{"total":12,"items":[{"citing_arxiv_id":"2606.28486","ref_index":46,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Spectral phase transitions and trainability in neural network learning dynamics","primary_cat":"cond-mat.dis-nn","submitted_at":"2026-06-26T18:00:01+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"SGD on neural network weights induces a BBP phase transition that detaches signal eigenvalues from the random bulk, yielding an analytically solvable phase diagram for trainability in a linear teacher-student model.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.06333","ref_index":42,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability","primary_cat":"cs.LG","submitted_at":"2026-06-04T16:08:25+00:00","verdict":"CONDITIONAL","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"SASA replaces single-vector decoders in SAEs with learned subspaces plus block sparsity and nuclear-norm regularization, proving that a single group becomes the global minimizer once block size meets intrinsic dimension and yielding polynomial rather than exponential sample complexity.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2606.04754","ref_index":115,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Beyond Structural Symmetries: Linear Mode Connectivity via Neuron Identifiability","primary_cat":"cs.LG","submitted_at":"2026-06-03T11:37:58+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":7.0,"formal_verification":"none","one_line_summary":"Neural networks admit large families of approximately equivalent solutions via neuron identifiability even without structural symmetry, enabling linear low-loss merging paths without prior alignment.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.15622","ref_index":83,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Position: Zeroth-Order Optimization in Deep Learning Is Underexplored, Not Underpowered","primary_cat":"cs.LG","submitted_at":"2026-05-15T05:11:43+00:00","verdict":"UNVERDICTED","verdict_confidence":"UNKNOWN","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Zeroth-order optimization is underexplored rather than underpowered in deep learning, with limitations stemming from full-space designs that can be addressed via subspace, spectral, and systems-aware approaches.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.12128","ref_index":48,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Metaphor Is Not All Attention Needs","primary_cat":"cs.CL","submitted_at":"2026-05-12T13:50:26+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"Poetic jailbreaks succeed because they induce distinct attention patterns in LLMs that are independent of harmful-content detection, not because models fail to recognize literary formatting.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2605.07878","ref_index":20,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Black-box model classification under the discriminative factorization","primary_cat":"cs.LG","submitted_at":"2026-05-08T15:32:35+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Discriminative factorization distinguishes high-quality query sets for black-box model classification, with chance-level error decaying exponentially in query budget and parameters predicting empirical decay rates on auditing tasks.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"of the 39th International Conference on Machine Learning, volume 162 ofProceedings of Machine Learning Research, pages 9525-9587. PMLR, 2022. [19] Mandar Joshi, Eunsol Choi, Daniel S Weld, and Luke Zettlemoyer. TriviaQA: A large scale distantly supervised challenge dataset for reading comprehension. InProceedings of the 55th Annual Meeting of the Association for Computational Linguistics, pages 1601-1611, 2017. [20] Prakhar Kaushik, Shravan Chaudhari, Ankit Vaidya, Rama Chellappa, and Alan Yuille. The universal weight subspace hypothesis, 2025. URLhttps://arxiv.org/abs/2512.05117. [21] Simon Kornblith, Mohammad Norouzi, Honglak Lee, and Geoffrey Hinton. Similarity of neural network representations revisited. InProceedings of the 36th International Conference on"},{"citing_arxiv_id":"2605.06523","ref_index":17,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"On the Implicit Reward Overfitting and the Low-rank Dynamics in RLVR","primary_cat":"cs.LG","submitted_at":"2026-05-07T16:30:28+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":5.0,"formal_verification":"none","one_line_summary":"RLVR exhibits implicit reward overfitting to training data and optimizes heavy-tailed singular spectra with rank-1 focus on reasoning capability.