Rather than treating training collapse as a black box, the paper analyzes <strong>token-level gradient dynamics</strong> and derives a simple taxonomy showing that the effect of an update depends jointly on the <strong>advantage sign</strong> and the token's probability under the current policy. This provides an intuitive explanation for why entropy sometimes collapses so abruptly.</p>\n<p>The resulting algorithm, <strong>Winner Advantage Policy Optimization (WAPO)</strong>, is almost surprisingly simple: perform clipped policy updates only on positive-advantage completions. Despite this small change, it consistently improves training stability while matching or exceeding GRPO-style baselines across math reasoning and multi-hop QA.</p>\n","updatedAt":"2026-06-17T02:00:39.159Z","author":{"_id":"65adf6fbae773a8112f4b9ad","avatarUrl":"/avatars/38d2eb81bf1f9b8c4a1903d21c48f365.svg","fullname":"Jesse Cresswell","name":"JesseCresswell","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8530111312866211},"editors":["JesseCresswell"],"editorAvatarUrls":["/avatars/38d2eb81bf1f9b8c4a1903d21c48f365.svg"],"reactions":[{"reaction":"👍","users":["satyakrishnagorti"],"count":1}],"isReport":false}}],"primaryEmailConfirmed":false,"paper":{"id":"2606.16154","authors":[{"_id":"6a31ba59bc818ff14e453c43","user":{"_id":"65a944aedf466514069f9055","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/wLYGl5acJXrMfN4WZ130v.jpeg","isPro":false,"fullname":"Prasanth YSS","user":"prasanthyss","type":"user","name":"prasanthyss"},"name":"Prasanth YSS","status":"claimed_verified","statusLastChangedAt":"2026-06-17T11:24:32.217Z","hidden":false},{"_id":"6a31ba59bc818ff14e453c44","name":"Zhichen Ren","hidden":false},{"_id":"6a31ba59bc818ff14e453c45","name":"Rasa Hosseinzadeh","hidden":false},{"_id":"6a31ba59bc818ff14e453c46","name":"Ilan Gofman","hidden":false},{"_id":"6a31ba59bc818ff14e453c47","name":"Yuqi Chen","hidden":false},{"_id":"6a31ba59bc818ff14e453c48","name":"Zhaoyan Liu","hidden":false},{"_id":"6a31ba59bc818ff14e453c49","name":"Guangwei Yu","hidden":false},{"_id":"6a31ba59bc818ff14e453c4a","user":{"_id":"65adf6fbae773a8112f4b9ad","avatarUrl":"/avatars/38d2eb81bf1f9b8c4a1903d21c48f365.svg","isPro":false,"fullname":"Jesse Cresswell","user":"JesseCresswell","type":"user","name":"JesseCresswell"},"name":"Jesse C. Cresswell","status":"claimed_verified","statusLastChangedAt":"2026-06-17T11:24:36.758Z","hidden":false},{"_id":"6a31ba59bc818ff14e453c4b","user":{"_id":"65bff4266a6b6de5e4193c30","avatarUrl":"/avatars/3ccee2824d36700edf840a6b83812a08.svg","isPro":false,"fullname":"Satya Krishna Gorti","user":"satyakrishnagorti","type":"user","name":"satyakrishnagorti"},"name":"Satya Krishna Gorti","status":"claimed_verified","statusLastChangedAt":"2026-06-17T11:24:34.582Z","hidden":false}],"publishedAt":"2026-06-15T00:00:00.000Z","submittedOnDailyAt":"2026-06-17T00:00:00.000Z","title":"A Gradient Perspective on RLVR Stability and Winner Advantage Policy Optimization","submittedOnDailyBy":{"_id":"65adf6fbae773a8112f4b9ad","avatarUrl":"/avatars/38d2eb81bf1f9b8c4a1903d21c48f365.svg","isPro":false,"fullname":"Jesse Cresswell","user":"JesseCresswell","type":"user","name":"JesseCresswell"},"summary":"Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but GRPO-style optimization remains prone to collapse. We analyse this instability through token-level gradient dynamics, deriving a taxonomy that predicts how updates affect next-token probabilities and entropy. The taxonomy shows that stability depends jointly on the advantage sign and token distribution under the current policy. Motivated by this finding, we propose Winner Advantage Policy Optimization (WAPO), a simple online clipped policy-gradient objective that updates only on positive-advantage completions. Across mathematical reasoning and multi-hop QA benchmarks, WAPO improves training stability and matches or outperforms baselines across multiple model families. Full code can be found at https://github.com/layer6ai-labs/wapo.","upvotes":6,"discussionId":"6a31ba59bc818ff14e453c4c","githubRepo":"https://github.com/layer6ai-labs/wapo","githubRepoAddedBy":"user","ai_summary":"Training instability in reinforcement learning with verifiable rewards is analyzed through token-level gradient dynamics, leading