<strong>Project Page:</strong> <a href=\"https://syjmelody.github.io/RankE/\" rel=\"nofollow\">https://syjmelody.github.io/RankE/</a><br><strong>GitHub:</strong> <a href=\"https://github.com/syjmelody/RankE\" rel=\"nofollow\">https://github.com/syjmelody/RankE</a></p>\n<h2 class=\"relative group flex items-baseline\">\n\t<a id=\"⚡-tldr\" class=\"block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full\" href=\"#⚡-tldr\" rel=\"nofollow\">\n\t\t<span class=\"header-link\"><svg class=\"text-gray-500 hover:text-black dark:hover:text-gray-200 w-4\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" aria-hidden=\"true\" role=\"img\" width=\"1em\" height=\"1em\" preserveAspectRatio=\"xMidYMid meet\" viewBox=\"0 0 256 256\"><path d=\"M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z\" fill=\"currentColor\"></path></svg></span>\n\t</a>\n\t<span>\n\t\t⚡ TL;DR\n\t</span>\n</h2>\n<p>RankE is the <strong>first end-to-end post-training framework for discrete text-to-image generation</strong> that jointly optimizes the <strong>Generator</strong> and the <strong>Decoder</strong>. Instead of improving reward scores at the cost of image quality, RankE improves both alignment and fidelity at the same time.</p>\n<h2 class=\"relative group flex items-baseline\">\n\t<a id=\"🤔-background-what-is-the-problem\" class=\"block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full\" href=\"#🤔-background-what-is-the-problem\" rel=\"nofollow\">\n\t\t<span class=\"header-link\"><svg class=\"text-gray-500 hover:text-black dark:hover:text-gray-200 w-4\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" aria-hidden=\"true\" role=\"img\" width=\"1em\" height=\"1em\" preserveAspectRatio=\"xMidYMid meet\" viewBox=\"0 0 256 256\"><path d=\"M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z\" fill=\"currentColor\"></path></svg></span>\n\t</a>\n\t<span>\n\t\t🤔 Background: What is the problem?\n\t</span>\n</h2>\n<p>Most discrete text-to-image models still follow a two-stage pipeline:</p>\n<ol>\n<li>train a VQ-VAE / tokenizer to map images into discrete visual tokens;</li>\n<li>train an autoregressive Generator to model those tokens.</li>\n</ol>\n<p>This pipeline works well for pretraining, but post-training is usually incomplete: existing methods optimize the <strong>Generator only</strong> and keep the <strong>Decoder frozen</strong>.</p>\n<p>That creates a mismatch. As the Generator is optimized to chase higher rewards, its output token distribution gradually drifts away from the real token distribution that the Decoder saw during tokenizer training. The result is a frustrating trade-off:</p>\n<ul>\n<li>reward scores go up,</li>\n<li>but decoded image quality can get worse.</li>\n</ul>\n<p>The paper identifies this issue as <strong>Latent Covariate Shift</strong>.</p>\n<h2 class=\"relative group flex items-baseline\">\n\t<a id=\"🔍-why-existing-solutions-are-not-enough\" class=\"block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full\" href=\"#🔍-why-existing-solutions-are-not-enough\" rel=\"nofollow\">\n\t\t<span class=\"header-link\"><svg class=\"text-gray-500 hover:text-black dark:hover:text-gray-200 w-4\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" aria-hidden=\"true\" role=\"img\" width=\"1em\" height=\"1em\" preserveAspectRatio=\"xMidYMid meet\" viewBox=\"0 0 256 256\"><path d=\"M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z\" fill=\"currentColor\"></path></svg></span>\n\t</a>\n\t<span>\n\t\t🔍 Why existing solutions are not enough\n\t</span>\n</h2>\n<p>Recent work such as <strong>REPA-E</strong> has shown that in continuous diffusion models, the autoencoder is not just a supporting module — it can be a real bottleneck for alignment and visual quality.