{"metadata":{"name":"Neural Compilation and Fuzzy-Function Programming","version":"1.3.0","updated_at":"2026-11-14","search_refreshed_at":"2026-10-19","core_record_count":14,"canonical_url":"https://programasweights.com/reading-list/","navigation_url":"https://programasweights.com/reading-list/data/navigation.json","citation_policy":"papers.bib and papers.csl.json use conference citations for PAW (NeurIPS 2026) and Compile by Training (EMNLP 2026 System Demonstrations), and arXiv citations for the other papers. Separate preprint citations remain available from the PAW and Compile by Training paper pages. publications.bib and publications.csl.json contain conference and journal citations."},"papers":[{"id":"1511.06279","title":"Neural Programmer-Interpreters","authors":"Scott Reed, Nando de Freitas","first_public":"2015-11-19","first_public_source":"https://arxiv.org/abs/1511.06279","arxiv_first_submission":"2015-11-19","date_label":"arXiv v1","arxiv":"1511.06279","paper_url":"https://arxiv.org/abs/1511.06279","venue":"ICLR 2016","lens":["neural-programs"],"related_terms":["neural programmer-interpreter","neural programs","program memory","compositional execution"],"summary":"Learns neural programs that share a recurrent interpreter and a program memory.","contribution":"Learns from execution traces and composes subprograms for addition, sorting, and 3D-object canonicalization.","limitations":"Training requires execution traces and task-specific interfaces.","program_form":"Program embeddings stored in a key-value memory.","conditioning":"A program identifier, arguments, and observations of the environment.","execution":"A recurrent interpreter calls subprograms and takes actions in the environment.","relationship":"PAW cites this work on learned program representations and shared interpreters.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","author_names":["Scott Reed","Nando de Freitas"],"record_url":"https://programasweights.com/reading-list/papers/1511.06279/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-1511-06279","citation_date":"2015-11-19","json_url":"https://programasweights.com/reading-list/papers/1511.06279/record.json","markdown_url":"https://programasweights.com/reading-list/papers/1511.06279/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/1511.06279/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/1511.06279/citation.csl.json","publication":{"title":"Neural Programmer-Interpreters","author_names":["Scott Reed","Nando de Freitas"],"year":2016,"type":"paper-conference","container_title":"International Conference on Learning Representations","url":"https://arxiv.org/abs/1511.06279","source_url":"https://arxiv.org/pdf/1511.06279","verified_at":"2026-10-19","source_note":"The paper PDF identifies ICLR 2016; the conference program confirms the presentation."},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/1511.06279/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/1511.06279/publication.csl.json","summary_zh":"学习由共享循环解释器执行的神经程序，并把程序表示存入程序记忆。","translations":{"zh-CN":{"summary":"学习由共享循环解释器执行的神经程序，并把程序表示存入程序记忆。","relationship":"PAW 引用这项关于学习程序表示和共享解释器的研究，并将编译器与解释器接口用于自然语言定义的模糊函数。","program_form":"存在键值记忆中的程序嵌入。","conditioning":"程序标识、参数和环境观测。","execution":"循环解释器调用子程序，并在环境中执行动作。"}}},{"id":"1609.09106","title":"HyperNetworks","authors":"David Ha, Andrew Dai, Quoc V. Le","first_public":"2016-09-27","first_public_source":"https://arxiv.org/abs/1609.09106","arxiv_first_submission":"2016-09-27","date_label":"arXiv v1","arxiv":"1609.09106","paper_url":"https://arxiv.org/abs/1609.09106","venue":"ICLR 2017","lens":["generated-adapters"],"related_terms":["hypernetwork","weight generation","fast weights","conditional parameters"],"summary":"Uses one neural network to generate the weights of another.","contribution":"Applies weight generation to convolutional and recurrent networks, relaxing weight sharing across layers and generating dynamic weights for sequence models.","limitations":"The hypernetwork is trained end to end for the target architecture.","program_form":"Weights for the target network.","conditioning":"Layer embeddings or the current recurrent state.","execution":"The target network runs with the generated weights.","relationship":"Provides the weight-generation mechanism used by later description-to-adapter methods and PAW.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","author_names":["David Ha","Andrew Dai","Quoc V. Le"],"record_url":"https://programasweights.com/reading-list/papers/1609.09106/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-1609-09106","citation_date":"2016-09-27","json_url":"https://programasweights.com/reading-list/papers/1609.09106/record.json","markdown_url":"https://programasweights.com/reading-list/papers/1609.09106/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/1609.09106/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/1609.09106/citation.csl.json","publication":{"title":"HyperNetworks","author_names":["David Ha","Andrew M. Dai","Quoc V. Le"],"year":2017,"type":"paper-conference","container_title":"International Conference on Learning Representations","url":"https://openreview.net/forum?id=rkpACe1lx","source_url":"https://openreview.net/pdf?id=rkpACe1lx","verified_at":"2026-10-19"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/1609.09106/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/1609.09106/publication.csl.json","summary_zh":"用一个神经网络生成另一个神经网络的权重。","translations":{"zh-CN":{"summary":"用一个神经网络生成另一个神经网络的权重。","relationship":"超网络提供权重生成机制。PAW 用这一机制从函数描述生成供小模型执行的适配器。","program_form":"目标网络的权重。","conditioning":"层嵌入或当前循环状态。","execution":"目标网络使用生成的权重完成计算。"}}},{"id":"2101.00190","title":"Prefix-Tuning: Optimizing Continuous Prompts for Generation","authors":"Xiang Lisa Li, Percy Liang","first_public":"2021-01-01","first_public_source":"https://arxiv.org/abs/2101.00190","arxiv_first_submission":"2021-01-01","date_label":"arXiv v1","arxiv":"2101.00190","paper_url":"https://arxiv.org/abs/2101.00190","venue":"ACL-IJCNLP 2021","lens":["peft"],"related_terms":["prefix-tuning","continuous prompts","virtual tokens","parameter-efficient fine-tuning"],"summary":"Adapts a frozen language model by learning a short sequence of continuous prefix vectors.","contribution":"Trains prefixes for table-to-text generation and summarization while updating far fewer parameters than full fine-tuning.","limitations":"A separate prefix is optimized for each task.","program_form":"Continuous prefix activations.","conditioning":"Training examples for the target task.","execution":"The frozen language model attends to the learned prefix.","relationship":"PAW evaluates compiler-generated prefixes alongside LoRA adapters.