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"Recently, a substantial body of researches [ 40] have observed that LLMs exhibit low-rank characteristics when adapting to downstream tasks, and RL appears to be no exception. A recent study [3] suggests that the reasoning improvements conferred by RL are primarily concentrated within the Rank-1 component of weight updates. Closely related work has boldly proposed the universal weight subspace hypothesis [ 17]. Furthermore, another disruptive study [39] recently found that even without RL training, the pass@k [5] metric of models continues to improve as k increases. This implies that RL may not be imparting new underlying reasoning logic ∗Corresponding Author Preprint. arXiv:2605.06523v1 [cs.LG] 7 May 2026 RL-trained Model Base Model"},{"citing_arxiv_id":"2604.24037","ref_index":3,"ref_count":2,"confidence":0.9,"is_internal_anchor":false,"paper_title":"A Limit Theory of Foundation Models: A Mathematical Approach to Understanding Emergent Intelligence and Scaling Laws","primary_cat":"cs.LG","submitted_at":"2026-04-27T04:43:42+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":3.0,"formal_verification":"none","one_line_summary":"Formalizes emergent intelligence in foundation models as the limit of E(N,P,K) as N,P,K approach infinity, proves existence conditions via nonlinear Lipschitz operators, and derives scaling laws from covering numbers.","context_count":1,"top_context_role":"background","top_context_polarity":"support","context_text":"(iii) the emergent intelligence is essentially determined by an infinite-dimensional system, yet can be effectively approximated in practice through a finite-dimensional architecture, offering a solid theoretical evidence for the practical applicability of foundation models in general, and the universal weight subspace hypothesis proposed by Kaushik et al. [ 3] in particular. We apply the established theories to the foundation models with GPT-2-like architectures, showing that the GPT-2-like foundation models do exhibit intelligent emergence, whereas GPT-1-like models do not, and for GPT-2 like models, their scaling law with respect to model size follows a power-law. We provide a series of experiments to support the rationality"},{"citing_arxiv_id":"2604.18487","ref_index":20,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"Adversarial Humanities Benchmark: Results on Stylistic Robustness in Frontier Model Safety","primary_cat":"cs.CL","submitted_at":"2026-04-20T16:37:27+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"Stylistic rewrites of harmful prompts raise attack success rates from 3.84% to 36.8-65% across 31 frontier models, indicating weak generalization in safety refusals.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.11947","ref_index":3,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"ResBM: Residual Bottleneck Models for Low-Bandwidth Pipeline Parallelism","primary_cat":"cs.LG","submitted_at":"2026-04-13T18:40:45+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"ResBM achieves 128x activation compression in pipeline-parallel transformer training by adding a residual bottleneck module that preserves a low-rank identity path, with no major loss in convergence or added overhead.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null},{"citing_arxiv_id":"2604.06377","ref_index":32,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"The Master Key Hypothesis: Unlocking Cross-Model Capability Transfer via Linear Subspace Alignment","primary_cat":"cs.LG","submitted_at":"2026-04-07T19:02:10+00:00","verdict":"UNVERDICTED","verdict_confidence":"LOW","novelty_score":6.0,"formal_verification":"none","one_line_summary":"The Master Key Hypothesis states that capabilities are low-dimensional directions transferable across models through linear subspace alignment, with UNLOCK demonstrating gains such as 12.1% accuracy improvement on MATH when transferring CoT from 14B to 7B models.","context_count":1,"top_context_role":"background","top_context_polarity":"background","context_text":"space in which this transformation is represented, and can be broadly categorized into three distinct frameworks - (1)Weight-space transfer[26]: The parameter-level difference between two Source models is added directly to the Target model, which typically requires architectural compatibility between the Source and Target models, or additional pruning or corrective training for cross-model alignment; (2)Output-space transfer[ 32]: The logit difference between two Source variants is applied at each generation step to adjust the output distribution of the Target model without mod- ifying its parameters. This avoids shape mismatches, but incurs substantial inference cost due to per-token logit computation for multiple models, and requires identical tokenization between the source and target; and (3)Latent-space transfer[42]: This strategy involves intervening on internal"},{"citing_arxiv_id":"2512.15742","ref_index":12,"ref_count":1,"confidence":0.9,"is_internal_anchor":false,"paper_title":"SHARe-KAN: Post-Training Vector Quantization for Cache-Resident KAN Inference","primary_cat":"cs.LG","submitted_at":"2025-12-10T04:10:13+00:00","verdict":"CONDITIONAL","verdict_confidence":"MODERATE","novelty_score":6.0,"formal_verification":"none","one_line_summary":"SHARe-KAN compresses KAN prediction-head storage by 9.3X via post-training vector quantization at a 2-point mAP cost on PASCAL VOC detection, with no retraining and good zero-shot transfer.","context_count":0,"top_context_role":null,"top_context_polarity":null,"context_text":null}],"limit":50,"offset":0}