[{"data":1,"prerenderedAt":-1},["ShallowReactive",2],{"article-program-as-weights-fuzzy-functions-en":3,"article-related-program-as-weights-fuzzy-functions-en":30,"series-research-93228acd-047c-403b-bbbb-15e1498522df":73},{"id":4,"slug":5,"title":6,"content":7,"summary":8,"source":9,"source_url":10,"author":11,"image_url":12,"cover_image":12,"category":13,"language":14,"translated_content":11,"related_article_id":15,"keywords":16,"key_takeaways":22,"views":26,"created_at":27,"published_at":28,"topic_cluster_id":29},"93228acd-047c-403b-bbbb-15e1498522df","program-as-weights-fuzzy-functions-en","Program-as-Weights turns prompts into reusable tools","\u003Cp data-speakable=\"summary\">PAW compiles natural-language task specs into small local neural artifacts that run cheaply and offline.\u003C\u002Fp>\u003Cul>\u003Cli>\u003Cstrong>Research org\u003C\u002Fstrong>: Unspecified in arXiv abstract\u003C\u002Fli>\u003Cli>\u003Cstrong>Core data\u003C\u002Fstrong>: 10M-example FuzzyBench dataset\u003C\u002Fli>\u003Cli>\u003Cstrong>Breakthrough\u003C\u002Fstrong>: Compile natural-language specs into parameter-efficient adapters\u003C\u002Fli>\u003C\u002Ful>\u003Cp>Some programming tasks do not fit neat rules, and that is exactly where teams often reach for large language model APIs. This paper argues for a different pattern: instead of asking a model to solve every input from scratch, compile the task once into a reusable artifact that can run locally.\u003C\u002Fp>\u003Cp>The practical appeal is straightforward. If the same fuzzy task appears over and over—think log filtering, malformed JSON repair, or intent-based ranking—you want something reproducible, cheaper to run, and less dependent on remote \u003Ca href=\"\u002Ftag\u002Finference\">inference\u003C\u002Fa>. Program-as-Weights, or PAW, is the paper’s answer to that problem.\u003C\u002Fp>\u003Ch2>What problem this paper is trying to fix\u003C\u002Fh2>\u003Cp>The abstract focuses on “everyday programming tasks” that resist clean rule-based implementation. These are tasks where the edge cases are messy and the logic is hard to encode directly, so developers often outsource them to \u003Ca href=\"\u002Ftag\u002Fllm\">LLM\u003C\u002Fa> APIs.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783062181621-asl2.png\" alt=\"Program-as-Weights turns prompts into reusable tools\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>That approach works, but the paper calls out three practical costs: locality, reproducibility, and price. In other words, if every call goes back to a hosted model, you pay on latency, infrastructure dependence, and operational consistency.\u003C\u002Fp>\u003Cp>So the paper reframes the task. Instead of treating the foundation model as a per-input problem solver, it treats it as a tool builder. The model is invoked once per function definition, not once per function call.\u003C\u002Fp>\u003Ch2>How PAW works in plain English\u003C\u002Fh2>\u003Cp>The broader idea is called fuzzy-function programming: compile a natural-language specification into a compact neural artifact that can be executed locally. That is the main shift. The “program” is no longer just source code in a traditional language; it becomes weights.\u003C\u002Fp>\u003Cp>PAW instantiates that idea with two parts. First, a 4B compiler is trained on FuzzyBench, the 10M-example dataset the authors release. Second, that compiler emits parameter-efficient adapters for a frozen, lightweight interpreter.\u003C\u002Fp>\u003Cp>In plain English, the compiler reads the natural-language description of the task and produces a small model add-on. The interpreter then executes that add-on. Because the interpreter is frozen and lightweight, the resulting artifact is meant to be reusable and cheap to run after the one-time compilation step.\u003C\u002Fp>\u003Cp>This is not the same as asking a general-purpose model to answer every request directly. The paper’s framing is closer to “compile a function, then call the function” than “prompt a model every time.”\u003C\u002Fp>\u003Ch2>What the paper actually shows\u003C\u002Fh2>\u003Cp>The concrete result in the abstract is strong: a 0.6B Qwen3 interpreter executing PAW programs matches the performance of direct prompting of Qwen3-32B. That is the headline comparison, and it is the main evidence that the approach can preserve capability while shrinking runtime cost.