Expander SAEs apply left-d-regular expander masks to TopK SAEs, learning only dn decoder parameters instead of mn and tracing a storage-fidelity frontier that reaches 293x compression with 84% retained performance on Qwen2.5-3B.
Canonical reference
A graph-based recovery and decomposition of Swanson’ s hypothesis using semantic predications
Canonical reference. 80% of citing Pith papers cite this work as background.
1095
Pith papers citing it
1
external citations · Crossref
Background
80% of classified citations
citation-role summary
background 254
method 24
baseline 6
dataset 5
extension 2
citation-polarity summary
authors
co-cited works
representative citing papers
The normalized orbit of a bounded normal operator can be a frame, providing a counterexample to Conjecture 3.
Introduces the largest freely available Italian clinical notes corpus with 4M notes and expert-annotated subset for a new CRF-filling benchmark.
Machine learning methods discover a new noncrossing-partition statistic interpreting q,t-Narayana polynomials and yield a combinatorial proof of their symmetry.
SLayerGen generates crystals invariant to any space or layer group via autoregressive lattice and Wyckoff sampling plus equivariant diffusion, achieving gains over bulk models on diperiodic materials after correcting a prior loss inconsistency for hexagonal groups.
Every proper minor-closed graph class admits an optimal (1+o(1)) log n bit adjacency labeling scheme.
A directed weighted two-graph model separates feasibility from movement in solution discovery and yields a detailed complexity classification for path and shortest-path discovery.
The method reformulates ALE mesh motion as independent multi-patch spline parameterizations per time step, using barrier functions, tangential-slip reparameterization, and constant-preserving quasi-interpolation to enable large-rotation FSI simulations.
Superconductivity in high-pressure MnB4 is induced by altermagnetic spin fluctuations, yielding extended-s pairing symmetry.
A new qubit-efficient HUBO encoding for graph partitioning problems like minimum coloring uses logarithmic bits and a lexicographic penalty to cut resources while providing provable optimality conditions.
A survey of 172 open educational datasets from 204 papers across LAK, EDM, and AIED conferences reveals trends, 143 previously uncatalogued datasets, field gaps, and an 8-item PRACTICE checklist for better data publication.
A microlocal lift of Navier-Stokes dynamics on manifolds yields an if-and-only-if geometric criterion for solution blow-up in terms of deformation integrability, directional entropy, and lifted energy.
A 9U CubeSat detector can identify a thermonuclear weapon on a satellite from 4 km away by observing spallation neutrons induced by GeV protons in roughly one week.
O(n log n) algorithm and matching Omega(n log n) lower bound for partitioning a simple polygon's boundary into the minimum number of contiguous visible segments.
Introduces a method to design structure-specific relational inductive biases for a base transformer architecture, enabling end-to-end transcription of documents with intrinsic structures, demonstrated on sheet music, shape drawings, and mechanical engineering drawings.
The paper introduces a probabilistic sign rule for quotients of positive series and integral transforms that reduces monotonicity, log-supermodularity, and log-convexity to kernel sign criteria via moment identities, and applies it to derive new inequalities for hypergeometric, Stieltjes, and Prabha
SPoILeR uses multimodal pre-training to enable accurate novel view synthesis of infrared, polarimetric, and multispectral data from RGB-supervised fine-tuning on new scenes.
NEvo performs evolutionary search guided by a dynamic voxel-level encoding model to synthesize videos that maximize predicted activity in target brain ROIs, recovering known selectivities and revealing temporal dynamics differences.
Every n-vertex H-minor-free graph admits a 3-coloring with monochromatic components of size O_H(n^{4/9}).
The Spin-MInt algorithm is proven symplectic for general K electronic states via explicit verification of the condition MJM^T = J on the coadjoint orbit of the su(K) Lie-Poisson algebra.
The authors synthesize a typology of fourteen OSS sub-genres from a review of 3,925 papers and present a research agenda on cross-sub-genre generalization.
