# How Does Controllability Emerge In Language Models During Pretraining?

arXiv: https://arxiv.org/abs/2508.01892 · HTML: https://mlsanity.com/paper/2508.01892/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Builds on 31 of its 40 references that are in the map.
- First posted on arXiv 3 Aug 2025.

## Builds on (most cited first, 30 of 31)

- [Language Models are Few-Shot Learners](https://mlsanity.com/paper/2005.14165/) (2020)
- [Think you have Solved Question Answering? Try ARC, the AI2 Reasoning Challenge](https://mlsanity.com/paper/1803.05457/) (2018)
- [Sparks of Artificial General Intelligence: Early experiments with GPT-4](https://mlsanity.com/paper/2303.12712/) (2023)
- [Deep Inside Convolutional Networks: Visualising Image Classification Models and Saliency Maps](https://mlsanity.com/paper/1312.6034/) (2013)
- [Can a Suit of Armor Conduct Electricity? A New Dataset for Open Book Question Answering](https://mlsanity.com/paper/1809.02789/) (2018)
- [Scaling LLM Test-Time Compute Optimally can be More Effective than Scaling Model Parameters](https://mlsanity.com/paper/2408.03314/) (2024)
- [BLOOM: A 176B-Parameter Open-Access Multilingual Language Model](https://mlsanity.com/paper/2211.05100/) (2022)
- [Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling](https://mlsanity.com/paper/2304.01373/) (2023)
- [Understanding intermediate layers using linear classifier probes](https://mlsanity.com/paper/1610.01644/) (2016)
- [Representation Engineering: A Top-Down Approach to AI Transparency](https://mlsanity.com/paper/2310.01405/) (2023)
- [Inference-Time Intervention: Eliciting Truthful Answers from a Language Model](https://mlsanity.com/paper/2306.03341/) (2023)
- [Steering Llama 2 via Contrastive Activation Addition](https://mlsanity.com/paper/2312.06681/) (2023)
- [RACE: Large-scale ReAding Comprehension Dataset From Examinations](https://mlsanity.com/paper/1704.04683/) (2017)
- [Progress measures for grokking via mechanistic interpretability](https://mlsanity.com/paper/2301.05217/) (2023)
- [Discovering Latent Knowledge in Language Models Without Supervision](https://mlsanity.com/paper/2212.03827/) (2022)
- [Are Emergent Abilities of Large Language Models a Mirage?](https://mlsanity.com/paper/2304.15004/) (2023)
- [UnifiedQA: Crossing Format Boundaries With a Single QA System](https://mlsanity.com/paper/2005.00700/) (2020)
- [What do you learn from context? Probing for sentence structure in contextualized word representations](https://mlsanity.com/paper/1905.06316/) (2019)
- [Emergent World Representations: Exploring a Sequence Model Trained on a Synthetic Task](https://mlsanity.com/paper/2210.13382/) (2022)
- [A Survey of Transformers](https://mlsanity.com/paper/2106.04554/) (2021)
- [Linguistic Knowledge and Transferability of Contextual Representations](https://mlsanity.com/paper/1903.08855/) (2019)
- [Visualizing and Understanding Neural Models in NLP](https://mlsanity.com/paper/1506.01066/) (2015)
- [Post-hoc Interpretability for Neural NLP: A Survey](https://mlsanity.com/paper/2108.04840/) (2021)
- [Inspecting and Editing Knowledge Representations in Language Models](https://mlsanity.com/paper/2304.00740/) (2023)
- [LLM360: Towards Fully Transparent Open-Source LLMs](https://mlsanity.com/paper/2312.06550/) (2023)
- [A Comprehensive Survey on Applications of Transformers for Deep Learning Tasks](https://mlsanity.com/paper/2306.07303/) (2023)
- [MAP-Neo: Highly Capable and Transparent Bilingual Large Language Model Series](https://mlsanity.com/paper/2405.19327/) (2024)
- [What's the Magic Word? A Control Theory of LLM Prompting](https://mlsanity.com/paper/2310.04444/) (2023)
- [Evaluating Saliency Methods for Neural Language Models](https://mlsanity.com/paper/2104.05824/) (2021)
- [LiSeCo: Linear Semantic Control for Language Generation](https://mlsanity.com/paper/2405.15454/) (2024)

## Abstract (arXiv)

Language models can be steered by modifying their internal representations to control concepts such as emotion, style, or truthfulness in generation. However, the conditions for an effective intervention remain unclear and are often validated through heuristics and trial-and-error. To fill this gap, we demonstrate that intervention efficacy, measured by linear steerability (i.e., the ability to adjust output via linear transformations of hidden states), emerges during intermediate stages of training. Moreover, even closely related concepts (e.g., anger and sadness) exhibit steerability emergence at distinct stages of training. To better interpret the dynamics of steerability during training, we adapt existing intervention techniques into a unified framework, referred to as the "Intervention Detector" (ID), which is designed to reveal how linear steerability evolves over the course of training through hidden state and representation analysis. ID reveals that concepts become increasingly linearly separable in the hidden space as training progresses, which strongly correlates with the emergence of linear steerability. We further introduce ID-based metrics, such as heatmaps, entropy trends, and cosine similarity, to help interpret how linear steerability evolves throughout training. In addition, we apply ID across different model families to ensure the generality of our findings on steerability dynamics.