\u003C\u002Fp>\n\u003Cfigure class=\"my-6\">\u003Cimg src=\"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783062178588-5pjm.png\" alt=\"Program-as-Weights turns prompts into reusable tools\" class=\"rounded-xl w-full\" loading=\"lazy\" \u002F>\u003C\u002Ffigure>\n\u003Cp>The resource story matters as much as the quality story. The paper says the PAW setup uses roughly one fiftieth of the inference memory and runs at 30 tokens per second on a MacBook M3. For developers, that is the kind of detail that changes whether a workflow stays in the cloud or can move onto a local machine.\u003C\u002Fp>\u003Cp>There are no additional \u003Ca href=\"\u002Ftag\u002Fbenchmark\">benchmark\u003C\u002Fa> tables or task-specific scores in the abstract, so the public evidence here is limited to that high-level comparison. If you are looking for a full matrix of datasets, failure modes, or ablation results, the abstract does not provide them.\u003C\u002Fp>\u003Cp>Still, the result is enough to show the intended tradeoff. PAW is not trying to beat frontier models on every task by brute force. It is trying to compress a task into a reusable form that makes repeated execution much cheaper.\u003C\u002Fp>\u003Ch2>Why developers should care\u003C\u002Fh2>\u003Cp>This is interesting if you build systems where the same ambiguous transformation gets applied many times. Examples in the abstract include alerting on important log lines, repairing malformed JSON, and ranking search results by intent. Those are exactly the kinds of tasks where a prompt-only setup can become expensive or brittle.\u003C\u002Fp>\u003Cp>PAW suggests a different deployment pattern: use a larger model once to synthesize a task-specific artifact, then run that artifact locally. That could make some LLM workflows more reproducible, since the behavior is tied to a compiled artifact rather than a live remote model that may shift over time.\u003C\u002Fp>\u003Cp>It also changes the economics of “AI features.” If the function can be compiled into a small adapter and executed on-device or on a local server, you may not need to \u003Ca href=\"\u002Fnews\u002Fcloudflare-monetization-gateway-pay-per-use-en\">pay per\u003C\u002Fa> request to a large hosted model. The paper’s memory and speed claims point in that direction.\u003C\u002Fp>\u003Ch2>What is still unclear\u003C\u002Fh2>\u003Cp>The abstract gives a compelling direction, but it leaves important questions open. It does not spell out how broad the task coverage is beyond the examples mentioned, how robust the compiled artifacts are to distribution shift, or how much manual effort is required to define the natural-language specification well.\u003C\u002Fp>\u003Cp>It also does not tell us how the approach behaves when the task itself changes over time. A compiled artifact is attractive when a function is stable, but less obviously useful when the policy or ranking logic is constantly evolving.\u003C\u002Fp>\u003Cp>Another open question is operational complexity. The method introduces a compiler, a dataset, and a frozen interpreter, which may be a good trade if you have repeated use cases—but it is still more moving parts than a single prompt to an \u003Ca href=\"\u002Ftag\u002Fapi\">API\u003C\u002Fa>.\u003C\u002Fp>\u003Ch2>The bottom line\u003C\u002Fh2>\u003Cp>PAW is a proposal to make fuzzy tasks look more like software and less like one-off prompts. The paper’s central claim is that a foundation model can be used as a compiler that produces a compact, reusable neural program instead of solving each request directly.\u003C\u002Fp>\u003Cp>For engineers, the promise is easy to understand: lower memory, local execution, and a more reusable artifact for messy tasks that do not fit clean rules. The abstract’s results suggest that this is not just a conceptual shift, but one that can preserve performance while cutting runtime cost.\u003C\u002Fp>\u003Cul>\u003Cli>PAW compiles natural-language specs into reusable neural artifacts.\u003C\u002Fli>\u003Cli>A 0.6B interpreter matches direct prompting of Qwen3-32B in the abstract.\u003C\u002Fli>\u003Cli>The approach targets local, reproducible, cheaper execution for fuzzy tasks.