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","author_names":["Xiang Lisa Li","Percy Liang"],"record_url":"https://programasweights.com/reading-list/papers/2101.00190/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2101-00190","citation_date":"2021-01-01","json_url":"https://programasweights.com/reading-list/papers/2101.00190/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2101.00190/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2101.00190/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2101.00190/citation.csl.json","publication":{"title":"Prefix-Tuning: Optimizing Continuous Prompts for Generation","author_names":["Xiang Lisa Li","Percy Liang"],"year":2021,"type":"paper-conference","container_title":"Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers)","url":"https://aclanthology.org/2021.acl-long.353/","source_url":"https://aclanthology.org/2021.acl-long.353/","verified_at":"2026-10-19","doi":"10.18653/v1/2021.acl-long.353","publisher":"Association for Computational Linguistics","pages":"4582-4597"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2101.00190/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2101.00190/publication.csl.json","method_comparisons":[{"title":"Program form and compiler","text":"LoRA represents adaptation as low-rank weight updates; prefix tuning uses continuous activations. Their original methods learn these parameters from examples. PAW generates both forms from specifications.","source_ids":["2106.09685","2101.00190","2607.02512"],"source_urls":["https://arxiv.org/abs/2106.09685","https://aclanthology.org/2021.acl-long.353/","https://arxiv.org/html/2607.02512v1#S3"],"source_labels":["LoRA paper","Prefix-tuning paper","PAW implementations"],"topic_id":"peft","topic_label":"Parameter-efficient building blocks","topic_url":"https://programasweights.com/reading-list/guides/compilation/#peft"}],"summary_zh":"训练一段连续前缀向量，让冻结的语言模型适应目标任务。","translations":{"zh-CN":{"summary":"训练一段连续前缀向量，让冻结的语言模型适应目标任务。","relationship":"PAW 把编译器生成的前缀作为神经程序的一种表示，并与生成的 LoRA 适配器进行比较。","program_form":"连续前缀激活值。","conditioning":"目标任务的训练样本。","execution":"冻结的语言模型在生成时读取学到的前缀。"}}},{"id":"2101.00420","title":"Learning to Generate Task-Specific Adapters from Task Description","authors":"Qinyuan Ye, Xiang Ren","first_public":"2021-01-02","first_public_source":"https://arxiv.org/abs/2101.00420","arxiv_first_submission":"2021-01-02","date_label":"arXiv v1","arxiv":"2101.00420","paper_url":"https://arxiv.org/abs/2101.00420","venue":"ACL-IJCNLP 2021","lens":["generated-adapters"],"related_terms":["Hypter","task descriptions","generated adapters","zero-shot learning"],"summary":"Introduces Hypter, which generates adapters from natural-language task descriptions.","contribution":"Combines a description encoder and adapter generator with BART. Tests generalization to unseen tasks on ZEST and synthetic SQuAD.","limitations":"Training requires a collection of tasks with descriptions and labeled examples.","program_form":"Bottleneck adapters inserted into a text-to-text model.","conditioning":"A natural-language task description.","execution":"BART performs the task with the generated adapters.","relationship":"An early description-to-adapter method cited by PAW. PAW develops the programming interface for saving and calling the resulting neural functions.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","author_names":["Qinyuan Ye","Xiang Ren"],"record_url":"https://programasweights.com/reading-list/papers/2101.00420/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2101-00420","citation_date":"2021-01-02","json_url":"https://programasweights.com/reading-list/papers/2101.00420/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2101.00420/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2101.00420/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2101.00420/citation.csl.json","publication":{"title":"Learning to Generate Task-Specific Adapters from Task Description","author_names":["Qinyuan Ye","Xiang Ren"],"year":2021,"type":"paper-conference","container_title":"Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 2: Short Papers)","url":"https://aclanthology.org/2021.acl-short.82/","source_url":"https://aclanthology.org/2021.acl-short.82/","verified_at":"2026-10-19","doi":"10.18653/v1/2021.acl-short.82","publisher":"Association for Computational Linguistics","pages":"646-653"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2101.00420/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2101.00420/publication.csl.json","method_comparisons":[{"title":"From generated adapters to reusable functions","text":"Hypter, HINT, HyperTuning, and Text-to-LoRA generate adaptations from descriptions or examples. PAW uses generated parameters as part of a program that can be saved, called from ordinary code, and reused by a fixed interpreter.","source_ids":["2101.00420","2212.10315","2211.12485","2506.06105","2607.02512"],"source_urls":["https://arxiv.org/html/2607.02512v1#S10","https://proceedings.mlr.press/v267/charakorn25a.html"],"source_labels":["PAW related work","Text-to-LoRA paper"],"topic_id":"generated-adapters","topic_label":"Generated weights and adapters","topic_url":"https://programasweights.com/reading-list/guides/compilation/#generated-adapters"}],"summary_zh":"提出 Hypter，根据自然语言任务描述生成适配器，并检验对未见任务的泛化。","translations":{"zh-CN":{"summary":"提出 Hypter，根据自然语言任务描述生成适配器，并检验对未见任务的泛化。","relationship":"PAW 引用这项从描述生成适配器的研究，并进一步建立保存、加载和调用神经函数的编程接口。","program_form":"插入文本到文本模型的瓶颈适配器。","conditioning":"自然语言任务描述。","execution":"BART 加载生成的适配器后执行任务。"}}},{"id":"2106.09685","title":"LoRA: Low-Rank Adaptation of Large Language Models","authors":"Edward J. Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen","first_public":"2021-06-17","first_public_source":"https://arxiv.org/abs/2106.09685","arxiv_first_submission":"2021-06-17","date_label":"arXiv v1","arxiv":"2106.09685","paper_url":"https://arxiv.org/abs/2106.09685","venue":"ICLR 2022","lens":["peft"],"related_terms":["LoRA","low-rank adaptation","parameter-efficient fine-tuning","adapters"],"summary":"Adapts a pretrained model with low-rank weight updates while keeping the original weights fixed.","contribution":"Reduces the parameters trained and stored for each task. Evaluates adaptation across several language-model families.","limitations":"A separate adapter is trained for each task.","program_form":"Low-rank matrices that update selected model weights.","conditioning":"Training examples for the target task.","execution":"The language model applies the low-rank updates during inference.","relationship":"PAW uses LoRA as a program representation and generates the adapter with a trained compiler.