AI Research Paper Archive & Publications
Browse through 321 peer-reviewed AI publications, arxiv preprints, technical abstracts, and key takeaway summaries covering LLM architectures, vision models, alignment, and multi-agent systems.
By Aho Yapi, Pierre Latouche, Arnaud Guillin et al.
Task-specific fine-tuning of French pre-trained language models achieves ~85.7% balanced accuracy in classifying accident-process roles across construction, metallurgy, and chemistry sectors, enabling cross-sector generalization without target-domain retraining.
By Hongyang Du, Lan Yan, Christian Flores et al.
A frozen frontier model drives 230+ design tools while an external natural-language procedural memory grows via widening/deepening + a matched replay gate, lifting execution success from 72.7% to 99.3% with no weight updates.
By Nitish Dashora, Douglas Chen, Idan Shenfeld et al.
Train-time VLM queries distill salient history into a lightweight workspace token, enabling robotic policies to solve memory-intensive tasks without in-the-loop VLM reasoning.
By Juri Opitz, Andrianos Michail
Text embedding models poorly encode physical measurements like mass and distance, and their similarity is dominated by superficial string patterns rather than objective magnitude.
By Bingxin Xu, Yuzhang Shang, Zhen Dong et al.
SafeHarness adds obstacle-aware route planning and contact execution to coding agents, forcing safety constraints into planning and boosting collision avoidance by 27% over prior SOTA.
By Kevin Qu, Tao Sun, Massimiliano Viola et al.
FAMOS uses a Multi-state Articulation Transformer to jointly process sparse, unordered point clouds, predicting movable-part segmentation and joint parameters, with a procedural generator to overcome data scarcity.
By Tapan Chugh, Vidushi Singh, Krish Jain et al.
Proposes a 'social harness' layer for inter-agent interactions to prevent failures and enable detection of invalid messages in agentic societies.
By Shuhan Xue, Jianyuan Zhong, Ziyuan Nan et al.
ScienceBuddy uses recursive-in-recursive self-improvement to jointly evolve AI harness and model via inner harness optimization and outer reinforcement learning, enabling continuously improving scientific agents.
By Atindra Jha, Margaret Li, Jure Leskovec et al.
Mixture-of-Experts models overfit more to repeated data than dense models, with degradation scaling with sparsity; masking-based regularization can mitigate but not eliminate the gap.
By William Zhou, Mayukha Siripuram, Xiao Yan et al.
Edge-deployable VLMs (2-8B) underperform a 300M specialist (BioCLIP) on species ID; domain gap (clean-to-field) is an image-quality property, not a general-purpose model flaw.
By Hongbo Chen, Li Charlie Xia
Proposes $\gamma^*$-concept shift using entropic optimal transport to unify covariate and concept shifts, with sample-based estimators and the DataShifts algorithm for rigorous error bound quantification under distribution shift.
By Boning Li, Longbo Huang
GPU-CFR compiles CFR game trees into static dataflow, precomputing indices and batching depth levels, then uses CUDA Graph Replay to cut kernel launches, achieving 80x speedup over prior GPU CFR.
Frontier Artificial Intelligence Research Disciplines
Our research archive systematically categorizes preprints and conference proceedings across core theoretical and applied machine learning domains:
Language Models & Scaling
Investigating transformer scaling laws, sparse Mixture-of-Experts (MoE), test-time compute scaling, reasoning token generation, and deep reasoning architectures.
Vision, Audio & Video
Advancements in unified multimodal tokenizers, autoregressive video synthesis, latent diffusion models, and real-time spatial scene understanding.
Alignment & Interpretability
Techniques in Direct Preference Optimization (DPO), Reinforcement Learning from AI Feedback (RLAIF), sparse autoencoders, and mechanistic circuit auditing.
Autonomous Agents & Tool Use
Multi-agent coordination protocols, external environment simulation, self-correcting code synthesis, and long-horizon planning frameworks.
Inference Optimization
Hardware-aware attention mechanisms, FlashAttention kernels, sub-4-bit quantization (AWQ/GPTQ), speculative decoding, and KV cache eviction policies.
Benchmark Integrity & Contamination
Methodologies addressing test-set leakage, dynamic synthetic benchmark generation, human-in-the-loop validation, and adversarial capability probes.
Frequently Asked Questions on AI Research
How frequently is the research paper archive updated?▼
Our indexing pipeline scans daily arXiv submissions in cs.CL, cs.AI, cs.CV, and cs.LG alongside major conference proceedings (NeurIPS, ICML, ICLR, ACL). Publications are updated continuously with fresh abstracts and takeaways.
Where can I access the original full-text PDF of a paper?▼
Every individual paper synopsis page on llmdb.app contains direct, verified links to the original arXiv abstract, PDF download mirror, and official GitHub source code repositories when published by the authors.