\u003C\u002Fli>\u003C\u002Ful>","PAW compiles natural-language task specs into small local neural artifacts that run cheaply and offline.","arxiv.org","https:\u002F\u002Farxiv.org\u002Fabs\u002F2607.02512",null,"https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783062181621-asl2.png","research","en","6cfddc0d-ce6e-4a14-baf7-3531bf32bc5d",[17,18,19,20,21],"LLM systems","neural compilation","local inference","parameter-efficient adapters","fuzzy functions",[23,24,25],"Compiles natural-language task specs into reusable neural programs.","Matches direct prompting of a much larger model in the abstract.","Aims to reduce memory use and enable local offline execution.",0,"2026-07-03T07:02:33.067987+00:00","2026-07-03T07:02:33.06+00:00","3103988e-c4fe-45e3-98ab-846500c9d507",{"tags":31,"relatedLang":32,"relatedPosts":36},[],{"id":15,"slug":33,"title":34,"language":35},"program-as-weights-fuzzy-functions-zh","PAW把提示詞編成可重用工具","zh",[37,43,49,55,61,67],{"id":38,"slug":39,"title":40,"cover_image":41,"image_url":41,"created_at":42,"category":13},"ca28a691-10df-40cc-86fa-4684b467c452","deepspec-data-regeneration-pipeline-qwen3-eagle3-en","DeepSpec should be treated as a data-regeneration pipeline, not a tra…","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783080168119-497r.png","2026-07-03T12:02:18.983093+00:00",{"id":44,"slug":45,"title":46,"cover_image":47,"image_url":47,"created_at":48,"category":13},"7c19a29b-70e8-4982-8b8d-9fff544d2984","lacuna-llm-unlearning-localization-testbed-en","LACUNA tests whether LLM unlearning really erases","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783060371861-ttwp.png","2026-07-03T06:32:31.852501+00:00",{"id":50,"slug":51,"title":52,"cover_image":53,"image_url":53,"created_at":54,"category":13},"6d92fdd6-ff88-4338-b7e5-00a05307d338","persistent-state-ai-agents-attack-surface-en","Persistent-state AI agents open a new attack surface","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1783058578300-fjsw.png","2026-07-03T06:02:30.780677+00:00",{"id":56,"slug":57,"title":58,"cover_image":59,"image_url":59,"created_at":60,"category":13},"cc12b2b9-0f6f-4dbf-8e2e-49d52008dda2","language-critiques-imitation-learning-en","Language critiques improve imitation learning","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782975783575-ibss.png","2026-07-02T07:02:29.283153+00:00",{"id":62,"slug":63,"title":64,"cover_image":65,"image_url":65,"created_at":66,"category":13},"b8167640-c431-4064-be79-10c877d15087","one-transformer-layer-can-carry-rl-gains-en","One Transformer Layer Can Carry RL Gains","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782973978159-8klr.png","2026-07-02T06:32:29.644564+00:00",{"id":68,"slug":69,"title":70,"cover_image":71,"image_url":71,"created_at":72,"category":13},"8d35bb8a-3563-4ac6-8c45-745d4e606f7f","bineval-binary-questions-llm-evals-en","BINEVAL uses binary questions to score LLM outputs","https:\u002F\u002Fxxdpdyhzhpamafnrdkyq.supabase.co\u002Fstorage\u002Fv1\u002Fobject\u002Fpublic\u002Fcovers\u002Finline-1782927166631-h8c1.png","2026-07-01T17:32:24.15899+00:00",[74,79,84,89,94,99,104,109,114,119],{"id":75,"slug":76,"title":77,"created_at":78},"a2715e72-1fe8-41b3-abb1-d0cf1f710189","ai-predictions-2026-big-changes-en","AI Predictions for 2026: Brace for Big Changes","2026-03-26T01:25:07.788356+00:00",{"id":80,"slug":81,"title":82,"created_at":83},"8404bd7b-4c2f-4109-9ec4-baf29d88af2b","ml-papers-of-the-week-github-research-desk-en","ML Papers of the Week Turns GitHub Into a Research Desk","2026-03-27T01:11:39.480259+00:00",{"id":85,"slug":86,"title":87,"created_at":88},"87897a94-8065-4464-a016-1f23e89e17cc","ai-ml-conferences-to-watch-in-2026-en","AI\u002FML Conferences to Watch in 2026","2026-03-27T01:51:54.184108+00:00",{"id":90,"slug":91,"title":92,"created_at":93},"6f1987cf-25f3-47a4-b3e6-db0997695be8","openclaw-agents-manipulated-self-sabotage-en","OpenClaw Agents Can Be Manipulated Into Failure","2026-03-28T03:03:18.899465+00:00",{"id":95,"slug":96,"title":97,"created_at":98},"a53571ad-735a-4178-9f93-cb09b699d99c","vega-driving-language-instructions-en","Vega: Driving with Natural Language Instructions","2026-03-28T14:54:04.698882+00:00",{"id":100,"slug":101,"title":102,"created_at":103},"a34581d6-f36e-46da-88bb-582fb3e7425c","personalizing-autonomous-driving-styles-en","Drive My Way: Personalizing Autonomous Driving Styles","2026-03-28T14:54:26.148181+00:00",{"id":105,"slug":106,"title":107,"created_at":108},"2bc1ad7f-26ce-4f02-9885-803b35fd229d","training-knowledge-bases-writeback-rag-en","Training Knowledge Bases with WriteBack-RAG","2026-03-28T14:54:45.643433+00:00",{"id":110,"slug":111,"title":112,"created_at":113},"71adc507-3c54-4605-bbe2-c966acd6187e","packforcing-long-video-generation-en","PackForcing: Efficient Long-Video Generation Method","2026-03-28T14:55:02.646943+00:00",{"id":115,"slug":116,"title":117,"created_at":118},"675942ef-b9ec-4c5f-a997-381250b6eacb","pixelsmile-facial-expression-editing-en","PixelSmile Framework Enhances Facial Expression Editing","2026-03-28T14:55:20.633463+00:00",{"id":120,"slug":121,"title":122,"created_at":123},"6954fa2b-8b66-4839-884b-e46f89fa1bc3","adaptive-block-scaled-data-types-en","IF4: Smarter 4-Bit Quantization That Adapts to Your Data","2026-03-31T06:00:36.65963+00:00"]