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","author_names":["Edward J. Hu","Yelong Shen","Phillip Wallis","Zeyuan Allen-Zhu","Yuanzhi Li","Shean Wang","Lu Wang","Weizhu Chen"],"record_url":"https://programasweights.com/reading-list/papers/2106.09685/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2106-09685","citation_date":"2021-06-17","json_url":"https://programasweights.com/reading-list/papers/2106.09685/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2106.09685/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2106.09685/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2106.09685/citation.csl.json","publication":{"title":"LoRA: Low-Rank Adaptation of Large Language Models","author_names":["Edward Hu","Yelong Shen","Phillip Wallis","Zeyuan Allen-Zhu","Yuanzhi Li","Shean Wang","Lu Wang","Weizhu Chen"],"year":2022,"type":"paper-conference","container_title":"International Conference on Learning Representations","url":"https://openreview.net/forum?id=nZeVKeeFYf9","source_url":"https://openreview.net/pdf?id=nZeVKeeFYf9","verified_at":"2026-10-19"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2106.09685/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2106.09685/publication.csl.json","method_comparisons":[{"title":"Program form and compiler","text":"LoRA represents adaptation as low-rank weight updates; prefix tuning uses continuous activations. Their original methods learn these parameters from examples. PAW generates both forms from specifications.","source_ids":["2106.09685","2101.00190","2607.02512"],"source_urls":["https://arxiv.org/abs/2106.09685","https://aclanthology.org/2021.acl-long.353/","https://arxiv.org/html/2607.02512v1#S3"],"source_labels":["LoRA paper","Prefix-tuning paper","PAW implementations"],"topic_id":"peft","topic_label":"Parameter-efficient building blocks","topic_url":"https://programasweights.com/reading-list/guides/compilation/#peft"}],"summary_zh":"冻结预训练模型的原始权重，训练低秩权重更新来适应任务。","translations":{"zh-CN":{"summary":"冻结预训练模型的原始权重，训练低秩权重更新来适应任务。","relationship":"PAW 用 LoRA 表示程序中的参数部分，由训练好的编译器生成适配器权重。","program_form":"用于更新指定模型权重的低秩矩阵。","conditioning":"目标任务的训练样本。","execution":"语言模型在推理时使用低秩权重更新。"}}},{"id":"2211.12485","title":"HyperTuning: Toward Adapting Large Language Models without Back-propagation","authors":"Jason Phang, Yi Mao, Pengcheng He, Weizhu Chen","first_public":"2022-11-22","first_public_source":"https://arxiv.org/abs/2211.12485","arxiv_first_submission":"2022-11-22","date_label":"arXiv v1","arxiv":"2211.12485","paper_url":"https://arxiv.org/abs/2211.12485","venue":"ICML 2023","lens":["generated-adapters"],"related_terms":["HyperTuning","HyperT5","few-shot adaptation","generated LoRA"],"summary":"Produces task-specific parameters for a frozen language model from a few demonstrations.","contribution":"Introduces HyperT5 models that generate prefixes or LoRA parameters. Hyperpretraining and multitask training prepare the generator for new tasks.","limitations":"The hypermodel accepts a limited number of demonstrations. The paper also studies further task-specific tuning.","program_form":"Generated prefix or LoRA parameters.","conditioning":"A few examples of the task.","execution":"A frozen T5 model runs with the generated parameters.","relationship":"PAW cites HyperTuning as prior work on generating adaptation parameters from examples.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","author_names":["Jason Phang","Yi Mao","Pengcheng He","Weizhu Chen"],"record_url":"https://programasweights.com/reading-list/papers/2211.12485/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2211-12485","citation_date":"2022-11-22","json_url":"https://programasweights.com/reading-list/papers/2211.12485/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2211.12485/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2211.12485/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2211.12485/citation.csl.json","publication":{"title":"HyperTuning: Toward Adapting Large Language Models without Back-propagation","author_names":["Jason Phang","Yi Mao","Pengcheng He","Weizhu Chen"],"year":2023,"type":"paper-conference","container_title":"Proceedings of the 40th International Conference on Machine Learning","url":"https://proceedings.mlr.press/v202/phang23a.html","source_url":"https://proceedings.mlr.press/v202/phang23a.html","verified_at":"2026-10-19","volume":"202","publisher":"PMLR","pages":"27854-27875"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2211.12485/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2211.12485/publication.csl.json","method_comparisons":[{"title":"From generated adapters to reusable functions","text":"Hypter, HINT, HyperTuning, and Text-to-LoRA generate adaptations from descriptions or examples. PAW uses generated parameters as part of a program that can be saved, called from ordinary code, and reused by a fixed interpreter.","source_ids":["2101.00420","2212.10315","2211.12485","2506.06105","2607.02512"],"source_urls":["https://arxiv.org/html/2607.02512v1#S10","https://proceedings.mlr.press/v267/charakorn25a.html"],"source_labels":["PAW related work","Text-to-LoRA paper"],"topic_id":"generated-adapters","topic_label":"Generated weights and adapters","topic_url":"https://programasweights.com/reading-list/guides/compilation/#generated-adapters"}],"summary_zh":"根据几个任务示例，为冻结的语言模型生成前缀或 LoRA 参数。","translations":{"zh-CN":{"summary":"根据几个任务示例，为冻结的语言模型生成前缀或 LoRA 参数。","relationship":"PAW 引用 HyperTuning，讨论从示例生成模型适配参数的方法。","program_form":"生成的前缀或 LoRA 参数。","conditioning":"少量任务示例。","execution":"冻结的 T5 模型使用生成的参数执行任务。"}}},{"id":"2212.10315","title":"HINT: Hypernetwork Instruction Tuning for Efficient Zero- & Few-Shot Generalisation","authors":"Hamish Ivison, Akshita Bhagia, Yizhong Wang, Hannaneh Hajishirzi, Matthew Peters","first_public":"2022-12-20","first_public_source":"https://arxiv.org/abs/2212.10315","arxiv_first_submission":"2022-12-20","date_label":"arXiv v1","arxiv":"2212.10315","paper_url":"https://arxiv.org/abs/2212.10315","venue":"ACL 2023","lens":["generated-adapters"],"related_terms":["HINT","instruction tuning","hypernetwork","instruction fusion"],"summary":"Turns instructions and examples into reusable prefixes, adapters, and encoded instruction representations.","contribution":"Combines a hypernetwork with instruction fusion so repeated calls can reuse the instruction computation. Evaluates zero-shot and few-shot generalization.","limitations":"The generated modules and cached instruction representations are tied to the trained encoder-decoder model.","program_form":"Prefixes, adapters, and encoded instruction representations.","conditioning":"Task instructions and optional examples.","execution":"An encoder-decoder model uses the generated modules and attends to cached instruction representations.","relationship":"HINT reuses generated modules and an encoded instruction. PAW reuses a program containing text and adaptation parameters.