DiVR - MBPP: leaderboard

DiVR v1 released unmodified generators only, under native Greedy, Confidence-guided and ReMDM resampling conditions. Canvas 256 tokens except MATH-500 512; 64 inference steps, dataset-specific native evaluation. DiVR-guided and DiVR best-of-four require trained LoRA+MLP verifiers without affirmative public weight artifacts and are retained in the independent source-identity hold packet, never attributed to the frozen generator. Base and Instruct checkpoints stay distinct. Native source table scores/conditions retained. Held DiVR guided configurations use at most three regenerations, theta0.5 and4alternatives attemperature0.7; these trained-verifier settings are not assigned to the retained frozen-generator baselines. Metric: Native task accuracy (%)..

Metric: Native task accuracy (%).. Source: arxiv.org. Saturation forecast: Rough model projection: around 2026. 7 models tracked. Updated 14 Nov 2026.

Top models

#ModelScoreOverall rank
1Dream-v0-Instruct-7B (Harness: DiVR SR-SMC (4))50.2
2Dream-v0-Instruct-7B (Harness: DiVR Prism)46.4
3Dream-v0-Instruct-7B (Harness: DiVR Greedy)43.8
4LLaDA-8B-Instruct (Harness: DiVR SR-SMC (4))34
5LLaDA-8B-Instruct (Harness: DiVR Greedy)33.2
6LLaDA-8B-Instruct (Harness: DiVR Prism)31.6
7LLaDA-8B-Base (Harness: DiVR Greedy)30.2

Interactive version: theaggregate.ai/benchmark?slug=divr-mbpp · How It Works · Data refreshed daily, snapshot 2026-11-16.