# Prototype Language Models

arXiv: https://arxiv.org/abs/2607.00510 · HTML: https://mlsanity.com/paper/2607.00510/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Builds on 21 of its 101 references that are in the map.
- First posted on arXiv 1 Jul 2026.

## Builds on (most cited first, 21 of 21)

- [Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge](https://mlsanity.com/paper/1803.05457/) (2018)
- [Eliciting Latent Predictions from Transformers with the Tuned Lens](https://mlsanity.com/paper/2303.08112/) (2023)
- [TinyStories: How Small Can Language Models Be and Still Speak Coherent English?](https://mlsanity.com/paper/2305.07759/) (2023)
- [Studying Large Language Model Generalization with Influence Functions](https://mlsanity.com/paper/2308.03296/) (2023)
- [If Influence Functions are the Answer, Then What is the Question?](https://mlsanity.com/paper/2209.05364/) (2022)
- [Scaling Up Influence Functions](https://mlsanity.com/paper/2112.03052/) (2021)
- [Adaptive Gradient Methods at the Edge of Stability](https://mlsanity.com/paper/2207.14484/) (2022)
- [Gradient Routing: Masking Gradients to Localize Computation in Neural Networks](https://mlsanity.com/paper/2410.04332/) (2024)
- [MAGIC: Near-Optimal Data Attribution for Deep Learning](https://mlsanity.com/paper/2504.16430/) (2025)
- [FlexOlmo: Open Language Models for Flexible Data Use](https://mlsanity.com/paper/2507.07024/) (2025)
- [This Looks Better than That: Better Interpretable Models with ProtoPNeXt](https://mlsanity.com/paper/2406.14675/) (2024)
- [Beyond Data Filtering: Knowledge Localization for Capability Removal in LLMs](https://mlsanity.com/paper/2512.05648/) (2025)
- [Revisiting inverse Hessian vector products for calculating influence functions](https://mlsanity.com/paper/2409.17357/) (2024)
- [Large-Scale Training Data Attribution for Music Generative Models via Unlearning](https://mlsanity.com/paper/2506.18312/) (2025)
- [Generalized Group Data Attribution](https://mlsanity.com/paper/2410.09940/) (2024)
- [Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods](https://mlsanity.com/paper/2412.03906/) (2024)
- [Automated Interpretability and Feature Discovery in Language Models with Agents](https://mlsanity.com/paper/2605.01555/) (2026)
- [Interpretable Next-token Prediction via the Generalized Induction Head](https://mlsanity.com/paper/2411.00066/) (2024)
- [LoRIF: Low-Rank Influence Functions for Scalable Training Data Attribution](https://mlsanity.com/paper/2601.21929/) (2026)
- [Cite Pretrain: Retrieval-Free Knowledge Attribution for Large Language Models](https://mlsanity.com/paper/2506.17585/) (2025)
- [Natively Unlearnable Large Language Models](https://mlsanity.com/paper/2606.13873/) (2026)

## Built on by (most cited first, 2 of 2)

- [Scaling Inherently Interpretable Language Models](https://mlsanity.com/paper/2608.07594/) (2026)
- [When Are Concept Bottleneck Model Explanations Faithful and Compact?](https://mlsanity.com/paper/2610.06285/) (2026)

## Abstract (arXiv)

Knowing which training examples drive outputs is fundamental to auditing, correcting, and understanding language models, yet for modern LLMs this remains expensive, approximate, and largely post-hoc. Standard language models generate tokens through a dense network pathway, causing training data's influence to be distributed across parameters rather than organized along explicit, traceable components. We introduce a prototype language model architecture, Prototypes for Interpretable Sequence Modeling (PRISM), that forms each prediction via a sparse, non-negative mixture of learned prototypes, trained with clustering objectives that anchor each prototype to coherent neighborhoods of training examples. Across architectures from 130M to 1.6B parameters trained on up to 50B tokens, prototype language models either surpass or remain within 2.5 percentage points on average downstream accuracy of matched dense baselines. We show that sparse prototype structure localizes curvature in the loss landscape, yielding a more tractable Hessian and enabling training data attribution that is ~500x faster than post hoc baselines when consuming equivalent memory. Calibrating linear prototype controllers can improve downstream accuracy by roughly 3 points while tracing those corrections back to training neighborhoods, and targeted prototype suppression can remove model behaviors without finetuning or measurable loss in generation quality.