</p>\n<p>But discrete T2I is harder.<br>Because token sampling and vector quantization are discrete operations, gradients cannot flow cleanly through the entire generation process. That is why most existing RL or preference-optimization methods for discrete generation still update only the Generator while leaving the Decoder unchanged.</p>\n<p>So the field already knows the decoder matters — but a practical end-to-end solution for discrete generation has been missing.</p>\n<h2 class=\"relative group flex items-baseline\">\n\t<a id=\"🚀-what-ranke-does\" class=\"block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full\" href=\"#🚀-what-ranke-does\" rel=\"nofollow\">\n\t\t<span class=\"header-link\"><svg class=\"text-gray-500 hover:text-black dark:hover:text-gray-200 w-4\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" aria-hidden=\"true\" role=\"img\" width=\"1em\" height=\"1em\" preserveAspectRatio=\"xMidYMid meet\" viewBox=\"0 0 256 256\"><path d=\"M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z\" fill=\"currentColor\"></path></svg></span>\n\t</a>\n\t<span>\n\t\t🚀 What RankE does\n\t</span>\n</h2>\n<p>RankE addresses this directly by making the <strong>Generator and Decoder co-evolve</strong>.</p>\n<p>Its core insight is simple:<br><strong>if Generator optimization already behaves like a ranking process over latent token sequences, why not extend the same ranking principle to pixel-space Decoder optimization?</strong></p>\n<p>RankE therefore uses <strong>alternating optimization</strong>:</p>\n<ul>\n<li><strong>Generator step:</strong> optimize the policy so that higher-reward latent token sequences receive stronger updates;</li>\n<li><strong>Decoder step:</strong> optimize the Decoder so it can better adapt to the Generator’s evolving token distribution, while also favoring higher-reward decoded images.</li>\n</ul>\n<p>In other words, RankE does not just make the model “better at scoring well.” It aligns optimization across <strong>both latent space and pixel space</strong>.</p>\n<p>This is the key difference from standard frozen-decoder RL.</p>\n<h2 class=\"relative group flex items-baseline\">\n\t<a id=\"🧠-why-this-matters\" class=\"block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full\" href=\"#🧠-why-this-matters\" rel=\"nofollow\">\n\t\t<span class=\"header-link\"><svg class=\"text-gray-500 hover:text-black dark:hover:text-gray-200 w-4\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" aria-hidden=\"true\" role=\"img\" width=\"1em\" height=\"1em\" preserveAspectRatio=\"xMidYMid meet\" viewBox=\"0 0 256 256\"><path d=\"M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z\" fill=\"currentColor\"></path></svg></span>\n\t</a>\n\t<span>\n\t\t🧠 Why this matters\n\t</span>\n</h2>\n<p>In standard RL post-training for discrete T2I, the Generator keeps changing, but the Decoder stays fixed. Over time, the Decoder is forced to decode token patterns it was never really trained to handle.</p>\n<p>RankE removes this bottleneck by continuously adapting the Decoder during post-training. This turns reward optimization into actual visual improvement, rather than reward hacking in latent space.