Antibody Language Models SAEs + Non-Linear Representation Dilemma = H2 2026 mech-interp expands across domains while questioning foundational methodology

Two interpretability papers reflect H2 2026 dual direction: domain-specific applications expand mech-interp scope (antibody language models), foundational-question papers challenge causal-abstraction sufficiency (non-linear representation dilemma). Both directions productive — methodology expansion alongside methodology reassessment.

Antibody language models SAE application + Non-Linear Representation Dilemma foundational question together demonstrate H2 2026 mech-interp dual research direction.

The methodology-expansion direction

Domain-specific applications extend mech-interp methodology beyond general-purpose LLMs to specialized scientific domains. Antibody language models, protein language models, code-correctness models, vision-language models — each domain provides distinct application context for SAE + concept-bottleneck + steerability methodology. Expansion produces methodology validation across diverse domains beyond the general-LLM baseline.

The methodology-reassessment direction

Foundational-question papers (causal abstraction sufficiency, falsifiability framework, concept-annotation evaluation) challenge implicit assumptions in mainstream methodology. If causal abstraction is insufficient for non-linear representations + activation-pattern correlations don't establish causal claims + proxy metrics don't measure semantic correspondence, substantial methodology investment needs reassessment.

The combined direction productive

Methodology expansion (validate methodology across diverse domains) + methodology reassessment (question methodology assumptions) together represent productive research direction. Yesterday's credibility-bar elevation analysis reinforces the pattern — H2 2026 mech-interp credibility-bar elevates substantively across multiple methodology dimensions simultaneously.

arXiv — Mechanistic Interpretability of Antibody Language Models Using SAEs → · arXiv — The Non-Linear Representation Dilemma →