{
    "componentChunkName": "component---gatsby-theme-academic-src-templates-tags-index-jsx",
    "path": "/tags/Efficient Inference",
    "result": {"data":{"allTag":{"edges":[{"node":{"color":"magenta","name":"Loop Models","description":null,"path":"/tags/Loop Models"}},{"node":{"color":"gold","name":"Fixed Points","description":null,"path":"/tags/Fixed Points"}},{"node":{"color":"orange","name":"Efficient Inference","description":null,"path":"/tags/Efficient Inference"}},{"node":{"color":"lime","name":"LLM Pre-training","description":null,"path":"/tags/LLM Pre-training"}},{"node":{"color":"lime","name":"Mixture-of-Experts","description":null,"path":"/tags/Mixture-of-Experts"}},{"node":{"color":"green","name":"Scalable Reasoning","description":null,"path":"/tags/Scalable Reasoning"}},{"node":{"color":"cyan","name":"Test-Time Compute","description":null,"path":"/tags/Test-Time Compute"}},{"node":{"color":"orange","name":"Equilibrium Models","description":null,"path":"/tags/Equilibrium Models"}},{"node":{"color":"green","name":"Parcae","description":null,"path":"/tags/Parcae"}},{"node":{"color":"orange","name":"Dynamical Systems","description":null,"path":"/tags/Dynamical Systems"}},{"node":{"color":"magenta","name":"Diffusion","description":null,"path":"/tags/Diffusion"}},{"node":{"color":"red","name":"LLM","description":null,"path":"/tags/LLM"}},{"node":{"color":"magenta","name":"Efficiency","description":null,"path":"/tags/Efficiency"}},{"node":{"color":"orange","name":"Data Attribution","description":null,"path":"/tags/Data Attribution"}},{"node":{"color":"cyan","name":"Survey","description":null,"path":"/tags/Survey"}},{"node":{"color":"gold","name":"GenAI","description":null,"path":"/tags/GenAI"}},{"node":{"color":"magenta","name":"Data-centric AI","description":null,"path":"/tags/Data-centric AI"}},{"node":{"color":"lime","name":"LLM Agent","description":null,"path":"/tags/LLM Agent"}},{"node":{"color":"magenta","name":"Benchmark","description":null,"path":"/tags/Benchmark"}},{"node":{"color":"volcano","name":"Adversarial Perturbations","description":null,"path":"/tags/Adversarial Perturbations"}},{"node":{"color":"volcano","name":"AI Safety","description":null,"path":"/tags/AI Safety"}},{"node":{"color":"geekblue","name":"writing","description":null,"path":"/tags/writing"}},{"node":{"color":"blue","name":"research","description":null,"path":"/tags/research"}},{"node":{"color":"volcano","name":"AI Interpretability","description":null,"path":"/tags/AI Interpretability"}},{"node":{"color":"red","name":"skills","description":null,"path":"/tags/skills"}},{"node":{"color":"red","name":"paper","description":null,"path":"/tags/paper"}},{"node":{"color":"blue","name":"influence function","description":null,"path":"/tags/influence function"}},{"node":{"color":"magenta","name":"World Model","description":null,"path":"/tags/World Model"}},{"node":{"color":"cyan","name":"Video Generation","description":null,"path":"/tags/Video Generation"}},{"node":{"color":"red","name":"Long-term Consistency","description":null,"path":"/tags/Long-term Consistency"}},{"node":{"color":"orange","name":"Equivariant Models","description":null,"path":"/tags/Equivariant Models"}},{"node":{"color":"gold","name":"Game Generation","description":null,"path":"/tags/Game Generation"}},{"node":{"color":"lime","name":"Image to Video","description":null,"path":"/tags/Image to Video"}},{"node":{"color":"lime","name":"ML","description":null,"path":"/tags/ML"}},{"node":{"color":"blue","name":"gatsby","description":null,"path":"/tags/gatsby"}},{"node":{"color":"cyan","name":"blog","description":null,"path":"/tags/blog"}},{"node":{"color":"blue","name":"test1","description":null,"path":"/tags/test1"}},{"node":{"color":"blue","name":"test2","description":null,"path":"/tags/test2"}},{"node":{"color":"geekblue","name":"test3","description":null,"path":"/tags/test3"}},{"node":{"color":"red","name":"3D-Reconstruction","description":null,"path":"/tags/3D-Reconstruction"}},{"node":{"color":"volcano","name":"CV","description":null,"path":"/tags/CV"}},{"node":{"color":"red","name":"AGI(Image)","description":null,"path":"/tags/AGI(Image)"}},{"node":{"color":"gold","name":"Evaluation","description":null,"path":"/tags/Evaluation"}},{"node":{"color":"green","name":"Model-Arch","description":null,"path":"/tags/Model-Arch"}},{"node":{"color":"volcano","name":"Coq","description":null,"path":"/tags/Coq"}},{"node":{"color":"gold","name":"FormalVerification","description":null,"path":"/tags/FormalVerification"}},{"node":{"color":"green","name":"RL","description":null,"path":"/tags/RL"}},{"node":{"color":"orange","name":"DDQN","description":null,"path":"/tags/DDQN"}},{"node":{"color":"lime","name":"MCTS","description":null,"path":"/tags/MCTS"}},{"node":{"color":"cyan","name":"Vue.js","description":null,"path":"/tags/Vue.js"}},{"node":{"color":"cyan","name":"echarts","description":null,"path":"/tags/echarts"}},{"node":{"color":"blue","name":"element-ui","description":null,"path":"/tags/element-ui"}},{"node":{"color":"volcano","name":"C++","description":null,"path":"/tags/C++"}},{"node":{"color":"gold","name":"Game Design","description":null,"path":"/tags/Game Design"}},{"node":{"color":"green","name":"React","description":null,"path":"/tags/React"}},{"node":{"color":"green","name":"MongoDB","description":null,"path":"/tags/MongoDB"}}]},"allMdx":{"edges":[{"node":{"frontmatter":{"cover":{"publicURL":"/static/d1391a1ae31785505e30a913e7e92d7e/preview.png"},"date":"2026-09-29","venue":"Institute of Foundation Models","conference":"","authors":["**Benhao Huang**","Chufan Shi","Junlin Chen","Shicheng Wen","Zhengzhong Liu","Eric Xing","Xuezhe Ma"],"path":"research/2026/looped-models-fixed-points","title":"Towards Looped Models Done Right — Part II: Rethinking at Fixed Points","tags":["Loop Models","Fixed Points","Efficient Inference"],"excerpt":"Looped LMs need not pay for every loop: near fixed points, they can scale FLOPs without scaling memory. Fixed-point shortcuts speed up training (pretraining, post-training) and inference (prefill, decoding), while a learned depth prior and orthogonal injection shape better fixed points. With a 3× smaller KV cache, the learned prior matches fixed-depth training at 1.6B.","highlight":null,"links":[{"name":"paper","url":"https://www.alphaxiv.org/abs/2610.looped-models-fixed-points"},{"name":"PDF","url":"https://github.com/ifm-ai/xllm-loop/blob/main/papers/part2.pdf"},{"name":"Github","url":"https://github.com/ifm-ai/xllm-loop"},{"name":"X thread","url":"https://x.com/huskydogewoof/status/2106130737007091939"},{"name":"Part I","url":"https://huskydoge.github.io/husky-blog/posts/recursive_models/towards-looped-models-done-right/"}],"type":"research"}}}]},"allFile":{"edges":[]}},"pageContext":{"fileAbsolutePath":"","tag":"Efficient Inference"}},
    "staticQueryHashes": ["3694448773","4097791827"]}