A comprehensive index of artificial intelligence and machine-learning research with AI-generated summaries, citation metrics, and direct links to papers and code.
Unknown
Proposes native reasoning models that train language models to reason on unverifiable data, moving beyond SFT and RLVR.
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.
Juan S. Santillana
VectraYX-Vision-1B is a sub-2B Spanish/LATAM cybersecurity VLM with structured reasoning and tool use, but current vision SFT yields near-zero visual grounding.
Desiree Cho, Cameron Tice, Bernie Hogan, et al.
Constitutional midtraining, inserting values-based content during pretraining, yields durable alignment gains at 120B scale, blunting SFT-induced blackmail propensity.
Siyan Zhao, Devaansh Gupta, Qinqing Zheng, et al.
Proposes d1, a framework adapting pre-trained masked diffusion LLMs into reasoning models via masked SFT and a novel critic-free RL algorithm diffu-GRPO.
Hongyi Jin, Wenhan Yang, Meysam Ghaffari, et al.
Shows that diverse and challenging reasoning examples for SFT can be identified using only initial reasoning tokens, enabling efficient data curation without strong reasoning models.
Juncheng Wu, Sheng Liu, Haoqin Tu, et al.
This paper introduces a fine-grained evaluation framework decomposing LLM reasoning into knowledge correctness and reasoning quality, revealing that SFT improves accuracy but degrades reasoning, while RL enhances medical reasoning by pruning inaccura
Unknown
LearnLM enhances Gemini for pedagogical instruction by co-training SFT and RLHF, outperforming leading LLMs in tutoring scenarios.
Unknown
A pipeline that generates system messages and aligned assistant responses for SFT datasets lacking them, using LLM-as-a-judge verification.
Unknown
RL improves generalization in text and visual tasks while SFT primarily memorizes training data, with inference-time compute scaling further enhancing RL gains.
Zihan Liu, Zhuoling Yang, Yang Chen, et al.
AceReason-Nemotron 1.1 scales SFT data and uses stage-wise RL with tuned temperatures to achieve state-of-the-art reasoning in 7B models.
Unknown
A 14B parameter reasoning language model derived from Deepseek-R1-Distilled-Qwen-14B, trained with SFT and model merging, fully open-sourced.