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","author_names":["Hamish Ivison","Akshita Bhagia","Yizhong Wang","Hannaneh Hajishirzi","Matthew Peters"],"record_url":"https://programasweights.com/reading-list/papers/2212.10315/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2212-10315","citation_date":"2022-12-20","json_url":"https://programasweights.com/reading-list/papers/2212.10315/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2212.10315/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2212.10315/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2212.10315/citation.csl.json","publication":{"title":"HINT: Hypernetwork Instruction Tuning for Efficient Zero- and Few-Shot Generalisation","author_names":["Hamish Ivison","Akshita Bhagia","Yizhong Wang","Hannaneh Hajishirzi","Matthew E. Peters"],"year":2023,"type":"paper-conference","container_title":"Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers)","url":"https://aclanthology.org/2023.acl-long.631/","source_url":"https://aclanthology.org/2023.acl-long.631/","verified_at":"2026-10-19","doi":"10.18653/v1/2023.acl-long.631","publisher":"Association for Computational Linguistics","pages":"11272-11288"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2212.10315/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2212.10315/publication.csl.json","method_comparisons":[{"title":"From generated adapters to reusable functions","text":"Hypter, HINT, HyperTuning, and Text-to-LoRA generate adaptations from descriptions or examples. PAW uses generated parameters as part of a program that can be saved, called from ordinary code, and reused by a fixed interpreter.","source_ids":["2101.00420","2212.10315","2211.12485","2506.06105","2607.02512"],"source_urls":["https://arxiv.org/html/2607.02512v1#S10","https://proceedings.mlr.press/v267/charakorn25a.html"],"source_labels":["PAW related work","Text-to-LoRA paper"],"topic_id":"generated-adapters","topic_label":"Generated weights and adapters","topic_url":"https://programasweights.com/reading-list/guides/compilation/#generated-adapters"}],"summary_zh":"把指令和示例转化为前缀、适配器及编码后的指令表示，供多次调用复用。","translations":{"zh-CN":{"summary":"把指令和示例转化为前缀、适配器及编码后的指令表示，供多次调用复用。","relationship":"HINT 复用生成的模块和编码后的指令；PAW 保存并复用包含文本与适配参数的程序。","program_form":"前缀、适配器和编码后的指令表示。","conditioning":"任务指令和提供的示例。","execution":"编码器—解码器模型使用生成的模块，并读取缓存的指令表示。"}}},{"id":"2304.08467","title":"Learning to Compress Prompts with Gist Tokens","authors":"Jesse Mu, Xiang Lisa Li, Noah Goodman","first_public":"2023-04-17","first_public_source":"https://arxiv.org/abs/2304.08467","arxiv_first_submission":"2023-04-17","date_label":"arXiv v1","arxiv":"2304.08467","paper_url":"https://arxiv.org/abs/2304.08467","venue":"NeurIPS 2023","lens":["context-compression"],"related_terms":["gist tokens","prompt compression","activation caching","amortized inference"],"summary":"Compresses an instruction into a few reusable token activations.","contribution":"Trains with an attention mask that forces later tokens to use gist activations in place of the original prompt.","limitations":"Compression can reduce output quality. The cached activations are tied to the trained model.","program_form":"Cached activations at a few gist-token positions.","conditioning":"The instruction or prompt.","execution":"The language model generates outputs while attending to the cached gist.","relationship":"PAW cites gisting as another way to reuse instruction computation. Gisting stores activations; PAW stores a program with text and adaptation parameters.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","author_names":["Jesse Mu","Xiang Lisa Li","Noah Goodman"],"record_url":"https://programasweights.com/reading-list/papers/2304.08467/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2304-08467","citation_date":"2023-04-17","json_url":"https://programasweights.com/reading-list/papers/2304.08467/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2304.08467/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2304.08467/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2304.08467/citation.csl.json","publication":{"title":"Learning to Compress Prompts with Gist Tokens","author_names":["Jesse Mu","Xiang Li","Noah Goodman"],"year":2023,"type":"paper-conference","container_title":"Advances in Neural Information Processing Systems","url":"https://proceedings.neurips.cc/paper_files/paper/2023/hash/3d77c6dcc7f143aa2154e7f4d5e22d68-Abstract-Conference.html","source_url":"https://proceedings.neurips.cc/paper_files/paper/2023/hash/3d77c6dcc7f143aa2154e7f4d5e22d68-Abstract-Conference.html","verified_at":"2026-10-19","volume":"36","doi":"10.52202/075280-0848","pages":"19327-19352"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2304.08467/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2304.08467/publication.csl.json","method_comparisons":[{"title":"What gets reused?","text":"Gisting caches a few instruction activations. Generative Adapter stores context in parameter updates. PAW saves a function as a program containing text and a neural module.","source_ids":["2304.08467","2411.05877","2607.02512"],"source_urls":["https://arxiv.org/abs/2304.08467","https://arxiv.org/abs/2411.05877","https://arxiv.org/html/2607.02512v1#S2"],"source_labels":["Gist paper","Generative Adapter paper","PAW program definition"],"topic_id":"context-compression","topic_label":"Context compression","topic_url":"https://programasweights.com/reading-list/guides/compilation/#context-compression"}],"summary_zh":"将一段指令压缩为几个可复用的 gist token 激活值。","translations":{"zh-CN":{"summary":"将一段指令压缩为几个可复用的 gist token 激活值。","relationship":"PAW 引用 gisting，讨论如何复用指令计算。Gisting 保存激活值，PAW 保存包含文本与适配参数的程序。","program_form":"少量 gist token 位置上的缓存激活值。","conditioning":"指令或提示词。","execution":"语言模型读取缓存的 gist 表示来生成输出。"