Steerling-8B remains competitive with open peer models trained on substantially 2-16x more compute, suggesting a different scaling paradigm: interpretability can be designed into training, and it improves with scale.
Abstract
Interpretability is often treated as a tax on capability: language models are trained as opaque systems, then explained after the fact, with methods whose reliability is difficult to establish. In this work, we challenge this premise. Rather than reverse-engineering a model, we make interpretability a constraint of the training pipeline, optimized alongside the language modeling objective. Across three orders of magnitude of compute, on both autoregressive and diffusion language models, interpretability scales with capability rather than against it. Surprisingly, model representations become more disentangled and aligned with human-understandable concepts with scale. We instantiate the training-time recipe with Steerling-8B, a diffusion language model with a causal attention mask. For any group of generated tokens, Steerling-8B attributes the output to relevant input tokens, human-understandable concepts, and training data. This enables closed-loop intervention: diagnose an output through its concept or feature attribution, retrieve similar training data, and correct the behavior through concept steering without retraining. Steerling-8B remains competitive with open peer models trained on substantially 2-16x more compute, suggesting a different scaling paradigm: interpretability can be designed into training, and it improves with scale.
This tutorial provides a comprehensive, end-to-end view of LLM interpretability, transitioning from microscopic neural analysis to macroscopic application and deployment, and explores how these interpretability paradigms scale and inspire the design of frontier architectures, agentic systems, and thinking models.
Wei Zhang, Zhengfu He, Lucia Zhang et al.· Proceedings of the 32nd ACM...· 0 citations
This survey reviews LLM interpretability through the lens of actionability, presenting a taxonomy of attributional and mechanistic approaches, along with emerging methods tailored to vision–language models (VLMs), and examining how actionable interpretability supports downstream objectives.
Jie Cai, Mafizur Rahman, James Enouen et al.· 0 citations
An integrated conceptual frame-work that couples attention- and perturbation-based explainability with lightweight hallucination-detection signals and token-efficient inference strategies is presented, and a set of cross-cutting consistency metrics are instrumented with a set of cross-cutting consistency metrics.
Sakshi Parate, Shreyans Sanyal· Advanced International Journ...· 0 citations
The results demonstrate that suppressing entity-linked supervision during pretraining induces a shift in epistemic behavior: KLLMs rely less on parametric knowledge and more on external evidence, leading to improved reliability under realistic conditions, suggesting that pretraining-time control over knowledge acquisition can complement retrieval-augmented and tool-based systems by providing a more evidence-sensitive base model.
Roi Cohen, Yvan Carr'e, Nick Lechtenborger et al.· 0 citations
This work presents the first systematic study of inverse relation directionality in LLMs, using a benchmark consisting of 5,457 instances spanning 27 distinct inverse relation labels and reveals systematic asymmetries in inverse relation classification across LLMs.
Large language models (LLMs) are increasingly used to generate synthetic data, in which tabular data constitute a fundamental data modality across a wide range of domains. Yet, current evaluation practices often provide limited insights into whether the synthetic data preserve real data-generating relationships or introduce plausible-looking artifacts. We present a conceptually simple, interpretable auditing framework that compares the explanatory structure induced by real versus synthetic data. The key idea is to use a transparent rule-based model as a shared explanatory language: we extract rules from real data to summarize how features relate to labels, then examine how this rule structure changes when explained us-ing LLM-generated data. Importantly, these rules are derived by an independent rule auditor rather than by the generator itself. The resulting “explanation shift” reveals which relationships are preserved, weakened, removed, or newly introduced by the generator, offering actionable diagnostics beyond aggregate fidelity scores. We further provide a theoretical perspective that links explanation shift and cross-domain predictive gaps to distribution mismatch within an interpretable hypothesis class. Overall, our approach turns synthetic data evaluation into a human-auditable comparison of explanations, improving transparency for LLM-based tabular synthesis.
Yue Yang, Fan Yang, Yu Bai et al.· Annual Meeting of the Associ...· 0 citations