Academic Publications
Academic Publications (Papers, formal publications and preprints) holds 101 entries from 9 sources. Larger sources have their own directory pages, 15 entries per page. Authority tiers in this category: A2–R.
Highlights by source
- arXiv (14 entries): BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding; Language Models are Few-Shot Learners; Training Compute-Optimal Large Language Models and more
- OpenReview (5 entries): Never Train from Scratch: Fair Comparison of Long-Sequence Models Requires Data-Driven Priors; Vision Transformers Need Registers; Learning Interactive Real-World Simulators and more
- ACL Anthology (52 entries): Reading Subtext: Evaluating Large Language Models on Short Story Summarization with Writers; HALoGEN: Fantastic LLM Hallucinations and Where to Find Them; HateDay: Insights from a Global Hate Speech Dataset Representative of a Day on Twitter and more
- PMLR (14 entries): Information Complexity of Stochastic Convex Optimization: Applications to Generalization, Memorization, and Tracing; Genie: Generative Interactive Environments; Stealing part of a production language model and more
- IEEE Xplore (4 entries): NoMaD: Goal Masked Diffusion Policies for Navigation and Exploration; Open X-Embodiment: Robotic Learning Datasets and RT-X Models; Marginalizing and Conditioning Gaussians onto Linear Approximations of Smooth Manifolds with Applications in Robotics and more
- ACM Digital Library (9 entries): WARP: An Efficient Engine for Multi-Vector Retrieval; A Framework for Auditing Chatbots for Dialect-Based Quality-of-Service Harms; External Evaluation of Discrimination Mitigation Efforts in Meta's Ad Delivery and more
NeurIPS Proceedings
Official NeurIPS proceedings (1 entry)
Stochastic Taylor Derivative Estimator: Efficient amortization for arbitrary differential operators
Uses a stochastic Taylor derivative estimator to efficiently contract arbitrary high-order derivative tensors of multivariate functions, making supervision with high-dimensional, high-order differential operators more feasible.
Evidence use: For checking STDE, the computational method for supervising high-order differential operators in PINNs, and the efficiency conclusions as scoped by the paper and the committee.
Limitations: The strongest evidence comes from PINNs and differential-operator supervision; it should not be described as a general speedup for all neural network training or backpropagation. The reported speed and memory gains depend on the specific operators, implementation, hardware and baselines; this site has not reproduced them independently.
Related concepts: Backpropagation · Supervised Learning · LLM Inference Optimization
Link last checked: 2026-09-24
CVF Open Access
CVPR, ICCV and WACV proceedings (1 entry)
VGGT: Visual Geometry Grounded Transformer
Evidence grade: for discovery only; must not by itself support a formal conclusion.
The paper proposes a feed-forward network that directly predicts camera parameters, depth, point maps and 3D point tracks from one, a few or many images of a scene. Its CVPR 2025 Best Paper award is verified, but there is currently no external expert statement showing a major contribution to AI, so it is listed only as a public entry awaiting further evidence.
Evidence use: For checking the paper's identity, authors, CVPR 2025 Best Paper result and original content; it cannot currently serve as evidence of a major AI contribution confirmed by this site.
Limitations: No external statement of a major AI contribution; being listed publicly records a gap in the review and does not mean the six-step importance review was passed. The official award shows recognition under that award process; until external expert commentary is added, it cannot be claimed that the paper significantly changed AI. Performance, speed and scope in the paper are as reported by the authors; this site has not reproduced the experiments independently.
Related concepts: World Models and 3D Generation · Multimodal Models · Transformer
Link last checked: 2026-09-24
Springer Nature
Nature journals and Springer scholarly publications (1 entry)
Minimalist Vision with Freeform Pixels
The paper proposes a minimalist vision system: the camera hardware is modeled as the first layer of a neural network, and task training jointly determines a small number of free-form physical pixels and the downstream inference network, performing vision tasks at a very small sensing scale while preserving privacy and enabling self-powered operation.
Evidence use: For understanding how learned sensing, computational imaging, privacy-preserving vision and edge inference are co-designed across hardware and models.
Limitations: The Best Paper award shows an outstanding contribution under the ECCV 2024 award process; it does not mean all of the paper's conclusions are permanently correct or apply to every vision task. This site verified the official award, the award citation, the Springer DOI and the public paper materials; experiments were not independently reproduced. The Springer page may offer only an abstract preview; this site shows metadata, review conclusions and the original link only, and does not copy restricted full text.
Related concepts: Neural Network · Convolutional Neural Network (CNN)
Link last checked: 2026-09-24