Jack Morris

Jack Morris, who publishes research as John X. Morris, studies the information retained in language models and text embeddings. His collaborative work includes reconstructing text from embeddings, incorporating neighboring documents into retrieval representations, estimating language-model memorization capacity, and approximating useful training data from model weights.[3][5][6][7][8]

In the May 2025 paper How much do language models memorize?, Morris and collaborators separate unintended memorization—information about a specific dataset—from generalization, information about the underlying data-generation process. Eliminating generalization allows them to estimate capacity through total memorization. Across hundreds of transformer models ranging from 500,000 to 1.5 billion parameters, they report approximately 3.6 bits per parameter for the studied GPT-style models. This is an empirical estimate under that measurement setup.[7]

The team also reports that, as training datasets grow, models memorize until their capacity fills; then “grokking” begins, and dataset-specific memorization declines as generalization develops. Morris's coauthors were Chawin Sitawarin, Chuan Guo, Narine Kokhlikyan, G. Edward Suh, Alexander M. Rush, Kamalika Chaudhuri, and Saeed Mahloujifar.[7]

In June 2025, Morris, Junjie Oscar Yin, Woojeong Kim, Vitaly Shmatikov, and Rush published Approximating Language Model Training Data from Weights. Their gradient-based method uses both original and finetuned model weights to select matching examples from a large public text corpus. It finds useful substitute training data without requiring knowledge of the original dataset. The authors report AG News classification accuracy of 80%, compared with 65% using random data and an 88% expert benchmark. For supervised finetuning on MSMARCO documents, they report perplexity of 2.3, compared with an expert LLAMA model's 2.0. These are task-specific performance comparisons, rather than exact-data recovery rates.[8]

Before these 2025 studies, Morris began research as a University of Virginia undergraduate and created TextAttack during 2019–2020. In his personal biography, he describes subsequently becoming a Google Brain Resident and receiving a Cornell University PhD at Cornell Tech. He worked closely with Sasha Rush and Vitaly Shmatikov, received NSF GRFP support, and spent about a third of his PhD as a part-time student researcher at Meta FAIR.[2]

His earlier embedding research supplies a concrete example of recovering information from a learned representation. In Text Embeddings Reveal (Almost) As Much As Text, submitted in October 2023 and accepted at EMNLP 2023, Morris, Volodymyr Kuleshov, Shmatikov, and Rush frame inversion as controlled generation: produce text whose embedding approaches a fixed target in latent space. Their method repeatedly corrects and re-embeds the generated text. The authors report exact recovery of 92% of 32-token inputs in their experiments involving two embedding models, and recovery of full names from clinical-note embeddings—a specific privacy risk demonstrated experimentally.[5]

With Rush, Morris then proposed Contextual Document Embeddings, first submitted in October 2024. Their design incorporates neighboring documents through two complementary methods: a contrastive training objective that includes neighbors in the intra-batch contextual loss, and an encoder architecture that puts neighbor information into the document representation. The paper reports improvements over biencoders, especially out of domain, and state-of-the-art MTEB results at the time without hard-negative mining, score distillation, dataset-specific instructions, intra-GPU example-sharing, or extremely large batches.[6]

In his April 9, 2025 essay There Are No New Ideas in AI… Only New Datasets, Morris argues that major advances opened access to new sources of learning: labeled images, internet text, human preferences, and verifiers such as calculators and compilers. He also credits systems research with making models cheaper. His proposed direction is to seek new data sources, with video and robotic sensor data offered as possibilities rather than certain next breakthroughs: “It’s hard to say whether YouTube or robots or something else will be the Next Big Thing for AI.”[9]

The essay includes a correction to its RLHF chronology, acknowledging Paul Christiano's 2017 paper and noting that aligning models to human preferences is older still. Morris thanks readers who pointed out the earlier work.[9]