{
  "abstract": "Reinforcement learning typically optimizes average reward. For generative policies, the average can hide an important distinction: two policies can achieve the same mean reward while having very different chances of producing a rare but high-reward rollout. This matters as sampling increases during training and inference, since its benefit depends on retaining probability mass on high-reward outcomes. We propose to optimize this coverage directly. Rather than considering only expected reward, we consider all of its upper tails: for each reward threshold, how likely is the policy to exceed it? This turns a continuous reward into a family of binary success events. We introduce Tail-Likelihood Reinforcement Learning (TailRL), which maximizes the log-probability of exceeding a randomly chosen reward threshold. Its gradient gives more weight to rare, high-reward rollouts and can be interpreted as a mixture of Best-of-k gradients. TailRL requires only a simple modification to the advantage function, making it compatible with existing reinforcement learning pipelines. Across object localization, maze navigation, GUI grounding, and code optimization, TailRL leverages rare high-reward training samples to avoid suboptimal solutions and yields models that benefit more from additional samples at inference time.",
  "abstract_shown": true,
  "affiliations": [],
  "article_type": null,
  "arxiv": "2609.02987",
  "author_total": 14,
  "authors": [
    {
      "name": "Shrinivas Ramasubramanian"
    },
    {
      "name": "Daman Arora"
    },
    {
      "name": "Fahim Tajwar"
    },
    {
      "name": "Guanning Zeng"
    },
    {
      "name": "Qingyang Wu"
    },
    {
      "name": "Zhongzhu Zhou"
    },
    {
      "name": "Chenfeng Xu"
    },
    {
      "name": "Haiwen Feng"
    },
    {
      "name": "Yuda Song"
    },
    {
      "name": "Aarti Singh"
    },
    {
      "name": "Ruslan Salakhutdinov"
    },
    {
      "name": "J. Andrew Bagnell"
    },
    {
      "name": "Jeff Schneider"
    },
    {
      "name": "Andrea Zanette"
    }
  ],
  "bibtex": "https://research.draconitic.com/paper/3czm7jfvpsmrm.bib",
  "cited_by": 0,
  "doi": "10.48550/arxiv.2609.02987",
  "fulltext_url": null,
  "id": "3czm7jfvpsmrm",
  "issns": [],
  "landing_url": "https://arxiv.org/abs/2609.02987",
  "language": null,
  "licence": "CC-BY-4.0",
  "licence_label": "CC BY 4.0",
  "licence_source": "article",
  "licence_url": "http://creativecommons.org/licenses/by/4.0/",
  "metadata_licence": "CC0-1.0",
  "native_subjects": [
    [
      "arxiv",
      "cs.LG"
    ],
    [
      "arxiv",
      "stat.ML"
    ],
    [
      "openalex",
      "1702|Reinforcement Learning in Robotics"
    ]
  ],
  "openalex": "W7208739650",
  "pdf_url": "https://arxiv.org/pdf/2609.02987",
  "pmcid": null,
  "pmid": null,
  "published": "2026-09-09",
  "publisher": null,
  "ris": "https://research.draconitic.com/paper/3czm7jfvpsmrm.ris",
  "sources": [
    {
      "first_seen": "2026-10-30T12:20:22Z",
      "native_id": "2609.02987",
      "source": "arxiv",
      "url": "https://arxiv.org/abs/2609.02987"
    }
  ],
  "subjects": [
    {
      "field": "mathematics",
      "origin": "arxiv:stat.ML",
      "subfield": "statistics"
    }
  ],
  "title": "Tail-Likelihood Reinforcement Learning",
  "title_text": "Tail-Likelihood Reinforcement Learning",
  "type": "preprint",
  "updated_at": "2026-10-30T12:20:22Z",
  "url": "https://research.draconitic.com/paper/3czm7jfvpsmrm",
  "venue": "arXiv",
  "versions": []
}