}}},{"id":"2407.11004","title":"The ALCHEmist: Automated Labeling 500x CHEaper Than LLM Data Annotators","authors":"Tzu-Heng Huang, Catherine Cao, Vaishnavi Bhargava, Frederic Sala","first_public":"2024-06-25","first_public_source":"https://arxiv.org/abs/2407.11004","arxiv_first_submission":"2024-06-25","date_label":"arXiv v1","arxiv":"2407.11004","paper_url":"https://arxiv.org/abs/2407.11004","venue":"NeurIPS 2024","lens":["neural-programs"],"related_terms":["ALCHEmist","program synthesis","automated labeling","reusable tools"],"summary":"Asks a language model to write labeling programs that can be saved and run locally.","contribution":"Reuses generated code across many labeling decisions and combines noisy program outputs with weak supervision.","limitations":"Label quality depends on the model’s ability to write useful programs. Complex tasks can produce code that is harder to inspect.","program_form":"Symbolic labeling programs.","conditioning":"A task description, labeling instructions, and a function signature.","execution":"Generated programs label examples locally; their outputs can be combined to train a downstream model.","relationship":"PAW evaluates an ALCHEmist-style approach. Both build reusable tools: ALCHEmist produces symbolic code, while PAW produces programs for a neural interpreter.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","author_names":["Tzu-Heng Huang","Catherine Cao","Vaishnavi Bhargava","Frederic Sala"],"record_url":"https://programasweights.com/reading-list/papers/2407.11004/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2407-11004","citation_date":"2024-06-25","json_url":"https://programasweights.com/reading-list/papers/2407.11004/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2407.11004/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2407.11004/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2407.11004/citation.csl.json","publication":{"title":"The ALCHEmist: Automated Labeling 500x CHEaper than LLM Data Annotators","author_names":["Tzu-Heng Huang","Catherine Cao","Vaishnavi Bhargava","Frederic Sala"],"year":2024,"type":"paper-conference","container_title":"Advances in Neural Information Processing Systems","url":"https://proceedings.neurips.cc/paper_files/paper/2024/hash/72802bef5cf1a3449e909b20c2ae18d5-Abstract-Conference.html","source_url":"https://proceedings.neurips.cc/paper_files/paper/2024/hash/72802bef5cf1a3449e909b20c2ae18d5-Abstract-Conference.html","verified_at":"2026-10-19","volume":"37","doi":"10.52202/079017-2003","pages":"62648-62672"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2407.11004/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2407.11004/publication.csl.json","method_comparisons":[{"title":"Code and neural programs","text":"ALCHEmist generates symbolic labeling programs. PAW compiles fuzzy functions into text and learned parameters. Both build an artifact once and reuse it; one runs as code, the other on a neural interpreter.","source_ids":["2407.11004","2607.02512"],"source_urls":["https://proceedings.neurips.cc/paper_files/paper/2024/hash/72802bef5cf1a3449e909b20c2ae18d5-Abstract-Conference.html","https://arxiv.org/html/2607.02512v1#S6"],"source_labels":["ALCHEmist paper","PAW evaluation"],"topic_id":"neural-programs","topic_label":"Programs and reusable artifacts","topic_url":"https://programasweights.com/reading-list/guides/fuzzy-functions/"}],"summary_zh":"让语言模型编写可保存、可本地执行的标注程序，并用弱监督组合多个程序的输出。","translations":{"zh-CN":{"summary":"让语言模型编写可保存、可本地执行的标注程序，并用弱监督组合多个程序的输出。","relationship":"PAW 评测了 ALCHEmist 风格的方法。ALCHEmist 生成符号代码，PAW 生成由神经解释器执行的程序，两者都构建可复用的工具。","program_form":"符号标注程序。","conditioning":"任务描述、标注要求和函数签名。","execution":"生成的程序在本地标注数据，其输出可进一步用于训练下游模型。"}}},{"id":"2411.05877","title":"Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass","authors":"Tong Chen, Hao Fang, Patrick Xia, Xiaodong Liu, Benjamin Van Durme, Luke Zettlemoyer, Jianfeng Gao, Hao Cheng","first_public":"2024-11-08","first_public_source":"https://arxiv.org/abs/2411.05877","arxiv_first_submission":"2024-11-08","date_label":"arXiv v1","arxiv":"2411.05877","paper_url":"https://arxiv.org/abs/2411.05877","venue":"ICLR 2025","lens":["generated-adapters","context-compression"],"related_terms":["Generative Adapter","contextualization","fast weights","streaming context"],"summary":"Converts new context into low-rank model updates in one forward pass.","contribution":"Trains an adapter generator with self-supervision. Tests learning from documents, demonstrations, and conversation history.","limitations":"The adapter generator is tied to its frozen base model.","program_form":"Additive low-rank adapters.","conditioning":"Documents, demonstrations, or conversation history, including successive chunks of context.","execution":"The adapted model answers new inputs using context stored in its parameter updates.","relationship":"PAW cites this context-to-adapter method. Generative Adapter stores information from context; PAW compiles the behavior of a function.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","first_public_note":"An AFM 2024 OpenReview version appeared before the first arXiv submission.","author_names":["Tong Chen","Hao Fang","Patrick Xia","Xiaodong Liu","Benjamin Van Durme","Luke Zettlemoyer","Jianfeng Gao","Hao Cheng"],"record_url":"https://programasweights.com/reading-list/papers/2411.05877/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2411-05877","citation_date":"2024-11-08","json_url":"https://programasweights.com/reading-list/papers/2411.05877/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2411.05877/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2411.05877/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2411.05877/citation.csl.json","publication":{"title":"Generative Adapter: Contextualizing Language Models in Parameters with A Single Forward Pass","author_names":["Tong Chen","Hao Fang","Patrick Xia","Xiaodong Liu","Benjamin Van Durme","Luke Zettlemoyer","Jianfeng Gao","Hao Cheng"],"year":2025,"type":"paper-conference","container_title":"International Conference on Learning Representations","url":"https://openreview.net/forum?id=bc3sUsS6ck","source_url":"https://openreview.net/pdf?id=bc3sUsS6ck","verified_at":"2026-10-19"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2411.05877/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2411.05877/publication.csl.json","method_comparisons":[{"title":"What gets reused?","text":"Gisting caches a few instruction activations. Generative Adapter stores context in parameter updates. PAW saves a function as a program containing text and a neural module.","source_ids":["2304.08467","2411.05877","2607.02512"],"source_urls":["https://arxiv.org/abs/2304.08467","https://arxiv.org/abs/2411.05877","https://arxiv.org/html/2607.02512v1#S2"],"source_labels":["Gist paper","Generative Adapter paper","PAW program definition"],"topic_id":"context-compression","topic_label":"Context compression","topic_url":"https://programasweights.com/reading-list/guides/compilation/#context-compression"}],"summary_zh":"通过一次前向传播，将新上下文转化为低秩参数更新。","translations":{"zh-CN":{"summary":"通过一次前向传播，将新上下文转化为低秩参数更新。","relationship":"Generative Adapter 将上下文信息写入适配器；PAW 把函数的行为编译成程序。","program_form":"可叠加的低秩适配器。","conditioning":"文档、示例或对话历史，也支持连续到来的上下文片段。","execution":"模型通过参数更新中保存的上下文信息回答新输入。"