</p>\n<h2 class=\"relative group flex items-baseline\">\n\t<a id=\"📈-results\" class=\"block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full\" href=\"#📈-results\" rel=\"nofollow\">\n\t\t<span class=\"header-link\"><svg class=\"text-gray-500 hover:text-black dark:hover:text-gray-200 w-4\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" aria-hidden=\"true\" role=\"img\" width=\"1em\" height=\"1em\" preserveAspectRatio=\"xMidYMid meet\" viewBox=\"0 0 256 256\"><path d=\"M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z\" fill=\"currentColor\"></path></svg></span>\n\t</a>\n\t<span>\n\t\t📈 Results\n\t</span>\n</h2>\n<p>The gains are clear.</p>\n<p>On <strong>LlamaGen-XL (775M)</strong> under CLIP-based optimization:</p>\n<ul>\n<li><strong>Standard RL:</strong> improves CLIP, but hurts FID</li>\n<li><strong>RankE:</strong> improves both</li>\n</ul>\n<p>Specifically:</p>\n<ul>\n<li><strong>CLIP:</strong> 32.45 → <strong>33.76</strong></li>\n<li><strong>FID:</strong> 17.76 → <strong>15.21</strong></li>\n</ul>\n<p>That is the main message of the paper:<br><strong>RankE breaks the common fidelity–alignment trade-off in discrete text-to-image post-training.</strong></p>\n<p>The improvements are also consistent across:</p>\n<ul>\n<li>different backbones,</li>\n<li>different reward functions,</li>\n<li>and multiple evaluation settings.</li>\n</ul>\n<h2 class=\"relative group flex items-baseline\">\n\t<a id=\"✨-one-line-takeaway\" class=\"block pr-1.5 text-lg md:absolute md:p-1.5 md:opacity-0 md:group-hover:opacity-100 md:right-full\" href=\"#✨-one-line-takeaway\" rel=\"nofollow\">\n\t\t<span class=\"header-link\"><svg class=\"text-gray-500 hover:text-black dark:hover:text-gray-200 w-4\" xmlns=\"http://www.w3.org/2000/svg\" xmlns:xlink=\"http://www.w3.org/1999/xlink\" aria-hidden=\"true\" role=\"img\" width=\"1em\" height=\"1em\" preserveAspectRatio=\"xMidYMid meet\" viewBox=\"0 0 256 256\"><path d=\"M167.594 88.393a8.001 8.001 0 0 1 0 11.314l-67.882 67.882a8 8 0 1 1-11.314-11.315l67.882-67.881a8.003 8.003 0 0 1 11.314 0zm-28.287 84.86l-28.284 28.284a40 40 0 0 1-56.567-56.567l28.284-28.284a8 8 0 0 0-11.315-11.315l-28.284 28.284a56 56 0 0 0 79.196 79.197l28.285-28.285a8 8 0 1 0-11.315-11.314zM212.852 43.14a56.002 56.002 0 0 0-79.196 0l-28.284 28.284a8 8 0 1 0 11.314 11.314l28.284-28.284a40 40 0 0 1 56.568 56.567l-28.285 28.285a8 8 0 0 0 11.315 11.314l28.284-28.284a56.065 56.065 0 0 0 0-79.196z\" fill=\"currentColor\"></path></svg></span>\n\t</a>\n\t<span>\n\t\t✨ One-line takeaway\n\t</span>\n</h2>\n<p><strong>RankE is a more natural way to post-train discrete text-to-image models: instead of optimizing only the Generator, it lets the Generator and Decoder improve together.</strong></p>\n","updatedAt":"2026-05-22T15:18:14.509Z","author":{"_id":"673ee069513dd08f78fde43d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/673ee069513dd08f78fde43d/-NyOWigCOjUCQR8O6Dm61.jpeg","fullname":"Siyong Jian","name":"syjian","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"identifiedLanguage":{"language":"en","probability":0.8599147200584412},"editors":["syjian"],"editorAvatarUrls":["https://cdn-avatars.huggingface.co/v1/production/uploads/673ee069513dd08f78fde43d/-NyOWigCOjUCQR8O6Dm61.jpeg"],"reactions":[{"reaction":"🔥","users":["Lupin1998","Jerry-98"],"count":2},{"reaction":"👍","users":["Lupin1998","Jerry-98"],"count":2},{"reaction":"🚀","users":["Lupin1998"],"count":1}],"isReport":false}},{"id":"6a13c07cc134b7b3c10a159f","author":{"_id":"673ee069513dd08f78fde43d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/673ee069513dd08f78fde43d/-NyOWigCOjUCQR8O6Dm61.jpeg","fullname":"Siyong Jian","name":"syjian","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false},"createdAt":"2026-05-25T03:22:36.000Z","type":"comment","data":{"edited":true,"hidden":true,"hiddenBy":"","hiddenReason":"Resolved","latest":{"raw":"This comment has been