Decoded
← Feed
arXiv:2609.18842AILG16 sep 2026

Infinite-Parameter LLMs: Generating and Adapting Weights from Live Data

1 min readHN ×1PDF ↗
TRAINED SELECTOR53.3%DENSE RETRIEVAL45.3%ZERO-SHOT CONFIDENCE22.7%BM2520.7%0100% TOP-1 ROUTING ACCURACY
fig. top-1 routing accuracyTrained selector picks the right code more often

Compile what users tell a chatbot into its weights instead of rereading it every prompt.

48.0F1

MS MARCO answer F1, weights

Beats putting the same ten noisy passages in the prompt

How it works

  1. Frozen models forget

    Facts users supply live in the prompt, get reread on every request, then are thrown away after the session.

  2. Generate weights, track belief

    A hypernetwork turns each passage into a low-rank weight tweak; a running belief picks which tweak to apply.

  3. Beats prompt and retrieval

    On long, noisy evidence, weights beat the prompt, and the trained selector out-routes dense retrieval on four datasets.

read onceone codescore codestop-1modulate FFNLive dataData-to-weightencoderCachedcode poolBeliefselectorLow-rankdeltaFrozenQwen3-8B answerread onceone codescore codestop-1modulate FFNLive dataData-to-weightencoderCachedcode poolBeliefselectorLow-rankdeltaFrozenQwen3-8B answer

What they found

Which comes out on top on Top-1 routing accuracy?

Packing more into one code

try it

Drag the pool size up and watch one big read lose accuracy.

Passages in poolAnswer F1Small pool, read holdsLarge pool, read dilutesBig read drops to 27.8
Answer F1
medium
here
Small pool, read holds
A fixed-size code dilutes, so the paper keeps one small code per passage and selects.Illustrative model, not the paper's data

Terms

6 terms

Hypernetwork A network that outputs weight changes for another model, here from a text passage.

Instead of rereading your notes before every answer, you memorize each note as a reflex card, then keep a running guess which card fits the conversation.

Where it breaks
Cards are generated weights, not memories; total knowledge capacity does not grow.

Limits

  • Prompt still wins on short clean evidence: SQuAD 85.3 vs 51.8.
  • Only a categorical top-1 belief tested; continuous Bayesian version left untested.
  • Multi-turn results use authored conversations over pools built for them.
From the paper