# PRIME-RL/TTRL

[NeurIPS 2025] TTRL: Test-Time Reinforcement Learning

Repository: https://github.com/PRIME-RL/TTRL
Canonical: https://ross.abutalabs.com/products/ttrl
Homepage: https://arxiv.org/abs/2504.16084
Language: Python
License: MIT
License Family: permissive
Topics: llm, reasoning, rl
Last push: 2026-04-15T02:47:09+00:00

## Health v2 (maintenance only)
Score: 51/100 (v2, computed 2026-09-03T02:20:16.233290+00:00)
- activity 77, release rhythm 28, longevity 35
- inputs: {"age_days": 498, "days_push": 140, "days_rel": 418, "gap_med": 49, "n_releases_24m": 2}
- flags: none
- formula: round(0.45*activity + 0.35*rhythm + 0.20*longevity); archived -> min(score, 10)

## Adoption (not part of the score)
Stars 1120, forks 81 (observed 2026-08-28T04:03:39.696901+00:00)

## What it is
TTRL is an open-source implementation of Test-Time Reinforcement Learning, a method for training LLMs with RL on unlabeled data using majority voting as the reward signal. It is built on the verl framework and accompanies a NeurIPS 2025 paper.

## Use cases
- train an LLM with reinforcement learning without ground-truth labels
- improve math reasoning of a model using only unlabeled test data
- use majority voting as a reward signal for RL fine-tuning
- reproduce TTRL experiments on AIME 2024 with Qwen models
- run online RL on test data with verl integration

## When to choose
- you want to fine-tune LLMs on unlabeled reasoning data
- you need a research-grade implementation of test-time RL on verl
- you lack ground-truth labels but have model priors suitable for majority voting

## When to avoid
- you need a production RLHF pipeline with human-labeled rewards
- you lack GPU resources for RL training
- your task has no verifiable or majority-votable outputs

## Facets
- artifact type: library
- maturity: active
- function: llm-training, reinforcement-learning, machine-learning
- domain: large-language-models, machine-learning, deep-learning
- platform: python
- tags: test-time-reinforcement-learning, rlvr, majority-voting-reward, verl, reasoning, research-code, neurips-2025, gpu, linux

## Member repositories
- PRIME-RL/TTRL (main) score 51

## Provenance
- Observed fields: from GitHub, fetched 2026-08-28T04:03:39.696901+00:00.
- Health v2: computed from the inputs above; adoption is never an input.
- Inferred fields (summary, facets, guidance): AI-extracted, prompt v1, taxonomy v1, on 2026-08-30T06:41:04.137972+00:00, confidence not recorded.
  - readme: https://github.com/PRIME-RL/TTRL (fetched 2026-08-28T04:03:39.696901+00:00, sha 263f4092ee48)
  - homepage: https://arxiv.org/abs/2504.16084 (fetched 2026-08-29T12:45:04.848376+00:00, sha c876a60812af)
  - site_page: https://info.arxiv.org/about/ourmembers.html (fetched 2026-08-29T12:45:04.944705+00:00, sha 47cbc55ff1de)
  - site_page: https://info.arxiv.org/about/donate.html (fetched 2026-08-29T12:45:04.880727+00:00, sha cca9c3a11c56)
  - site_page: https://info.arxiv.org/about (fetched 2026-08-29T12:45:05.042729+00:00, sha a1f16f915a9a)
  - site_page: https://info.arxiv.org/labs/index.html (fetched 2026-08-29T12:45:04.912639+00:00, sha b14a8d05a0ec)
- Data as of 2026-08-30T08:39:29.467469+00:00.