# Categorical Prior Lock-in: Why In-Context Learning Fails for Structured Data

arXiv: https://arxiv.org/abs/2606.11961 · HTML: https://mlsanity.com/paper/2606.11961/

## Key facts

- Cited by 1 papers in the map (data through 14 Oct 2026).
- Builds on 8 of its 13 references that are in the map.
- First posted on arXiv 10 Jun 2026.

## Builds on (most cited first, 8 of 8)

- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [LoRA: Low-Rank Adaptation of Large Language Models](https://mlsanity.com/paper/2106.09685/) (2021)
- [Finetuned Language Models Are Zero-Shot Learners](https://mlsanity.com/paper/2109.01652/) (2021)
- [Modeling Tabular data using Conditional GAN](https://mlsanity.com/paper/1907.00503/) (2019)
- [Language Models are Realistic Tabular Data Generators](https://mlsanity.com/paper/2210.06280/) (2022)
- [REaLTabFormer: Generating Realistic Relational and Tabular Data using Transformers](https://mlsanity.com/paper/2302.02041/) (2023)
- [TabuLa: Harnessing Language Models for Tabular Data Synthesis](https://mlsanity.com/paper/2310.12746/) (2023)
- [A Comparative Analysis of Instruction Fine-Tuning LLMs for Financial Text Classification](https://mlsanity.com/paper/2411.02476/) (2024)

## Built on by (most cited first, 1 of 1)

- [Attention Sensitivity Is Not Enough: Dissociating Attention-Level and Behavioural In-Context Learning under Fine-Tuning](https://mlsanity.com/paper/2609.00064/) (2026)

## Abstract (arXiv)

Large language models (LLMs) are increasingly used as conditional generators for structured data, relying on in-context learning (ICL) to adapt to new distributions without parameter updates. We investigate the limits of ICL for structured generation under distribution mismatch, using high-cardinality tabular data as a controlled test case, and identify a structural failure mode we term textitcategorical prior lock-in: the inability of ICL to update the model's prior over token distributions inherited from pre-training. Across two 7B-parameter open-weight models, ICL improves numerical fidelity with additional examples but exhibits a sharp ceiling on categorical distributions, failing to reproduce rare classes entirely. Parameter-efficient fine-tuning (LoRA) overcomes these limitations but introduces measurable memorization risk and, in some cases, destabilizes structured output generation, highlighting a fundamental trade-off between adaptability and privacy.