Structural identifiability analysis shows point sources restore identifiability for inferring spatial stochastic dynamics parameters from static snapshots, unlike distributed sources, with limits depending on modeling choices.
Maximal quantum leakage upper-bounds quantum inference accuracy; optimal encodings are pure states, with tight frames and equiangular tight frames optimal when system dimension is small.
In multistage SI(k)R models, the relationship between prevalence peak and weighted stage functional maxima varies with scaling of progression rates, converging under Erlang scaling to a delay model that justifies the factor-two approximation with error bounds and corrections.
citing papers explorer
-
eCream-MedCorpus A Large-Scale Corpus of Clinical Notes for Italian
Introduces the largest freely available Italian clinical notes corpus with 4M notes and expert-annotated subset for a new CRF-filling benchmark.
-
PolySpeech-100: A Large-Scale Benchmark for Speech Understanding Across 100+ Languages and Dialects
PolySpeech-100 is a new benchmark for native-level speech comprehension across 110 linguistic variants that evaluates 22 models and reports E2E advantages on dialects, robustness gaps on low-resource languages, and degradation from Chain-of-Thought prompting.
-
Can LLMs Use Linguistic Uncertainty Markers to Reliably Reflect Intrinsic Confidence?
LLMs struggle to associate epistemic markers with stable internal confidence levels across distributions, even under model-centric interpretations, while maintaining somewhat consistent marker rankings.
-
Ishigaki-IDS-Bench: A Benchmark for Generating Information Delivery Specification from BIM Information Requirements
Presents Ishigaki-IDS-Bench, the first benchmark for LLM-based IDS generation from BIM requirements, with baseline results showing max 65.6% Facet F1 and 33.1% content pass rate across 10 models.
-
Reinforcement Learning with Metacognitive Feedback Elicits Faithful Uncertainty Expression in LLMs
RLMF uses quality of model self-judgments to refine RL rankings and select training data, achieving SOTA faithful calibration while preserving accuracy and outperforming standard RL by up to 63%.
-
Redact or Keep? A Fully Local AI Cascade for Educational Dialogue De-Identification
A local cascade framework for educational dialogue de-identification reaches 0.958 macro F1 on math tutoring transcripts, outperforming same-family LLM-only and commercial baselines while remaining fully on-device.
-
The Dynamics of Human and AI-Generated Language: How Semantics Fluctuates across Different Timescales
Develops ACW-based semantic timescale features showing longer autocorrelation windows associate with generic vocabulary and shorter ones with specific words in both human and LLM speech, with the pattern abolished by randomizing word order and timing.
-
Keyphrase Generative Representation of Youth Crisis Conversations Beyond Static Taxonomies
KGR, a constrained LLM for generating keyphrases from crisis SMS, expands a static taxonomy and raises topic-retrieval accuracy from 0.25 to 0.70 while surfacing new themes like immigration problems.
-
AstroMind: A High-Fidelity Benchmark for Spacecraft Behavior Reasoning Based on Large Language Models
AstroMind is a new physics-grounded benchmark for LLM reasoning on spacecraft behavior across intent inference, maneuver estimation, and threat assessment, evaluated on several open-weight models.
-
SymbolicLight V1: Spike-Gated Dual-Path Language Modeling with High Activation Sparsity and Sub-Billion-Scale Pre-Training Evidence
A 194M-parameter spiking dual-path model trained on 3B Chinese-English tokens achieves held-out PPL 8.88-8.93 at >89% per-element sparsity, trailing GPT-2 201M by 7.7% while showing that LIF temporal integration outperforms simple top-k masking at matched sparsity.
-
Exploiting Pre-trained Encoder-Decoder Transformers for Sequence-to-Sequence Constituent Parsing
Pre-trained encoder-decoder transformers fine-tuned for sequence-to-sequence constituent parsing outperform prior seq2seq models and compete with specialized parsers on continuous treebanks.