}}},{"id":"2506.06105","title":"Text-to-LoRA: Instant Transformer Adaption","authors":"Rujikorn Charakorn, Edoardo Cetin, Yujin Tang, Robert Tjarko Lange","first_public":"2025-06-06","first_public_source":"https://arxiv.org/abs/2506.06105","arxiv_first_submission":"2025-06-06","date_label":"arXiv v1","arxiv":"2506.06105","paper_url":"https://arxiv.org/abs/2506.06105","venue":"ICML 2025","lens":["generated-adapters"],"related_terms":["Text-to-LoRA","T2L","task description","adapter generation"],"summary":"Uses a hypernetwork to generate LoRA adapters from task descriptions.","contribution":"Trains hypernetworks to reconstruct existing adapters or learn from supervised task data. Tests adapter compression and generation for unseen tasks.","limitations":"Reconstruction-trained T2L underperforms the supervised variant on unseen tasks.","program_form":"Generated LoRA matrices.","conditioning":"A natural-language task description.","execution":"A base transformer performs the task with the generated LoRA.","relationship":"PAW cites Text-to-LoRA as prior work on generating adapters. PAW develops fuzzy-function programming, FuzzyBench, and an interface for reusable neural programs.","relationship_source":"https://arxiv.org/html/2607.02512v1#S10","provenance":"arXiv record and paper.","author_names":["Rujikorn Charakorn","Edoardo Cetin","Yujin Tang","Robert Tjarko Lange"],"record_url":"https://programasweights.com/reading-list/papers/2506.06105/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2506-06105","citation_date":"2025-06-06","json_url":"https://programasweights.com/reading-list/papers/2506.06105/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2506.06105/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2506.06105/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2506.06105/citation.csl.json","publication":{"title":"Text-to-LoRA: Instant Transformer Adaption","author_names":["Rujikorn Charakorn","Edoardo Cetin","Yujin Tang","Robert Tjarko Lange"],"year":2025,"type":"paper-conference","container_title":"Proceedings of the 42nd International Conference on Machine Learning","url":"https://proceedings.mlr.press/v267/charakorn25a.html","source_url":"https://proceedings.mlr.press/v267/charakorn25a.html","verified_at":"2026-10-19","volume":"267","publisher":"PMLR","pages":"7485-7514"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2506.06105/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2506.06105/publication.csl.json","method_comparisons":[{"title":"From generated adapters to reusable functions","text":"Hypter, HINT, HyperTuning, and Text-to-LoRA generate adaptations from descriptions or examples. PAW uses generated parameters as part of a program that can be saved, called from ordinary code, and reused by a fixed interpreter.","source_ids":["2101.00420","2212.10315","2211.12485","2506.06105","2607.02512"],"source_urls":["https://arxiv.org/html/2607.02512v1#S10","https://proceedings.mlr.press/v267/charakorn25a.html"],"source_labels":["PAW related work","Text-to-LoRA paper"],"topic_id":"generated-adapters","topic_label":"Generated weights and adapters","topic_url":"https://programasweights.com/reading-list/guides/compilation/#generated-adapters"}],"summary_zh":"用超网络根据任务描述生成 LoRA 适配器，研究适配器压缩和未见任务的适配。","translations":{"zh-CN":{"summary":"用超网络根据任务描述生成 LoRA 适配器，研究适配器压缩和未见任务的适配。","relationship":"PAW 引用 Text-to-LoRA 的适配器生成方法，并发展模糊函数编程、FuzzyBench 和可复用神经程序的接口。","program_form":"生成的 LoRA 矩阵。","conditioning":"自然语言任务描述。","execution":"基础 Transformer 使用生成的 LoRA 执行任务。"}}},{"id":"2606.06492","title":"Code2LoRA: Hypernetwork-Generated Adapters for Code Language Models under Software Evolution","authors":"Liliana Hotsko, Yinxi Li, Yuntian Deng, Pengyu Nie","first_public":"2026-06-04","first_public_source":"https://arxiv.org/abs/2606.06492","arxiv_first_submission":"2026-06-04","date_label":"arXiv v1","arxiv":"2606.06492","paper_url":"https://arxiv.org/abs/2606.06492","full_text":"https://arxiv.org/html/2606.06492v2","method_source":"https://arxiv.org/html/2606.06492v2#S3","project_url":"https://lilianahotsko.com/Code2LoRA/","venue":"arXiv","lens":["generated-adapters"],"related_terms":["Code2LoRA","repository-conditioned adapters","commit-streaming adapters","hypernetworks","weight-space programming"],"summary":"Trains a hypernetwork to generate repository-specific LoRA adapters from source code and update them as commits arrive.","contribution":"Code2LoRA-Static maps a repository snapshot to an adapter. Code2LoRA-Evo updates a GRU state with each code diff and generates the adapter from that state.","limitations":"The paper evaluates Python test-assertion completion with Qwen2.5-Coder-1.5B.","program_form":"Generated LoRA matrices for a frozen code language model.","conditioning":"A repository snapshot; the evolution variant also receives successive commit diffs.","execution":"The code model uses the generated adapter without adding repository context tokens to each request.","relationship":"Code2LoRA and PAW train models to generate reusable weights. Code2LoRA puts repository knowledge into adapters; PAW compiles function specifications into neural programs with a compiler–interpreter interface.","provenance":"June 4, 2026 arXiv release; method description from the October 27, 2026 revision.","author_names":["Liliana Hotsko","Yinxi Li","Yuntian Deng","Pengyu Nie"],"record_url":"https://programasweights.com/reading-list/papers/2606.06492/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2606-06492","citation_date":"2026-06-04","json_url":"https://programasweights.com/reading-list/papers/2606.06492/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2606.06492/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2606.06492/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2606.06492/citation.csl.json","summary_zh":"训练超网络，根据代码仓库生成 LoRA 适配器，并随新的提交更新适配器。","translations":{"zh-CN":{"summary":"训练超网络，根据代码仓库生成 LoRA 适配器，并随新的提交更新适配器。","relationship":"Code2LoRA 和 PAW 都训练模型来生成可复用的权重。Code2LoRA 将仓库知识写入适配器；PAW 将函数说明编译成神经程序，并建立编译器与解释器接口。","program_form":"供冻结的代码语言模型使用的 LoRA 矩阵。","conditioning":"代码仓库快照；演化版本还接收后续提交的代码差异。","execution":"代码模型使用生成的适配器，无需在每次请求中加入仓库上下文。"