Xuqin Zhang

Xuqin Zhang, Quan He, Zhenrui Zheng et al.

Reinforcement learning (RL) has emerged as a dominant paradigm for eliciting long-horizon reasoning in Large Language Models (LLMs). However, scaling Tool-Integrated Reasoning (TIR) via RL remains challenging due to interaction collapse: a pathological state where models fail to sustain multi-turn tool usage, instead degenerating into heavy internal reasoning with only trivial, post-hoc code verification. We systematically study three questions: (i) how cold-start SFT induces an agentic, tool-using behavioral prior, (ii) how the interaction density of cold-start trajectories shapes exploration and downstream RL outcomes, and (iii) how the RL interaction budget affects learning dynamics and generalization under varying inference-time budgets. We then introduce ASTER (Agentic Scaling with Tool-integrated Extended Reasoning), a framework that circumvents this collapse through a targeted cold-start strategy prioritizing interaction-dense trajectories. We find that a small expert cold-start set of just 4K interaction-dense trajectories yields the strongest downstream performance, establishing a robust prior that enables superior exploration during extended RL training. Extensive evaluations demonstrate that ASTER-4B achieves state-of-the-art results on competitive mathematical benchmarks, reaching 90.0% on AIME 2025, surpassing leading frontier open-source models, including DeepSeek-V3.2-Exp.

Yi-Chen Li, Tian Xu, Yang Yu et al.

The alignment of Large Language Models (LLMs) is critically dependent on reward models trained on costly human preference data. While recent work explores bypassing this cost with AI feedback, these methods often lack a rigorous theoretical foundation. In this paper, we discover that a powerful generalist reward model is already latently present within any LLM trained via standard next-token prediction. We prove that this endogenous reward is not a heuristic, but is theoretically equivalent to a reward function learned through offline inverse reinforcement learning. This connection allows us to directly elicit a high-quality reward signal from a base (pre-trained or supervised fine-tuned) model without any further training. Critically, we also prove that subsequent reinforcement learning using this endogenous reward leads to a policy with a provably superior error bound compared to the base model. To our best knowledge, this is the first theoretical proof of the effectiveness of reinforcement learning for LLMs. Our experiments validate this theory, demonstrating that our method not only outperforms existing LLM-as-a-judge approaches but can also surpass explicitly trained reward models. These findings suggest that the reward modeling stage can be replaced by a principled method of eliciting the knowledge already captured during pre-training, heralding a more efficient, powerful, and scalable paradigm for LLMs alignment as well as multi-modal models.

Wenjie Qiu, Yi-Chen Li, Xuqin Zhang et al.

Learning reward models from human preference datasets and subsequently optimizing language models via reinforcement learning has emerged as a fundamental paradigm for aligning LLMs with human preferences. The performance of the reward model plays a crucial role in the effectiveness of alignment. Previous reward models operate at a coarse-grained level, requiring the generation of a complete response to obtain a reward value. The sparse reward may present challenges for downstream reinforcement learning. While recent efforts have attempted to learn token-level reward models, the lack of explicit semantic information makes it difficult to model the credit of every individual token. In this paper, we propose assigning scores to every sentence, introducing an intermediate-grained reward model. By segmenting the complete response into sentences and applying differential operations to reward output at the start and end positions of each sentence, we can effectively model the rewards of sentences. Moreover, a novel attention mechanism is introduced to aggregate the scores of all sentences into a response-level score, which allows it to be trained using the Bradley-Terry model. On common benchmarks, our method outperforms the response-level reward model by 2.7% on RewardBench (for reward modeling evaluation) and surpasses all baselines on AlpacaEval (for alignment evaluation).