hidden","html":"This comment has been hidden","updatedAt":"2026-05-25T03:23:46.073Z","author":{"_id":"673ee069513dd08f78fde43d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/673ee069513dd08f78fde43d/-NyOWigCOjUCQR8O6Dm61.jpeg","fullname":"Siyong Jian","name":"syjian","type":"user","isPro":false,"isHf":false,"isHfAdmin":false,"isMod":false,"followerCount":1,"isUserFollowing":false}},"numEdits":0,"editors":[],"editorAvatarUrls":[],"reactions":[]}}],"primaryEmailConfirmed":false,"paper":{"id":"2605.21195","authors":[{"_id":"6a1057bfa53a61ce2e422fdf","user":{"_id":"673ee069513dd08f78fde43d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/673ee069513dd08f78fde43d/-NyOWigCOjUCQR8O6Dm61.jpeg","isPro":false,"fullname":"Siyong Jian","user":"syjian","type":"user","name":"syjian"},"name":"Siyong Jian","status":"claimed_verified","statusLastChangedAt":"2026-05-22T15:58:59.915Z","hidden":false},{"_id":"6a1057bfa53a61ce2e422fe0","name":"Siyuan Li","hidden":false},{"_id":"6a1057bfa53a61ce2e422fe1","name":"Luyuan Zhang","hidden":false},{"_id":"6a1057bfa53a61ce2e422fe2","name":"Zedong Wang","hidden":false},{"_id":"6a1057bfa53a61ce2e422fe3","name":"Xin Jin","hidden":false},{"_id":"6a1057bfa53a61ce2e422fe4","name":"Ying Li","hidden":false},{"_id":"6a1057bfa53a61ce2e422fe5","name":"Cheng Tan","hidden":false},{"_id":"6a1057bfa53a61ce2e422fe6","name":"Huan Wang","hidden":false}],"publishedAt":"2026-05-20T00:00:00.000Z","submittedOnDailyAt":"2026-05-25T00:00:00.000Z","title":"RankE: End-to-End Post-Training for Discrete Text-to-Image Generation with Decoder Co-Evolution","submittedOnDailyBy":{"_id":"673ee069513dd08f78fde43d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/673ee069513dd08f78fde43d/-NyOWigCOjUCQR8O6Dm61.jpeg","isPro":false,"fullname":"Siyong Jian","user":"syjian","type":"user","name":"syjian"},"summary":"Discrete autoregressive (AR) text-to-image (T2I) models pair a VQ tokenizer with an AR policy, and current post-training pipelines optimize only the policy while keeping the VQ decoder frozen. Recent diffusion T2I work, exemplified by REPA-E, has shown that the VAE itself constitutes a key alignment bottleneck, yet no analogous investigation exists for discrete AR models. We show that policy-only optimization induces Latent Covariate Shift: as the policy evolves, the resulting token distribution diverges from the ground-truth distribution on which the decoder was trained, such that reward scores improve while decoded image quality degrades. To address this mismatch, we propose RankE, the first end-to-end post-training framework for discrete T2I generation. Rather than optimizing the policy against a fixed decoder, RankE co-evolves both components through alternating optimization: each module maximizes a ranking-based alignment objective while being regularized by a stability-preserving anchor suited to its parameter space. This co-evolution breaks the fidelity--alignment trade-off that plagues frozen-decoder approaches: on LlamaGen-XL (775M), standard RL improves CLIP but degrades FID, whereas RankE improves both simultaneously (FID 15.21, CLIP 33.76 on MS-COCO 30K). Consistent gains on Janus-Pro (1B) confirm that decoder co-evolution reliably converts reward optimization into pixel-space quality improvements.","upvotes":13,"discussionId":"6a1057c0a53a61ce2e422fe7","projectPage":"https://arxiv.org/pdf/2605.21195","githubRepo":"https://github.com/syjmelody/RankE","githubRepoAddedBy":"user","ai_summary":"Discrete autoregressive text-to-image models suffer from latent covariate shift during policy optimization, which RankE addresses through end-to-end