}}},{"id":"2607.02512","title":"Program-as-Weights: A Programming Paradigm for Fuzzy Functions","authors":"Wentao Zhang, Liliana Hotsko, Woojeong Kim, Pengyu Nie, Stuart Shieber, Yuntian Deng","first_public":"2026-07-02","first_public_source":"https://arxiv.org/abs/2607.02512","arxiv_first_submission":"2026-07-02","date_label":"arXiv v1","arxiv":"2607.02512","paper_url":"https://arxiv.org/abs/2607.02512","venue":"NeurIPS 2026","lens":["neural-programs","generated-adapters"],"related_terms":["Program-as-Weights","ProgramAsWeights","PAW","fuzzy-function programming","neural compiler","neural interpreter","FuzzyBench"],"summary":"Proposes fuzzy-function programming: compile a natural-language function specification into a neural program, then reuse it locally.","contribution":"Develops the compiler–interpreter abstraction and programs that combine text with learned parameters. Introduces the 10M-example FuzzyBench dataset, local execution interfaces, and image-conditioned compilation. Prefix tuning and LoRA are two implementations.","program_form":"A discrete pseudo-program and a compiler-generated parameter-efficient module.","conditioning":"A function specification, optionally with examples. The image extension also accepts visual information.","execution":"A frozen lightweight interpreter loads the program and runs new inputs locally.","relationship":"Compile by Training builds on this programming abstraction. It keeps the same interface and refines the fast compiler’s prediction using synthesized examples.","relationship_source":"https://arxiv.org/html/2609.04199v1#S1","project_url":"https://programasweights.com/","code_url":"https://github.com/programasweights","equal_contribution":["Wentao Zhang","Liliana Hotsko","Woojeong Kim"],"venue_source":"https://arxiv.org/abs/2607.02512","provenance":"July 2026 arXiv paper; NeurIPS 2026.","author_names":["Wentao Zhang","Liliana Hotsko","Woojeong Kim","Pengyu Nie","Stuart Shieber","Yuntian Deng"],"record_url":"https://programasweights.com/reading-list/papers/2607.02512/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2607-02512","citation_date":"2026","json_url":"https://programasweights.com/reading-list/papers/2607.02512/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2607.02512/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2607.02512/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2607.02512/citation.csl.json","publication":{"title":"Program-as-Weights: A Programming Paradigm for Fuzzy Functions","author_names":["Wentao Zhang","Liliana Hotsko","Woojeong Kim","Pengyu Nie","Stuart Shieber","Yuntian Deng"],"year":2026,"type":"paper-conference","container_title":"The Fortieth Annual Conference on Neural Information Processing Systems","url":"https://arxiv.org/abs/2607.02512","source_url":"https://arxiv.org/abs/2607.02512","verified_at":"2026-10-24","preferred_citation":true,"citation_key":"zhang2026programasweights"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2607.02512/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2607.02512/publication.csl.json","preferred_citation":{"title":"Program-as-Weights: A Programming Paradigm for Fuzzy Functions","author_names":["Wentao Zhang","Liliana Hotsko","Woojeong Kim","Pengyu Nie","Stuart Shieber","Yuntian Deng"],"year":2026,"type":"paper-conference","container_title":"The Fortieth Annual Conference on Neural Information Processing Systems","url":"https://arxiv.org/abs/2607.02512","source_url":"https://arxiv.org/abs/2607.02512","verified_at":"2026-10-24","preferred_citation":true,"citation_key":"zhang2026programasweights"},"citation_url":"https://arxiv.org/abs/2607.02512","preprint_citation_date":"2026-07-02","preprint_bibtex_url":"https://programasweights.com/reading-list/papers/2607.02512/preprint.bib","preprint_csl_url":"https://programasweights.com/reading-list/papers/2607.02512/preprint.csl.json","method_comparisons":[{"title":"From generated adapters to reusable functions","text":"Hypter, HINT, HyperTuning, and Text-to-LoRA generate adaptations from descriptions or examples. PAW uses generated parameters as part of a program that can be saved, called from ordinary code, and reused by a fixed interpreter.","source_ids":["2101.00420","2212.10315","2211.12485","2506.06105","2607.02512"],"source_urls":["https://arxiv.org/html/2607.02512v1#S10","https://proceedings.mlr.press/v267/charakorn25a.html"],"source_labels":["PAW related work","Text-to-LoRA paper"],"topic_id":"generated-adapters","topic_label":"Generated weights and adapters","topic_url":"https://programasweights.com/reading-list/guides/compilation/#generated-adapters"},{"title":"Code and neural programs","text":"ALCHEmist generates symbolic labeling programs. PAW compiles fuzzy functions into text and learned parameters. Both build an artifact once and reuse it; one runs as code, the other on a neural interpreter.","source_ids":["2407.11004","2607.02512"],"source_urls":["https://proceedings.neurips.cc/paper_files/paper/2024/hash/72802bef5cf1a3449e909b20c2ae18d5-Abstract-Conference.html","https://arxiv.org/html/2607.02512v1#S6"],"source_labels":["ALCHEmist paper","PAW evaluation"],"topic_id":"neural-programs","topic_label":"Programs and reusable artifacts","topic_url":"https://programasweights.com/reading-list/guides/fuzzy-functions/"},{"title":"What gets reused?","text":"Gisting caches a few instruction activations. Generative Adapter stores context in parameter updates. PAW saves a function as a program containing text and a neural module.","source_ids":["2304.08467","2411.05877","2607.02512"],"source_urls":["https://arxiv.org/abs/2304.08467","https://arxiv.org/abs/2411.05877","https://arxiv.org/html/2607.02512v1#S2"],"source_labels":["Gist paper","Generative Adapter paper","PAW program definition"],"topic_id":"context-compression","topic_label":"Context compression","topic_url":"https://programasweights.com/reading-list/guides/compilation/#context-compression"},{"title":"Program form and compiler","text":"LoRA represents adaptation as low-rank weight updates; prefix tuning uses continuous activations. Their original methods learn these parameters from examples. PAW generates both forms from specifications.","source_ids":["2106.09685","2101.00190","2607.02512"],"source_urls":["https://arxiv.org/abs/2106.09685","https://aclanthology.org/2021.acl-long.353/","https://arxiv.org/html/2607.02512v1#S3"],"source_labels":["LoRA paper","Prefix-tuning paper","PAW implementations"],"topic_id":"peft","topic_label":"Parameter-efficient building blocks","topic_url":"https://programasweights.com/reading-list/guides/compilation/#peft"},{"title":"Fast prediction and training","text":"PAW’s fast compiler predicts a program. Compile by Training starts from that prediction, generates examples, and trains an adapter. Both use the same runtime interface; the extra training improves semantic accuracy on FuzzyBench-Hard at a higher compilation cost.","source_ids":["2607.02512","2609.04199"],"source_urls":["https://arxiv.org/html/2607.02512v1#S3","https://arxiv.org/html/2609.04199v1#S3","https://arxiv.org/html/2609.04199v1#S5"],"source_labels":["PAW compiler implementations","Compile by Training method","Compile by Training evaluation"],"topic_id":"training-compilation","topic_label":"Compilation through training","topic_url":"https://programasweights.com/reading-list/guides/compilation/#training-compilation"}],"summary_zh":"提出模糊函数编程：将自然语言函数说明编译成神经程序，保存后由小模型在本地反复执行。","translations":{"zh-CN":{"summary":"提出模糊函数编程：将自然语言函数说明编译成神经程序，保存后由小模型在本地反复执行。","relationship":"PAW 建立编译器与解释器接口，程序同时包含文本和生成的参数，并提供 FuzzyBench、本地运行接口及图像条件编译。