Absolute Zero: A self-generated curriculum without human data¶

Core candidate mechanism, not the full reasoning benchmark.

Paper information¶

Field	Details
Paper	Absolute Zero (arXiv 2505.03335)
Institution	Tsinghua University / LeapLab.
First public release	2025-05-06 (arXiv v1)
Author code	LeapLabTHU/Absolute-Zero-Reasoner.
Local adapter or method key	`absolute-zero`
Local reproduction code	`src/auto_research/post_training/`

Paper method and evidence¶

A proposer generates verifiable tasks, a solver attempts them, and an executable verifier supplies rewards. Adapt task difficulty to the current capability boundary.

flowchart LR
  P[Propose verifiable tasks] --> S[Solve]
  S --> V[Programmatic reward]
  V --> C[Adjust curriculum]
  C --> P

Original paper figure¶

The original figure shows the proposer–solver–verifier loop. From the original paper; copyright remains with the authors.

Core equation¶

\[ x\sim\pi_P,\ y\sim\pi_S(\cdot|x),\ r=V(x,y),\quad\max_{P,S}\mathbb E[r+\lambda H_{curriculum}(x)]. \]

Reported results¶

Absolute Zero Reasoner bootstraps code/math reasoning from zero data and beats multiple human-data RLVR baselines. No production A/B.

Local reproduction¶

Arithmetic candidates, 120 steps, 256 examples, seed 42: 0.2344→0.5938 (+153.33%) with matched reward, KL, length and candidate budgets.

auto-research post-train --algorithm absolute-zero --dataset arithmetic-smoke --maximum-examples 256 --steps 120 --seed 42
auto-research evolve --model post-training --dataset arithmetic-smoke --direction "Combine absolute-zero with installed paper operators" --generations 2 --population 4

Fixed-seed metrics. Only the method-specific objective, state updates and matched budget are tested. No large model, multi-GPU RL, private environment, live web or full benchmark.