co-evolution of policy and decoder components.","ai_keywords":["VQ tokenizer","autoregressive policy","VAE","latent covariate shift","RankE","alternating optimization","ranking-based alignment objective","stability-preserving anchor","CLIP","FID"],"githubStars":14,"organization":{"_id":"643cb10025681c3afab0f1a6","name":"Westlake-University","fullname":"Westlake University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6437eca0819f3ab20d162e14/SQRCHUyjPRyqdtV3um42X.jpeg"}},"canReadDatabase":false,"canManagePapers":false,"canSubmit":false,"hasHfLevelAccess":false,"upvoted":false,"upvoters":[{"_id":"66aa39349238d9c3a1c7f9dc","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/66aa39349238d9c3a1c7f9dc/mj6r7uxEYXM502x296UMf.jpeg","isPro":false,"fullname":"Xin Jin","user":"Xin1118","type":"user"},{"_id":"673ee069513dd08f78fde43d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/673ee069513dd08f78fde43d/-NyOWigCOjUCQR8O6Dm61.jpeg","isPro":false,"fullname":"Siyong Jian","user":"syjian","type":"user"},{"_id":"67d5848f179ad2756600eca3","avatarUrl":"/avatars/158168a753271b6e024e1fbdf52c9e73.svg","isPro":false,"fullname":"Junhan ZHU","user":"Alrightlone","type":"user"},{"_id":"6694aa3f286045211d4b86dd","avatarUrl":"/avatars/b6fd2f95264f710039d1a7a497be6ed2.svg","isPro":false,"fullname":"graenys","user":"graenys","type":"user"},{"_id":"66ebba2b1a537888d23af2b9","avatarUrl":"/avatars/83cd3c87e8c21f5dcc145b9556b8bbfe.svg","isPro":false,"fullname":"Xueyi Chen","user":"YIGE24","type":"user"},{"_id":"66966286ad7167254c4bb5d6","avatarUrl":"/avatars/1a3136918a74d7ce778dcee0ca93c411.svg","isPro":false,"fullname":"Kele Shao","user":"cokeshao","type":"user"},{"_id":"62b624f3b52bef716e248fd7","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/62b624f3b52bef716e248fd7/AllcccKH-eBWduA8KVnOQ.png","isPro":false,"fullname":"Huan Wang","user":"Huan-WhoRegisteredMyName","type":"user"},{"_id":"691469212ac382b46b1f1d47","avatarUrl":"/avatars/3c8b5e6b1fe8e6071e129aace2ed910b.svg","isPro":false,"fullname":"fate bai","user":"ffffate","type":"user"},{"_id":"6793290257b2fe2b1ef434b5","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/no-auth/COZJKTHI4RPqjs7h9iz19.png","isPro":false,"fullname":"Billy","user":"Billy1377","type":"user"},{"_id":"640f7083208821a59b74c757","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/1678735253848-640f7083208821a59b74c757.jpeg","isPro":false,"fullname":"Siyuan Li","user":"Lupin1998","type":"user"},{"_id":"683f2e9fa073d45457ce420d","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/g2WWieHoqAeG8gb1qWL5J.png","isPro":false,"fullname":"Jason Lee","user":"Jerry-98","type":"user"},{"_id":"69a12e3a663af458fd86167c","avatarUrl":"https://cdn-avatars.huggingface.co/v1/production/uploads/noauth/7GMxFuHV0GfbCxtl3nVfS.png","isPro":false,"fullname":"Cheng Lizhong","user":"LZChen9917","type":"user"}],"acceptLanguages":["en"],"dailyPaperRank":0,"organization":{"_id":"643cb10025681c3afab0f1a6","name":"Westlake-University","fullname":"Westlake University","avatar":"https://cdn-avatars.huggingface.co/v1/production/uploads/6437eca0819f3ab20d162e14/SQRCHUyjPRyqdtV3um42X.jpeg"},"markdownContentUrl":"https://huggingface.co/buckets/huggingchat/papers-content/resolve/2605/2605.21195.md"}">

Get this paper in your agent:

Don't have the latest CLI?

Cite arxiv.org/abs/2605.21195 in a model README.md to link it from this page.

Cite arxiv.org/abs/2605.21195 in a dataset README.md to link it from this page.

Cite arxiv.org/abs/2605.21195 in a Space README.md to link it from this page.

Add this paper to a collection to link it from this page.