publications

2026

1. NeurIPS Workshop

   MechInterp

   Circuit-Diff: Factual Edit-based Intervention Method for Localizing Knowledge in Attribution Graphs

   E. G. Friedman and X. Song

   In NeurIPS 2026 Workshop on Interpretability as a Science, Sydney, Australia, Dec 2026

   Accepted

   Abs arXiv PDF Supp Code Website

   Mechanistic interpretability defines features as the fundamental units of a neural network and circuits as the weighted subgraphs that carry out its computation. Because individual neurons are polysemantic, Cross-Layer Transcoders (CLTs) were introduced as a way to approximate a model’s circuits by generating an attribution graph. The nodes of that graph, however, are unlabeled features: reading a graph means pruning it and then working out by hand what each surviving node means. To make CLTs easier to use for circuit discovery, we introduce Circuit-Diff, which intervenes on the model itself with a low-rank factual edit and takes the features whose role in the attribution graph changes under that edit as related to the edited knowledge. On the edits we examine, the flagged nodes are not only detectors of the object token: read off the CLT’s released feature dashboards, they include features for the history, geography and associations surrounding the old and new objects. We formalize the method, measure how reliable a frozen CLT remains after a factual edit, test the selected nodes causally by patching them on up to 24 CounterFact edits, give a case study, and release an open-source implementation built on the circuit-tracer package, together with two further tools (multi-prompt aggregation and rule-based supernode labeling).

2. NeurIPS

   AI For Science

   InformedXRD: Reproducible Benchmarks and Physics-Informed Evaluation for Powder Diffraction Symmetry Classification

   E. G. Friedman, E. J. Baggett, A. Shetty, and 5 more authors

   In Advances in Neural Information Processing Systems 40 (NeurIPS 2026), Evaluations & Datasets Track, Sydney, Australia, Dec 2026

   Accepted

   Abs Website

   Machine learning for powder X-ray diffraction (PXRD) symmetry classification is being trained at scale, yet the field lacks a common evaluation standard: papers report results on different subsets of the RRUFF mineral diffraction database, under different target taxonomies, with different preprocessing and different summary metrics, and published models are rarely released in a form that permits rescoring. We present InformedXRD, an evaluation framework with four parts: (i) a machine-readable mapping from 230 space groups to 99 extinction groups, the finest classes that powder diffraction can identify from systematic absences alone; (ii) two algorithmically curated real-data benchmarks, RRUFF-473 and RRUFF-325, whose inclusion rules are specified as code and reproduce from a frozen upstream snapshot; (iii) a prior-only frequency baseline and stratified reporting protocol that expose when apparent gains are driven by label imbalance or nuisance-fit severity; and (iv) a hierarchy-aware error metric on the condensed translationengleiche (same-lattice) subgroup graph that measures the crystallographic severity of misclassifications. Using four publicly released ViT checkpoints from Baggett et al. (2026) and an in-house 1D residual CNN as case studies, we show that evaluation design changes model ranking: standard paired tests point in different directions depending on whether one evaluates Top-1 or Top-5, and stratified analysis exposes a label-entropy confound that aggregate accuracy hides. All mapping tables, benchmark code, evaluation scripts, and case-study checkpoints are released as shared infrastructure for future PXRD-ML evaluation.

3. APS PRX Intelligence

   AI For Science

   Attention Is Not All You Need for Diffraction

   E. J. Baggett, E. G. Friedman, A. Shetty, and 5 more authors

   PRX Intelligence, Sep 2026

   Accepted 9 September 2026; in press

   Abs DOI arXiv HTML PDF Supp Code

   Determining crystal symmetry from powder X-ray diffraction (PXRD) is a central problem in materials characterization, yet multiple space groups produce indistinguishable patterns, and automated methods plateau at low accuracy on real data. The open question is where the recoverable symmetry information is lost. We cast symmetry determination as decoding through a noisy measurement channel and measure its capacity directly. On our real-data benchmark, a Bayes-optimal decoder that is given the pattern of systematically absent reflections recovers 67% Top-1 at the benchmark’s class distribution, yet end-to-end model accuracy is only  10%. An explicit two-stage decomposition shows that nearly all of this gap comes from determining which reflections are truly absent in a noisy, degraded diffraction pattern, not from mapping a known absence pattern to the correct symmetry. To probe this limit with a strong decoder, we introduce a physics-informed transformer, with an explicit sin²θ coordinate channel, physics-aware positional encoding, and a structured multi-task decoder that separates geometric rule learning from holistic pattern recognition, trained on a synthetic-distribution curriculum chosen to reduce train-test overlap risk. This model is competitive with or stronger than published baselines and still hits the same wall, showing that the limit is structural rather than a deficiency of any one architecture. We also show that several recent state-of-the-art models share this wall. Mapping residual errors onto the directed acyclic graph of maximal symmetry-reducing (translationengleiche) subgroups shows that the errors are physically structured, local on the hierarchy and biased toward lower-symmetry descendants, a signature of erased absence cues, and that this hierarchy makes classical Pawley verification tractable by bounding the search. The bottleneck is measurement, not classification: the productive lever is better absence recovery, not larger classifiers.