Compile by Training 沿用这一程序格式和运行时。","program_form":"文本伪程序与编译器生成的参数高效模块，如 LoRA 适配器或前缀向量。","conditioning":"函数说明，也支持输入输出示例；图像扩展还接收视觉信息。","execution":"冻结的轻量解释器加载保存的程序，在本地处理新输入。"}}},{"id":"2609.04199","title":"Compile by Training: Turning Natural-Language Specifications into Local Neural Functions","authors":"Yuntian Deng, Pengyu Nie, Stuart Shieber","first_public":"2026-09-03","first_public_source":"https://arxiv.org/abs/2609.04199","arxiv_first_submission":"2026-09-03","date_label":"arXiv v1","arxiv":"2609.04199","paper_url":"https://arxiv.org/abs/2609.04199","original_url":"https://arxiv.org/html/2609.04199v1","venue":"EMNLP 2026 System Demonstrations","venue_source":"https://aclanthology.org/2026.emnlp-demos.78/","lens":["neural-programs","training-compilation"],"related_terms":["compile by training","training-based compilation","specification-specific optimization","synthetic supervision","local neural function","FuzzyBench-Hard","ProgramAsWeights"],"summary":"Compiles a function description into a reusable local program by generating training examples and optimizing an adapter.","contribution":"Makes task-specific training a software build step. Raises FuzzyBench-Hard semantic accuracy (LEM) from 22.4% to 83.6% and demonstrates composition in a website helper, avatar control, and English–Claudish translation.","program_form":"A LoRA adapter, runtime prompt, specification, and interpreter metadata in a PAW program.","conditioning":"The specification guides teacher synthesis. PAW’s fast compiler supplies the initial adapter and runtime prompt.","execution":"A frozen Qwen3-0.6B interpreter runs the saved function locally without further teacher calls.","relationship":"Extends PAW with training at compile time while keeping its program format and runtime interface.","relationship_source":"https://arxiv.org/html/2609.04199v1#S1","project_url":"https://programasweights.com/playground?compiler=paw-ft-bs48","demo_url":"https://programasweights.com/playground?compiler=paw-ft-bs48","code_url":"https://github.com/programasweights/compile-by-training","provenance":"September 2026 arXiv paper; published in EMNLP 2026 System Demonstrations, pages 830–839.","first_public_note":"The paper reports public-service measurements from May 2026.","author_names":["Yuntian Deng","Pengyu Nie","Stuart Shieber"],"record_url":"https://programasweights.com/reading-list/papers/2609.04199/","legacy_record_url":"https://programasweights.com/reading-list/papers/#paper-2609-04199","citation_date":"2026","json_url":"https://programasweights.com/reading-list/papers/2609.04199/record.json","markdown_url":"https://programasweights.com/reading-list/papers/2609.04199/index.md","bibtex_url":"https://programasweights.com/reading-list/papers/2609.04199/citation.bib","csl_url":"https://programasweights.com/reading-list/papers/2609.04199/citation.csl.json","publication":{"title":"Compile by Training: Turning Natural-Language Specifications into Local Neural Functions","author_names":["Yuntian Deng","Pengyu Nie","Stuart Shieber"],"year":2026,"type":"paper-conference","container_title":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations","editor_names":["Sweta Agrawal","Alberto Abad"],"month":10,"month_name":"October","address":"Budapest, Hungary","publisher":"Association for Computational Linguistics","pages":"830-839","bibtex_pages":"830--839","doi":"10.18653/2026.emnlp-demos.78","isbn":"979-8-89176-461-3","url":"https://aclanthology.org/2026.emnlp-demos.78/","source_url":"https://aclanthology.org/2026.emnlp-demos.78/","verified_at":"2026-11-06","preferred_citation":true,"citation_key":"deng-etal-2026-compile"},"publication_bibtex_url":"https://programasweights.com/reading-list/papers/2609.04199/publication.bib","publication_csl_url":"https://programasweights.com/reading-list/papers/2609.04199/publication.csl.json","preferred_citation":{"title":"Compile by Training: Turning Natural-Language Specifications into Local Neural Functions","author_names":["Yuntian Deng","Pengyu Nie","Stuart Shieber"],"year":2026,"type":"paper-conference","container_title":"Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing: System Demonstrations","editor_names":["Sweta Agrawal","Alberto Abad"],"month":10,"month_name":"October","address":"Budapest, Hungary","publisher":"Association for Computational Linguistics","pages":"830-839","bibtex_pages":"830--839","doi":"10.18653/2026.emnlp-demos.78","isbn":"979-8-89176-461-3","url":"https://aclanthology.org/2026.emnlp-demos.78/","source_url":"https://aclanthology.org/2026.emnlp-demos.78/","verified_at":"2026-11-06","preferred_citation":true,"citation_key":"deng-etal-2026-compile"},"citation_url":"https://aclanthology.org/2026.emnlp-demos.78/","preprint_citation_date":"2026-09-03","preprint_bibtex_url":"https://programasweights.com/reading-list/papers/2609.04199/preprint.bib","preprint_csl_url":"https://programasweights.com/reading-list/papers/2609.04199/preprint.csl.json","method_comparisons":[{"title":"Fast prediction and training","text":"PAW’s fast compiler predicts a program. Compile by Training starts from that prediction, generates examples, and trains an adapter. Both use the same runtime interface; the extra training improves semantic accuracy on FuzzyBench-Hard at a higher compilation cost.","source_ids":["2607.02512","2609.04199"],"source_urls":["https://arxiv.org/html/2607.02512v1#S3","https://arxiv.org/html/2609.04199v1#S3","https://arxiv.org/html/2609.04199v1#S5"],"source_labels":["PAW compiler implementations","Compile by Training method","Compile by Training evaluation"],"topic_id":"training-compilation","topic_label":"Compilation through training","topic_url":"https://programasweights.com/reading-list/guides/compilation/#training-compilation"}],"summary_zh":"根据函数说明生成训练样本，再训练 LoRA 适配器，将一次函数定义构建成可复用的本地神经程序。","translations":{"zh-CN":{"summary":"根据函数说明生成训练样本，再训练 LoRA 适配器，将一次函数定义构建成可复用的本地神经程序。","relationship":"在 PAW 的快速编译结果上增加针对当前函数的训练，保留 PAW 的程序格式和解释器接口。","program_form":"包含 LoRA 适配器、运行提示词、函数说明和解释器元数据的 PAW 程序。","conditioning":"函数说明指导教师模型生成样本，PAW 快速编译器提供初始适配器与运行提示词。","execution":"冻结的 Qwen3-0.6B 解释器执行保存的函数，不再调用教师模型。"}}}],"precursors":[]}