Steering Language Models in Multi-Token Generation: A Case Study on Tense and Aspect
Using linear discriminant analysis (LDA), the study identifies orthogonal directions encoding tense and aspect in LLM residual space, enabling causal control during multi-token generation.
Key Findings
Methodology
The research employs linear probes combined with linear discriminant analysis (LDA) to identify feature directions for tense and aspect in the residual space of large language models (LLMs). By training classifiers on aggregated hidden states across layers, the study extracts vectors representing categorical features. These vectors are orthogonal, reflecting the independence of tense and aspect. Subsequently, concept steering manipulates these vectors during generation, enabling causal control over grammatical features. Experiments involve Llama-3.1-8B and Qwen-2.5-7B models across tasks like random sentence generation, repetition, and translation, validating the method's effectiveness.
Key Results
- The feature directions for tense (past, present, future) and aspect (simple, progressive, perfect, perfect progressive) are identified with over 90% classification accuracy, and their vectors are nearly orthogonal (cosine similarity ~0.02), confirming structured encoding.
- Concept steering successfully controls grammatical features during multi-word generation, achieving success rates above 94% for tense and 50-70% for aspect, across various tasks, with parameter tuning affecting control strength and stability.
- Adjusting steering parameters such as position, strength, and duration significantly reduces undesired effects like topic shift and degeneration. The projection subtraction method enhances selectivity, maintaining output quality while improving control precision.
Significance
This work uncovers the structured, human-like encoding of tense and aspect in LLMs, advancing understanding of internal syntactic representations. The causal control framework opens pathways for precise grammatical manipulation, improving model interpretability and reliability in applications like machine translation, dialogue, and low-resource language processing. It bridges the gap between correlational analysis and causal intervention, contributing to more transparent and controllable AI systems.
Technical Contribution
The study introduces a novel integration of LDA with residual space probing to identify independent, orthogonal feature directions for complex grammatical phenomena. It develops a concept steering approach that applies targeted, parameterized interventions at the token level during generation. The combination of these techniques offers a new paradigm for interpretable, causal manipulation of model outputs, with broad applicability to linguistic features beyond syntax.
Novelty
This is the first work to systematically identify and leverage orthogonal feature directions for multi-dimensional grammatical phenomena like tense and aspect in large language models. Unlike prior binary or token-level analyses, it achieves sentence-level causal control through residual space manipulation, representing a significant step forward in interpretability and controllability of deep language models.
Limitations
- Control effectiveness varies with model architecture, layer depth, and task complexity, requiring manual tuning of parameters, which limits scalability and automation.
- The approach relies on the assumption of linear separability, which may not hold for more nuanced or ambiguous grammatical features, especially in low-resource or multilingual settings.
- Computational overhead increases with the number of features and layers involved, posing challenges for real-time or large-scale deployment.
Future Work
Future research will focus on automating parameter tuning via reinforcement learning or Bayesian optimization, extending the framework to more complex grammatical structures (e.g., conditionals, passive voice), and exploring multi-lingual generalization. Integrating semantic features and contextual cues could further enhance control robustness and applicability across diverse NLP tasks.
AI Executive Summary
Large language models (LLMs) have demonstrated remarkable capabilities in generating coherent and grammatically correct text. However, understanding how they internally encode complex syntactic phenomena like tense and aspect remains an open challenge. Traditional evaluation methods focus on behavioral metrics such as perplexity or acceptability scores, which do not reveal the underlying internal representations. Recent interpretability studies have begun to probe specific linguistic features, but these often target binary distinctions or single-token predictions, limiting insights into multi-word, hierarchical phenomena.
This study advances the field by applying linear discriminant analysis (LDA) to the residual space of LLMs, systematically identifying orthogonal directions that encode tense and aspect. These directions are discovered through probing classifiers trained on aggregated hidden states across layers, revealing that the model’s internal representations form structured subspaces for these grammatical categories. The vectors corresponding to tense (past, present, future) and aspect (simple, progressive, perfect, perfect progressive) are nearly orthogonal, indicating independent encoding.
Building on this, the authors introduce a concept steering framework that manipulates these feature directions during multi-token generation. By adjusting the residual activations at specific layers and positions, they achieve causal control over the grammatical properties of the generated text. Experimental results show high success rates (>94%) in controlling tense and moderate success (50-70%) for aspect, across tasks like random sentence generation, repetition, and translation. The effectiveness depends on parameters such as steering strength, location, and duration, with strategies like projection subtraction improving control precision.
These findings have significant implications for interpretability and controllability in NLP. They demonstrate that models encode hierarchical, human-like grammatical structures in a structured, manipulable manner. The proposed methods pave the way for more transparent, adjustable language generation systems, with potential applications in machine translation, dialogue, and low-resource language processing. Future work will focus on automating parameter tuning, extending to more complex grammatical phenomena, and enhancing multi-lingual generalization, aiming for robust, scalable grammatical control in AI language systems.
Deep Dive
Abstract
Large language models (LLMs) are able to generate grammatically well-formed text, but how do they encode their syntactic knowledge internally? While prior work has focused largely on binary grammatical contrasts, in this work, we study the representation and control of two multidimensional hierarchical grammar phenomena - verb tense and aspect - and for each, identify distinct, orthogonal directions in residual space using linear discriminant analysis. Next, we demonstrate causal control over both grammatical features through concept steering across three generation tasks. Then, we use these identified features in a case study to investigate factors influencing effective steering in multi-token generation. We find that steering strength, location, and duration are crucial parameters for reducing undesirable side effects such as topic shift and degeneration. Our findings suggest that models encode tense and aspect in structurally organized, human-like ways, but effective control of such features during generation is sensitive to multiple factors and requires manual tuning or automated optimization.