From $f(x)$ and $g(x)$ to $f(g(x))$: LLMs Learn New Skills in RL by Composing Old Ones

• URL: http://arxiv.org/abs/2509.25123v2
• Date: Mon, 13 Oct 2025 13:03:40 GMT
• Title: From $f(x)$ and $g(x)$ to $f(g(x))$: LLMs Learn New Skills in RL by Composing Old Ones
• Authors: Lifan Yuan, Weize Chen, Yuchen Zhang, Ganqu Cui, Hanbin Wang, Ziming You, Ning Ding, Zhiyuan Liu, Maosong Sun, Hao Peng,
• Abstract summary: We show that LLMs can acquire genuinely new skills during RL by composing existing ones.<n>Our experiments show that compositional skill acquired on a source task transfers to a different target task.<n>This transfer happens even without compositional training on the target, requiring only prior knowledge of the target's atomic skills.
• Score: 68.68686526804909
• License: http://creativecommons.org/licenses/by-nc-sa/4.0/
• Abstract: Does RL teach LLMs genuinely new skills, or does it merely activate existing ones? This question lies at the core of ongoing debates about the role of RL in LLM post-training. On one side, strong empirical results can be achieved with RL even without preceding supervised finetuning; on the other, critics argue that RL contributes little beyond reweighting existing reasoning strategies. This work provides concrete evidence that LLMs can acquire genuinely new skills during RL by composing existing ones, mirroring one of the central mechanisms by which humans acquire new cognitive skills. To mitigate data contamination and other confounding factors, and to allow precise control over task complexity, we develop a synthetic framework for our investigation. Specifically, we define a skill as the ability to infer the output of a string transformation function f(x) given x. When an LLM has already learned f and g prior to RL, our experiments reveal that RL enables it to learn unseen compositions of them h(x)=g(f(x)). Further, this compositional ability generalizes to more difficult problems such as compositions of >2 functions unseen during RL training. Surprisingly, our experiments show that compositional skill acquired on a source task transfers to a different target task. This transfer happens even without compositional training on the target, requiring only prior knowledge of the target's atomic skills. Our qualitative analysis shows that RL fundamentally changes the reasoning behaviors of the models. In contrast, next-token training with the same data yields none of these findings. Our systematic experiments provide fresh insights into LLM learning, suggesting the value of first building base models with basic skills, then using RL to incentivize advanced, generalizable skills for complex problems.

Related papers

• DELTA-Code: How Does RL Unlock and Transfer New Programming Algorithms in LLMs? [92.4931695205957]
  We introduce a benchmark of synthetic coding problem families designed to probe two fundamental aspects: learnability and transferrability.<n>Our experiments reveal a striking grokking phase transition: after an extended period with near-zero reward, RL-trained models abruptly climb to near-perfect accuracy.<n>Results show solid gains within families and for recomposed skills, but persistent weaknesses in transformative cases.
  arXiv  Detail & Related papers  (2025-09-25T11:20:56Z)
• Curriculum Reinforcement Learning from Easy to Hard Tasks Improves LLM Reasoning [52.32193550674408]
  We aim to improve the reasoning capabilities of language models via reinforcement learning (RL)<n>We propose to schedule tasks from easy to hard (E2H), allowing LLMs to build reasoning skills gradually.<n>E2H Reasoner significantly improves the reasoning ability of small LLMs (1.5B to 3B)
  arXiv  Detail & Related papers  (2025-06-07T02:41:54Z)
• Beyond Accuracy: Dissecting Mathematical Reasoning for LLMs Under Reinforcement Learning [93.00629872970364]
  Reinforcement learning (RL) has become the dominant paradigm for improving the performance of language models on complex reasoning tasks.<n>We introduce SPARKLE, a fine-grained analytic framework to dissect the effects of RL across three key dimensions.<n>We study whether difficult problems -- those yielding no RL signals and mixed-quality reasoning traces -- can still be effectively used for training.
  arXiv  Detail & Related papers  (2025-06-05T07:53:59Z)
• Decomposing Elements of Problem Solving: What "Math" Does RL Teach? [22.517954679764244]
  We decompose problem solving into fundamental capabilities: Plan, Execute, and Verify.<n>We show that RL-trained models struggle with fundamentally new problems, hitting a 'coverage wall' due to insufficient planning skills.<n>Our findings provide insights into the role of RL in enhancing LLM reasoning, expose key limitations, and suggest a path toward overcoming these barriers.
  arXiv  Detail & Related papers  (2025-05-28T18:18:49Z)
• R1-Searcher: Incentivizing the Search Capability in LLMs via Reinforcement Learning [87.30285670315334]
  textbfR1-Searcher is a novel two-stage outcome-based RL approach designed to enhance the search capabilities of Large Language Models.<n>Our framework relies exclusively on RL, without requiring process rewards or distillation for a cold start.<n>Our experiments demonstrate that our method significantly outperforms previous strong RAG methods, even when compared to the closed-source GPT-4o-mini.
  arXiv  Detail & Related papers  (2025-03-07T17:14:44Z)
• On the Emergence of Thinking in LLMs I: Searching for the Right Intuition [34.32871896067864]
  We propose a post-training framework called Reinforcement Learning via Self-Play (RLSP)<n> RLSP involves three steps: supervised fine-tuning with human or synthetic demonstrations of the reasoning process, using an exploration reward signal to encourage diverse and efficient reasoning behaviors, and RL training with an outcome verifier to ensure correctness while preventing reward hacking.<n> Empirical studies in the math domain show that RLSP improves reasoning.
  arXiv  Detail & Related papers  (2025-02-10T18:52:04Z)
• Teaching Large Language Models to Reason with Reinforcement Learning [38.17625148525193]
  Reinforcement Learning from Human Feedback (textbfRLHF) has emerged as a dominant approach for aligning LLM outputs with human preferences. Inspired by the success of RLHF, we study the performance of multiple algorithms that learn from feedback.
  arXiv  Detail & Related papers  (2024-03-07T16:36:29Z)
• RvS: What is Essential for Offline RL via Supervised Learning? [77.91045677562802]
  Recent work has shown that supervised learning alone, without temporal difference (TD) learning, can be remarkably effective for offline RL. In every environment suite we consider simply maximizing likelihood with two-layer feedforward is competitive. They also probe the limits of existing RvS methods, which are comparatively weak on random data.
  arXiv  Detail & Related papers  (2021-12-20T18:55:16Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.