Mechanistic interpretability research faces a gap between analyzing simple circuits in toy tasks and discovering features in large models. To bridge this gap, we propose text-to-SQL generation as an ideal task to study, as it combines the formal structure of toy tasks with real-world complexity. We introduce TinySQL, a synthetic dataset, progressing from basic to advanced SQL operations, and train models ranging from 33M to 1B parameters to establish a comprehensive testbed for interpretability. We apply multiple complementary interpretability techniques, including Edge Attribution Patching and Sparse Autoencoders, to identify minimal circuits and components supporting SQL generation. We compare circuits for different SQL subskills, evaluating their minimality, reliability, and identifiability. Finally, we conduct a layerwise logit lens analysis to reveal how models compose SQL queries across layers: from intent recognition to schema resolution to structured generation. Our work provides a robust framework for probing and comparing interpretability methods in a structured, progressively complex setting.

Controlling multiple behavioral attributes in large language models (LLMs) at inference time is a challenging problem due to interference between attributes and the limitations of linear steering methods, which assume additive behavior in activation space and require per-attribute tuning. We introduce K-Steering, a unified and flexible approach that trains a single non-linear multi-label classifier on hidden activations and computes intervention directions via gradients at inference time. This avoids linearity assumptions, removes the need for storing and tuning separate attribute vectors, and allows dynamic composition of behaviors without retraining. To evaluate our method, we propose two new benchmarks, TONEBANK and DEBATEMIX, targeting compositional behavioral control. Empirical results across 3 model families, validated by both activation-based classifiers and LLM-based judges, demonstrate that K-Steering outperforms strong baselines in accurately steering multiple behaviors.

Intent detection is a major task in Natural Language Understanding (NLU) and is the component of dialogue systems for interpreting users’ intentions based on their utterances. Many works have explored detecting intents by assuming that each utterance represents only a single intent. Such systems have achieved very good results; however, intent detection is a far more challenging task in typical real-world scenarios, where each user utterance can be highly complex and express multiple intents. Therefore, in this paper, we propose PCMID, a novel Multi-Intent Detection framework enabled by Prototypical Contrastive Learning under a supervised setting. The PCMID model can learn multiple semantic representations of a given user utterance under the context of different intent labels in an optimized semantic space. Our experiments show that PCMID achieves the current state-of-the-art performance on both multiple public benchmark datasets and a private real-world dataset for the multi-intent detection task.