Papers10

#Qwen3

Learn Hard Problems During RL with Reference Guided Fine-tuning

ReGFT is a simple pre-RL step that shows the model partial human hints, then makes it solve problems in its own words, creating correct, model-style solutions for hard questions.

#Reference-Guided Fine-Tuning#ReGFT#ReFT

DataChef: Cooking Up Optimal Data Recipes for LLM Adaptation via Reinforcement Learning

Intermediate

Yicheng Chen, Zerun Ma et al.Feb 11arXiv

DataChef teaches a large language model to be a smart data chef: it plans and codes full data pipelines that turn messy datasets into great training meals for other models.

#data recipe#data processing pipeline#reinforcement learning

DFlash: Block Diffusion for Flash Speculative Decoding

Intermediate

Jian Chen, Yesheng Liang et al.Feb 5arXiv

DFlash is a new way to make big language models answer much faster without changing the final answers.

#DFlash#speculative decoding#diffusion language model

THINKSAFE: Self-Generated Safety Alignment for Reasoning Models

Intermediate

Seanie Lee, Sangwoo Park et al.Jan 30arXiv

Large reasoning models got very good at thinking step-by-step, but that sometimes made them too eager to follow harmful instructions.

#THINKSAFE#self-generated safety alignment#refusal steering

Fast KVzip: Efficient and Accurate LLM Inference with Gated KV Eviction

Intermediate

Jang-Hyun Kim, Dongyoon Han et al.Jan 25arXiv

Fast KVzip is a new way to shrink an LLM’s memory (the KV cache) while keeping answers just as accurate.

#KV cache compression#gated KV eviction#sink attention

Mecellem Models: Turkish Models Trained from Scratch and Continually Pre-trained for the Legal Domain

Intermediate

Özgür Uğur, Mahmut Göksu et al.Jan 22arXiv

The paper builds special Turkish legal AI models called Mecellem by teaching them from the ground up and then giving them more law-focused lessons.

#Turkish legal NLP#ModernBERT#Continual pre-training

Robust Tool Use via Fission-GRPO: Learning to Recover from Execution Errors

Beginner

Zhiwei Zhang, Fei Zhao et al.Jan 22arXiv

Small AI models often stumble when a tool call fails and then get stuck repeating bad calls instead of fixing the mistake.

#FISSION-GRPO#error recovery#tool use

GlimpRouter: Efficient Collaborative Inference by Glimpsing One Token of Thoughts

Beginner

Wenhao Zeng, Xuteng Zhang et al.Jan 8arXiv

Big reasoning AIs think in many steps, which is slow and costly.

#collaborative inference#initial token entropy#step-level routing

Nemotron-Math: Efficient Long-Context Distillation of Mathematical Reasoning from Multi-Mode Supervision

Intermediate

Wei Du, Shubham Toshniwal et al.Dec 17arXiv

Nemotron-Math is a giant math dataset with 7.5 million step-by-step solutions created in three thinking styles and with or without Python help.

#mathematical reasoning#long-context fine-tuning#multi-mode supervision

Efficient-DLM: From Autoregressive to Diffusion Language Models, and Beyond in Speed

Intermediate

Yonggan Fu, Lexington Whalen et al.Dec 16arXiv

Autoregressive (AR) models write one word at a time, which is accurate but slow, especially when your computer or GPU can’t keep many tasks in memory at once.

#diffusion language models#autoregressive models#AR-to-dLM conversion