to a stable policy optimization method that updates only on positive-advantage completions.","ai_keywords":["reinforcement learning","verifiable rewards","language-model reasoning","GRPO-style optimization","token-level gradient dynamics","advantage sign","token distribution","policy-gradient objective","positive-advantage completions"],"ai_summary_model":"Qwen/Qwen2.5-Coder-32B-Instruct","githubStars":1,"organization":{"_id":"6435e64e2d0ed796668fac8d","name":"Layer6","fullname":"Layer 6 AI","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6435e521a4bd75c62cc0a64c/4S8zHoJeYjBq0L0IDW0nu.png"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"65adf6fbae773a8112f4b9ad","avatarUrl":"/avatars/38d2eb81bf1f9b8c4a1903d21c48f365.svg","isPro":false,"fullname":"Jesse Cresswell","user":"JesseCresswell","type":"user"},{"_id":"631109dbfa95534e2189dd9c","avatarUrl":"/avatars/a852580d729600bd26d21ff45ada45eb.svg","isPro":false,"fullname":"Rasa Hosseinzadeh","user":"rasaHusen","type":"user"},{"_id":"65bff4266a6b6de5e4193c30","avatarUrl":"/avatars/3ccee2824d36700edf840a6b83812a08.svg","isPro":false,"fullname":"Satya Krishna Gorti","user":"satyakrishnagorti","type":"user"},{"_id":"6509a3c26d2284f73246e79c","avatarUrl":"/avatars/9a0c8eed456e908ce2d65e5c1e57620d.svg","isPro":false,"fullname":"ilan gofman","user":"ilangofman","type":"user"},{"_id":"65a944aedf466514069f9055","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/wLYGl5acJXrMfN4WZ130v.jpeg","isPro":false,"fullname":"Prasanth YSS","user":"prasanthyss","type":"user"},{"_id":"66d8512c54209e9101811e8e","avatarUrl":"/avatars/62dfd8e6261108f2508efe678d5a2a57.svg","isPro":false,"fullname":"M Saad Salman","user":"MSS444","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"6435e64e2d0ed796668fac8d","name":"Layer6","fullname":"Layer 6 AI","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6435e521a4bd75c62cc0a64c/4S8zHoJeYjBq0L0IDW0nu.png"},"query":{}}">
A Gradient Perspective on RLVR Stability and Winner Advantage Policy Optimization
Abstract
Training instability in reinforcement learning with verifiable rewards is analyzed through token-level gradient dynamics, leading to a stable policy optimization method that updates only on positive-advantage completions.
Reinforcement learning with verifiable rewards (RLVR) improves language-model reasoning, but GRPO-style optimization remains prone to collapse. We analyse this instability through token-level gradient dynamics, deriving a taxonomy that predicts how updates affect next-token probabilities and entropy. The taxonomy shows that stability depends jointly on the advantage sign and token distribution under the current policy. Motivated by this finding, we propose Winner Advantage Policy Optimization (WAPO), a simple online clipped policy-gradient objective that updates only on positive-advantage completions. Across mathematical reasoning and multi-hop QA benchmarks, WAPO improves training stability and matches or outperforms baselines across multiple model families. Full code can be found at https://github.com/layer6ai-labs/wapo.
Community
Rather than treating training collapse as a black box, the paper analyzes token-level gradient dynamics and derives a simple taxonomy showing that the effect of an update depends jointly on the advantage sign and the token's probability under the current policy. This provides an intuitive explanation for why entropy sometimes collapses so abruptly.
The resulting algorithm, Winner Advantage Policy Optimization (WAPO), is almost surprisingly simple: perform clipped policy updates only on positive-advantage completions. Despite this small change, it consistently improves training stability while matching or exceeding GRPO-style baselines across math reasoning and multi-hop QA.
Upload images, audio, and videos by dragging in the text input, pasting, or clicking here.
Tap or paste here to upload images
Cite arxiv.org/abs/2606.16154 in a model README.md to link it from this page.
Cite arxiv.org/abs/2606.16154 in a dataset README.md to link it from this page.
Cite arxiv.org/abs/2606.16154 in a Space README.md to link it from this page.
Discussion (0)
Sign in to join the discussion. Free account, 30 seconds — email code or GitHub.
Sign in →No comments yet. Sign in and be the first to say something.