Grafy News
Try Grafy free

Test‑time symmetrization reveals new scaling laws for model‑agnostic inference

A new arXiv paper shows that applying symmetry‑preserving rewrites at inference time yields predictable accuracy gains.

Illustration for the news story: Test‑time symmetrization reveals new scaling laws for model‑agnostic inferenceSource: Grafy Research Lab

Test‑time symmetrization, a technique that runs a model on several equivalent representations of the same problem and aggregates the results, has been shown to follow clear scaling laws, according to a paper posted on arXiv on 19 Nov 2026. The finding matters because it extends the well‑known test‑time compute improvements—where accuracy rises with extra forward passes—to scenarios where the input can be rewritten without changing its semantics.

What is test‑time symmetrization?

Test‑time symmetrization exploits the fact that many tasks admit many interchangeable forms. A graph, for example, is unchanged by renumbering its nodes, and a piece of code retains its meaning when variable names are swapped. The authors propose to generate a set of such symmetry‑preserving transformations, feed each version to the same language model, and then combine the predictions (e.g., by majority vote or probability averaging). The process is model‑agnostic: it works with any decoder that can be called repeatedly at inference.

How does it differ from existing test‑time compute scaling laws?

Previous scaling laws for test‑time compute measured accuracy as a function of the number of independent forward passes on an unchanged prompt. Those works assumed the prompt itself was fixed, so the only source of variation was stochastic sampling. In contrast, symmetrization introduces structured diversity: each pass sees a different, but mathematically equivalent, encoding of the problem. This additional diversity can reduce systematic biases that arise from the model’s sensitivity to superficial token choices.

What scaling behavior did the authors observe?

By running extensive experiments on graph‑reasoning, program‑synthesis, and textual reasoning benchmarks, the paper reports that the error rate E(N)E(N) after NN symmetrized evaluations follows a power‑law decay of roughly E(N)≈CNE(N) \approx \frac{C}{\sqrt{N}} where CC is a task‑dependent constant. The relationship mirrors classic ensemble theory but emerges without training multiple models. The authors also note diminishing returns beyond a modest number of symmetries (often 8–12), aligning with the 1/N1/\sqrt{N} trend.

Practical implications for creators and developers

  • More reliable answers – For tasks where correctness is critical (e.g., code generation, data‑structure queries), applying a handful of symmetry transforms can noticeably boost confidence without changing the underlying model.
  • Cost‑effective accuracy – Because the same model is reused, the approach avoids the expense of larger models while still gaining a measurable lift, which is valuable for budget‑constrained deployments.
  • Automation potential – The technique can be wrapped into a thin inference layer that automatically detects applicable symmetries (node relabeling for graphs, variable renaming for code, paraphrasing for text) and orchestrates the multiple calls.

How to experiment with test‑time symmetrization in Grafy

Grafy provides the tools to prototype this workflow without leaving the platform:

  1. Generate alternative representations – Use the Grafy Graph app to create different node labelings of a graph or the Grafy Code studio to rename variables programmatically.
  2. Run the model repeatedly – Invoke the same language model through Grafy Chat, feeding each transformed input and collecting the raw completions.
  3. Aggregate results – Export the outputs to a spreadsheet via Grafy Files or use a simple Python script in the Grafy Code environment to compute majority votes or probability averages.
  4. Inspect the scaling – Plot error versus the number of symmetrizations directly in the Grafy Graph view to verify the 1/N1/\sqrt{N} trend on your own data.

Researchers who prefer to read the full methodology can upload the PDF to Grafy Reader, which lets you ask natural‑language questions about the paper and retrieve specific sections instantly.


Source: https://arxiv.org/abs/2611.24032


Source: Grafy Research Lab — Test-Time Symmetrization: Scaling Laws for Model-Agnostic Learning under Structure