{"arxiv_id":"2602.14111","incoming":{"papers":9,"sentences":12,"groups":[{"kind":"supports_this_paper","label":"Cites it as support","papers":4,"sentences":4,"quotes":[{"arxiv_id":"2607.20652","title":"Scaling Interpretable Transformers with Parity Bottleneck Layers","url":"/paper/2607.20652","section":"Fine-grained causal intervention","sentence":"For post-hoc SAEs, this is not guaranteed: SAEs trained on randomly initialized transformers produce auto-interpretability scores approaching those obtained from trained models [ref] , and random baselines that replace SAE feature directions or activation patterns with noise match fully-trained SAEs on auto-interpretability, sparse probing, and causal editing [ref] .","intent_p":0.96,"contests_p":null},{"arxiv_id":"2607.19386","title":"Building Fast, Evaluating Slow: Pipeline Choices Dominate Autointerpretability Score Variance","url":"/paper/2607.19386","section":"Introduction","sentence":"Proxy metrics such as reconstruction error and sparsity reflect optimization behavior rather than semantic quality [ref] , and SAEs that perform well on these metrics have been shown to detect only a small subset of true features [ref] .","intent_p":0.98,"contests_p":null},{"arxiv_id":"2608.13337","title":"Where You Measure Decides What You Measure: Position Selection in Ablation-Based SAE Evaluation","url":"/paper/2608.13337","section":"Related work","sentence":"[ref] show that seed alone substantially changes which features are learned, TopK more than ReLU ([ref]); [ref] find automated interpretability scores do not separate trained transformers from random ones; [ref] show frozen-random baselines match fully-trained SAEs across reconstruction, interpretability, sparse probing, and causal editing ($0.73$ vs $0.72$); and [ref] finds SAEBench's TPP and SCR fail three independent reliability lenses.","intent_p":0.96,"contests_p":null},{"arxiv_id":"2608.07594","title":"Scaling Inherently Interpretable Language Models","url":"/paper/2608.07594","section":"Why post-hoc methods fail to satisfy the interpretability requirements","sentence":"More concerning, recent work shows that SAEs trained with different random seeds produce different feature sets, and that random SAE baselines match fully-trained SAEs on sparse probing and causal editing metrics [ref] .","intent_p":0.99,"contests_p":null}]}],"unlabelled_sentences":8},"outgoing":{"papers":0,"sentences":0,"groups":[],"unlabelled_sentences":0},"claims":[],"thresholds":{"intent_p":0.8,"supports_p":0.95,"contests_p":0.7},"method":"Citing sentences come from each paper's LaTeX source; every \\cite marks the sentence and the works it cites. Pith's classifier typed each sentence. A label is shown at intent probability 0.8 or more (0.95 for support), a contest at 0.7 or more. Measured accuracy against two blind model families is at https://pith.science/accuracy. The quoted sentences are the citing authors' own.","held_up":null,"_links":{"html":"https://pith.science/paper/2602.14111"}}