A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Mehrnaz Mofakhami, Ananya Sahu, Alejandro R. Salamanca, et al.
Tiny Aya L2-Thinker achieves over 93% in-language reasoning across 60 languages via data mixing in SFT, without reasoning supervision in every target language.
Hmrishav Bandyopadhyay, Xuanchi Ren, Zijian Huang, et al.
Introduces Context-Matched Distillation (CMD), a causal DMD framework that aligns teacher supervision with the information available during autoregressive video generation, improving control adherence and state-of-the-art performance.
Yijiang Li, Bingyang Wang, Yijun Liang, et al.
Proposes U-OPSD, an unsupervised on-policy self-distillation method that uses only a model's own generations and majority voting to improve reasoning without external supervision.
Changle Qu, Sunhao Dai, Hengyi Cai, et al.
TurnSight is a turn-level hindsight self-distillation framework for tool-integrated reasoning that derives supervision from execution-conditioned hindsight and cross-horizon agreement to improve credit assignment.
Unknown
This paper proposes Chain-of-Thought (CoT) supervision to improve long-context language models by training them to reason over dispersed information before generating answers.
Unknown
This paper explores how large language models can self-improve through iterative fine-tuning on their own outputs, reducing the need for extensive human supervision.
Unknown
This paper introduces unsupervised process reward models that generate step-level supervision for LLM reasoning without human labels, improving performance on math reasoning tasks.
Tej Deep Pala, Panshul Sharma, Amir Zadeh, et al.
This paper introduces error-aware hierarchical supervision for process reward models, improving reasoning step evaluation by incorporating error type information.
Unknown
An Image Transformer pre-trained via self-supervision on document images for improved representation learning.
Unknown
An encoder-decoder architecture that reconstructs input images by masking random patches and leveraging a high proportion of masking for self-supervision.
Unknown
ColBERTv2 improves late-interaction retrieval by combining aggressive residual compression with denoised supervision to enhance quality and reduce storage.
Unknown
Shows that language models can learn diverse language tasks without explicit supervision.