Cover Image

How Does RL Post-training Induce Skill Composition? A Case Study on Countdown

Fuente: arXiv

Saved in:

Bibliographic Details
Main Authors:	Park, Simon, Kaur, Simran, Arora, Sanjeev
Format:	Preprint
Published:	2025
Subjects:	Machine Learning
Online Access:	Acceder al recurso
Tags:	Add Tag No Tags, Be the first to tag this record!

Similar Items

Instruct-SkillMix: A Powerful Pipeline for LLM Instruction Tuning
by: Kaur, Simran, et al.
Published: (2024)

Can Models Learn Skill Composition from Examples?
by: Zhao, Haoyu, et al.
Published: (2024)

Skill-Targeted Adaptive Training
by: He, Yinghui, et al.
Published: (2025)

Prioritized Replay for RL Post-training
by: Fatemi, Mehdi
Published: (2026)

Countdown-Code: A Testbed for Studying The Emergence and Generalization of Reward Hacking in RLVR
by: Khalifa, Muhammad, et al.
Published: (2026)

RL Grokking Recipe: How Does RL Unlock and Transfer New Algorithms in LLMs?
by: Sun, Yiyou, et al.
Published: (2025)

Contextual Drag: How Errors in the Context Affect LLM Reasoning
by: Cheng, Yun, et al.
Published: (2026)

On the Power of Context-Enhanced Learning in LLMs
by: Zhu, Xingyu, et al.
Published: (2025)

Optimistic Verifiable Training by Controlling Hardware Nondeterminism
by: Srivastava, Megha, et al.
Published: (2024)

SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training
by: Chu, Tianzhe, et al.
Published: (2025)

When and How Does CLIP Enable Domain and Compositional Generalization?
by: Kempf, Elias, et al.
Published: (2025)

Echo Chamber: RL Post-training Amplifies Behaviors Learned in Pretraining
by: Zhao, Rosie, et al.
Published: (2025)

VAM: Verbalized Action Masking for Controllable Exploration in RL Post-Training -- A Chess Case Study
by: Zhang, Zhicheng, et al.
Published: (2026)

CRePE: Convolution-aware Relative Importance in Post-training Pruning with Efficient Search
by: Park, Cheonjun
Published: (2026)

DUMP: Automated Distribution-Level Curriculum Learning for RL-based LLM Post-training
by: Wang, Zhenting, et al.
Published: (2025)

The unregulated plant‐based ‘milk’ industry: A threat to nutrition, health and safety?
by: Simran Kaur Arora
Published: (2024)

On the Impossibility of Retrain Equivalence in Machine Unlearning
by: Yu, Jiatong, et al.
Published: (2025)

On the SDEs and Scaling Rules for Adaptive Gradient Algorithms
by: Malladi, Sadhika, et al.
Published: (2022)

Generalizing from SIMPLE to HARD Visual Reasoning: Can We Mitigate Modality Imbalance in VLMs?
by: Park, Simon, et al.
Published: (2025)

SkillRL: Evolving Agents via Recursive Skill-Augmented Reinforcement Learning
by: Xia, Peng, et al.
Published: (2026)

JigsawRL: Assembling RL Pipelines for Efficient LLM Post-Training
by: Hu, Zhengding, et al.
Published: (2026)

Skill Reuse as Compression in Agentic RL
by: Xu, Zhikun, et al.
Published: (2026)

SCALAR: Learning and Composing Skills through LLM Guided Symbolic Planning and Deep RL Grounding
by: Zabounidis, Renos, et al.
Published: (2026)

A Quadratic Synchronization Rule for Distributed Deep Learning
by: Gu, Xinran, et al.
Published: (2023)

SCOPE-RL: Stable and Quantitative Control of Policy Entropy in RL Post-Training
by: Wang, Chen, et al.
Published: (2025)

ReSkill: Reconciling Skill Creation with Policy Optimization in Agentic RL
by: He, Zelin, et al.
Published: (2026)

ASAP: Unsupervised Post-training with Label Distribution Shift Adaptive Learning Rate
by: Park, Heewon, et al.
Published: (2025)

LLMStinger: Jailbreaking LLMs using RL fine-tuned LLMs
by: Jha, Piyush, et al.
Published: (2024)

SPEED-RL: Faster Training of Reasoning Models via Online Curriculum Learning
by: Zhang, Ruiqi, et al.
Published: (2025)

How Does Critical Batch Size Scale in Pre-training?
by: Zhang, Hanlin, et al.
Published: (2024)

Mirage or Method? How Model-Task Alignment Induces Divergent RL Conclusions
by: Wu, Haoze, et al.
Published: (2025)

Metacognitive Reuse: Turning Recurring LLM Reasoning Into Concise Behaviors
by: Didolkar, Aniket, et al.
Published: (2025)

Trainable Transformer in Transformer
by: Panigrahi, Abhishek, et al.
Published: (2023)

Provable unlearning in topic modeling and downstream tasks
by: Wei, Stanley, et al.
Published: (2024)

Why Does RL Generalize Better Than SFT? A Data-Centric Perspective on VLM Post-Training
by: Lu, Aojun, et al.
Published: (2026)

Unrealized Expectations: Comparing AI Methods vs Classical Algorithms for Maximum Independent Set
by: Wu, Yikai, et al.
Published: (2025)

Learning Dynamics in RL Post-Training for Language Models
by: Tomihari, Akiyoshi
Published: (2026)

RL in Name Only? Analyzing the Structural Assumptions in RL post-training for LLMs
by: Samineni, Soumya Rani, et al.
Published: (2025)

Decomposing Elements of Problem Solving: What "Math" Does RL Teach?
by: Qin, Tian, et al.
Published: (2025)

When Errors Can Be Beneficial: A Categorization of Imperfect Rewards for Policy Gradient
by: Shang, Shuning, et al.
Published: (2026)