, up front, verifying a solution, then reuses that verified state cheaply afterward.

The paper's own framing calls this a "verified-knowledge flywheel": every new problem the system solves and verifies can be added to the cache library, so a static, frozen model effectively keeps getting better at the problems the library covers without ever being retrained. It sits alongside a wave of July 2026 on-policy-distillation research -- including SEED, Demystifying OPD, and On-Policy Delta Distillation -- all built around training on what a model actually does rather than generic supervision. Ground Truth has already covered the broader on-policy-distillation story in a piece on a weak teacher producing a stronger student; this result is a distinct variant within that cluster, trading retraining for a cache you can graft in and out.

The honest caveat: this is a single paper's self-reported result on its own benchmark setup, not yet independently reproduced by outside researchers, and the "flywheel" language is the authors' own characterization of what the technique could enable rather than an established industry term. Whether the approach holds up on tasks beyond math competitions, and at what scale a verified-solution library becomes unwieldy to build and maintain, are open questions the paper does not settle. Still, byte-exact restoration backed by matching SHA-256 logits is a concrete, checkable claim, not a vague promise -- and it points at a genuinely different lever for making frozen models cheaper and smarter at once, one that does not run through the usual retraining pipeline at all.

Primary source, verified: read the paper → (arXiv 2607.14431)

Ground Truth. (2026, July 18). Grafting a Verified Solution Cache Lets a Frozen Model Skip Fine-Tuning Entirely. Ground Truth. https://groundtruth.day/news/kv-cache-grafting-verified-knowledge-flywheel.html

@misc{groundtruth:kv-cache-grafting-verified-knowledge-flywheel,
  title  = {Grafting a Verified Solution Cache Lets a Frozen Model Skip Fine-Tuning Entirely},
  author = {{Ground Truth}},
  year   = {2026},
  month  = {jul},
  url    = {https://groundtruth.day/news/kv-cache-grafting-verified-knowledge-flywheel.html}
}

Topics: kv-cache · distillation · efficiency · post-training · reasoning