Papers20

#RLVR

LSRIF: Logic-Structured Reinforcement Learning for Instruction Following

Real instructions often have logic like and first-then and if-else and this paper teaches models to notice and obey that logic.

#instruction following#logical structures#parallel constraints

RelayLLM: Efficient Reasoning via Collaborative Decoding

Intermediate

Chengsong Huang, Tong Zheng et al.Jan 8arXiv

RelayLLM lets a small model do the talking and only asks a big model for help on a few, truly hard tokens.

#token-level collaboration#<call>n</call> command#collaborative decoding

Evaluating Parameter Efficient Methods for RLVR

Intermediate

Qingyu Yin, Yulun Wu et al.Dec 29arXiv

The paper asks which small, add-on training tricks (PEFT) work best when we teach language models with yes/no rewards we can check (RLVR).

#RLVR#parameter-efficient fine-tuning#LoRA

Exploration vs Exploitation: Rethinking RLVR through Clipping, Entropy, and Spurious Reward

Intermediate

Peter Chen, Xiaopeng Li et al.Dec 18arXiv

The paper studies why two opposite-sounding tricks in RL for reasoning—adding random (spurious) rewards and reducing randomness (entropy)—can both seem to help large language models think better.

#RLVR#Group Relative Policy Optimization#ratio clipping

Puzzle Curriculum GRPO for Vision-Centric Reasoning

Intermediate

Ahmadreza Jeddi, Hakki Can Karaimer et al.Dec 16arXiv

This paper teaches vision-language models to reason about pictures using puzzles instead of expensive human labels.

#vision-language models#reinforcement learning#group-relative policy optimization

GTR-Turbo: Merged Checkpoint is Secretly a Free Teacher for Agentic VLM Training

Intermediate

Tong Wei, Yijun Yang et al.Dec 15arXiv

GTR-Turbo teaches a vision-language agent using a 'free teacher' made by merging its own past checkpoints, so no costly external model is needed.

#GTR-Turbo#checkpoint merging#TIES-merging

Long-horizon Reasoning Agent for Olympiad-Level Mathematical Problem Solving

Intermediate

Songyang Gao, Yuzhe Gu et al.Dec 11arXiv

This paper builds a math problem–solving agent, Intern-S1-MO, that thinks in multiple rounds and remembers proven mini-results called lemmas so it can solve very long, Olympiad-level problems.

#long-horizon reasoning#lemma-based memory#multi-agent reasoning

SPARK: Stepwise Process-Aware Rewards for Reference-Free Reinforcement Learning

Intermediate

Salman Rahman, Sruthi Gorantla et al.Dec 2arXiv

SPARK teaches AI to grade its own steps without needing the right answers written down anywhere.

#SPARK#Process Reward Model#PRM-CoT

1 2