-
IfcLLM: Natural Language Querying of IFC Models through Complementary Relational and Graph Representations
IfcLLM combines relational and graph representations of IFC models with an LLM agent to achieve 93.3-100% first-attempt accuracy on natural language queries across three models and 30 scenarios.
-
Meow-Omni 1: A Multimodal Large Language Model for Feline Ethology
Meow-Omni 1 is a quad-modal MLLM that fuses video, audio, physiological time-series, and text to achieve 71.16% accuracy on feline intent recognition in the new MeowBench benchmark.
-
Learning Evidence of Depression Symptoms via Prompt Induction
Symptom Induction compresses labeled data into interpretable guidelines that improve LLM classification of depression symptoms in text, outperforming zero-shot, in-context, and fine-tuning approaches with gains on rare symptoms and cross-disease generalization.
-
CoGate-LSTM: Prototype-Guided Feature-Space Gating for Mitigating Gradient Dilution in Imbalanced Toxic Comment Classification
CoGate-LSTM adds prototype-guided cosine feature-space gating to a character-level BiLSTM with multi-source embeddings and focal loss, reaching 0.881 macro-F1 on Jigsaw toxic comments while using 7.3M parameters and outperforming fine-tuned BERT by 6.9 points on minority labels.
-
Privacy-Preserving RAG via Multi-Agent Semantic Rewriting: Achieving Confidentiality Without Compromising Contextual Fidelity
A multi-agent semantic rewriting system for RAG cuts targeted privacy leakage from 144 to 1 instances on LLaMA-3-8B while raising BLEU-1 to 0.122 over SAGE's 0.117, with offline preprocessing.
-
Automatic Reflection Level Classification in Hungarian Student Essays
Classical machine learning models outperform Hungarian transformers slightly in overall performance (71% vs 68% average score) for classifying reflection levels in student essays, though transformers handle rare classes better.
-
Mapping how LLMs debate societal issues when shadowing human personality traits, sociodemographics and social media behavior
CDS is a new synthetic corpus of LLM-generated texts on vaccines, disinformation, gender gaps, and STEM stereotypes, linked to persona attributes to enable bias and alignment audits.
-
Domain Fine-Tuning FinBERT on Finnish Histopathological Reports: Train-Time Signals and Downstream Correlations
Fine-tuning FinBERT on Finnish medical text produces embedding geometry shifts whose correlation with downstream performance the authors attempt to measure as a potential early signal for domain adaptation benefit.
-
LITcoder: A General-Purpose Library for Building and Comparing Encoding Models
LITcoder introduces a modular open-source library for constructing, benchmarking, and comparing neural encoding models that map continuous stimuli such as stories to fMRI brain data.
-
The Master-Slave Encoder Model for Improving Patent Text Summarization: A New Approach to Combining Specifications and Claims
MSEA uses a master-slave encoder architecture on patent specifications and claims, enhanced with pointer networks and repetition suppression, to generate better summaries as measured by small ROUGE score gains.
-
Automatically Learning Construction Injury Precursors from Text
Standard NLP classifiers can surface valid injury precursors from raw construction safety reports.
-
A Synthetic Reliability-Aware PINN Benchmark for Offshore Wind Turbine Support-Structure Monitoring with Bayesian Inverse Identification
Digi Turbine is a synthetic PINN benchmark integrating Euler-Bernoulli beam theory with Winkler foundation, Bayesian inverse identification, and FORM screening for OWT monopile monitoring, validated on synthetic data with analytical ground truth.
-
Peacemaker at ATE-IT: Automatic term extraction from Italian text for waste management data using encoder model
An encoder-based model is fine-tuned for automatic term extraction on Italian waste management texts and reports balanced type-level and micro-level F1 scores in a shared task.
-
Finding New Connections between Concepts from Medline Database Incorporating Domain Knowledge
An extended Swanson ABC model using MetaMap and TF-IDF on MEDLINE abstracts recovers known intermediate linking concepts between medical topic pairs and finds some new ones.
- DialToM: A Theory of Mind Benchmark for Forecasting State-Driven Dialogue Trajectories