Xiangchi Yuan, Xingjian Diao, Zheyuan Liu, Chunhui Zhang, Dachuan Shi, Kejing Xia, Zhenbang Du, Shenglong Yao 等 10 位作者
LLM post-training often optimizes multiple objectives with unequal priorities, such as correctness and reasoning efficiency. However, the widely used additive reward aggregation in RL can violate these priorities: when primary and auxiliary advantages conflict, a sufficiently large auxiliary signal can overturn the preference induced by the primary objective. We term this problem \emph{advantage priority inversion}. Although reward-level gating can protect priority more directly, it still removes potentially useful auxiliary distinctions among rollout samples that fail the primary condition. To address this trade-off, we propose Primary Advantage Policy Optimization (\textbf{PAPO}), a priority-aware advantage aggregation method that asymmetrically combines a primary and an auxiliary advantage. PAPO retains auxiliary supervision when the objectives agree and smoothly attenuates conflicting auxiliary signals, preserving the primary sign during fusion for \emph{primary-consistent ordering}. This rule applies to sequence-level advantages in GRPO and token-level advantages in actor--critic PPO. Experiments on multiple domains and models show that PAPO consistently achieves the best primary objective while maintaining best or competitive auxiliary-objective performance across both GRPO- and PPO-based optimization. \href{https://xiangchi-yuan.github.io/papo-project/}{Code is available}.
Xiangchi Yuan ByteDance Inc., Dachuan Shi Meta Superintelligence Labs and Georgia Institute of Technology, Chunhui Zhang Dartmouth College, Zheyuan Liu University of Notre Dame, Shenglong Yao, Soroush Vosoughi Dartmouth College, Wenke Lee Georgia Institute of Technology
Reinforcement learning (RL) is central to post-training, particularly for agentic models that require specialized reasoning behaviors. In this setting, model merging offers a practical mechanism for integrating multiple RL-trained agents from different tasks into a single generalist model. However, existing merging methods are designed for supervised fine-tuning (SFT), and they are suboptimal to preserve task-specific capabilities on RL-trained agentic models. The root is a task-vector mismatch between RL and SFT: on-policy RL induces task vectors that are highly sparse and heterogeneous, whereas SFT-style merging implicitly assumes dense and globally comparable task vectors. When standard global averaging is applied under this mismatch, RL’s non-overlapping task vectors that encode critical task-specific behaviors are reduced and parameter updates are diluted. To address this issue, we propose Reinforced Agent Merging (RAM), a distribution-aware merging framework explicitly designed for RL-trained agentic models. RAM disentangles shared and task-specific unique parameter updates, averaging shared components while selectively preserving and rescaling unique ones to counteract parameter update dilution. Experiments across multiple agent domains and model architectures demonstrate that RAM not only surpasses merging baselines, but also unlocks synergistic potential among agents to achieve performance superior to that of specialized agents in their domains.
尚无 AI 解读，展示原文摘要
cs.CL · cs.AI
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Curriculum learning (CL), which orders training data from easy to hard, has become a popular strategy for improving reasoning in large language models (LLMs). Yet prior work employs disparate difficulty metrics and training setups, leaving open fundamental questions: When does curriculum help? Which direction—forward or reverse—is better? And does the answer depend on what we measure? We address these questions through a unified offline evaluation framework that decomposes curriculum difficulty into five complementary dimensions: Problem Difficulty, Model Surprisal, Confidence Margin, Predictive Uncertainty, and Decision Variability. Through controlled post-training experiments on mathematical reasoning benchmarks with Llama3.1-8B, Mistral-7B, and Gemma3-4B, we find that: (i) no curriculum strategy dominates universally—the relative effectiveness of forward versus reverse CL depends jointly on model capability and task complexity; (ii) even within a single metric, samples at different difficulty levels produce distinct gains depending on task demands; and (iii) Task-aligned curricula focus on shaping the model’s final representations and generalization, whereas inner-state curricula modulate internal states such as confidence and uncertainty. Our findings challenge the notion of a universal curriculum strategy and offer actionable guidance across model and task regimes, with some metrics indicating that prioritizing decision-uncertain samples can further enhance learning outcomes.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Pengqi Li, Lizhong Ding Beijing Institute of Technology, Zhehao Zhou, Chunhui Zhang, Jiarun Fu, Hao Li Beijing Institute of Technology, Ye Yuan Beijing Institute of Technology, Guoren Wang Beijing Institute of Technology
Hallucinations arise when large language models (LLMs) guess rather than acknowledge their underlying uncertainty. Existing static strategies for mitigating hallucinations have been only partially successful, largely because they do not explicitly model the information gain from interacting with the external environment. Researchers need a general method to proactively steer users toward informative clarifications, thereby unlocking the model’s effective capacity under underspecified inputs. We model the uncertainty of LLMs in interactive settings and uncover the mechanism of active calibration between model concepts and human evaluations, improving reliability. We show that calibration error in LLMs density estimation admits a non-vanishing lower bound under non-interactive learning, while interaction empirically reduces it. We further characterize that calibration error identifies informative queries and that calibration can be accelerated by shifting query distributions from imbalanced to balanced regimes. Guided by these insights, we propose a calibration-driven Interactive Learning Strategy (ILS) that selects clarification queries by optimizing calibration error, providing both theoretical guarantees and empirical gains for reliability. Code and data are available at https://github.com/zhouyeah215/Demystifying_Uncertainty.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表
Zhongyu Ouyang, Qianlong Wen ByteDance Inc., Chunhui Zhang Dartmouth College, Yanfang Ye University of Notre Dame, Soroush Vosoughi Dartmouth College
What enables large language models (LLMs) to effectively model user preferences in sequential recommendation? Our investigation reveals that existing preference-alignment approaches largely rely on binary pairwise comparisons, overlooking two critical factors: preference intensity (the structured strength of affinity or aversion) and temporal context (the extent to which recent interactions better reflect a user’s current intent). Through controlled experiments, we show that leveraging comprehensive feedback with structured preference signals substantially improves recommendation performance, indicating that binary modeling discards essential information. Motivated by these findings, we propose RecPO, a unified preference optimization framework that maps both explicit and implicit feedback into a common preference signal and constructs adaptive reward margins that jointly account for preference intensity and interaction recency. Experiments across five datasets show that RecPO consistently outperforms state-of-the-art baselines while exhibiting behavioral patterns aligned with human decision-making, including favoring immediate satisfaction, maintaining preference coherence, and avoiding dispreferred items. Our results highlight that preference intensity and temporal context are fundamental ingredients for effective LLM-based recommendation. Code: https://github.com/zyouyang/RecPO
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers) · 2026 · 正式发表