Towards Intrinsic Interpretability of Large Language Models: A Survey of Design Principles and Architectures

Published in Annual Meeting of the Association for Computational Linguistics (ACL 2026 Main Conference, Oral), 2026

We present a systematic survey on intrinsic interpretability for Large Language Models. Unlike existing surveys focusing on post-hoc explanation methods, we focus on transparency built directly into model architectures and computations. We categorize existing approaches into five design paradigms: functional transparency, concept alignment, representational decomposability, explicit modularization, and latent sparsity induction.

Oral presentation

Selected for an Oral presentation at ACL 2026. The final official schedule lists 486 unique main-conference Oral papers: approximately 4.00% of the 12,148 submissions used in the conference’s acceptance-rate calculation, or 21.17% of the 2,296 accepted main-conference papers. These percentages are calculated from the official counts, rather than a separately announced Oral acceptance rate.

Sources: official detailed schedule (unique MAIN paper IDs assigned to Oral sessions, excluding CL/TACL papers) and Program Chairs’ report, p. vii (submission and acceptance totals).

Yutong Gao*, Qinglin Meng*, Yuan Zhou, Liangming Pan
Download Paper