Ensuring consistent persona in interactive AI systems presents a significant challenge, especially in diverse application scenarios ranging from virtual assistants to customer service bots. Such capability is often constrained by the system's understanding of direct and explicit persona conflicts. Traditional approaches primarily focus on detecting discrepancies between machine responses and its predefined profile, or the contextual inconsistencies between the responses at the semantic level rather than the persona level. Due to the lack of a comprehensive persona-specific Commonsense Knowledge Graph, some indirect and implicit persona inconsistencies between machine responses can hardly be identified. In this paper, we build the first persona commonsense knowledge graph (PersonaKG), based on which we then construct a large-scale persona consistency dialogue dataset (PersonaCOM) containing both explicit and implicit persona conflicts between machine responses. With the guidance of the persona commonsense knowledge, we propose a Recognize-Rewrite framework (R2) which first recognizes the responses that are inconsistent in persona with the previous responses, and then rewrites them into consistent ones. The empirical study demonstrates that utilizing R2 method on PersonaCOM with PersonaKG results in a significant improvement of 12.20% in automatic metrics and 10.09% in manual evaluation compared to not using the R2 method and PersonaKG.

In recent years, large language models (LLMs) have demonstrated outstanding capabilities in various tasks. However, LLMs also have various drawbacks, especially hallucination. Hallucination refers to the generation of content that does not align with the user input, contradicts previously generated content or world knowledge. Current research on hallucination mainly include knowledge retrieval, prompt engineering, training data improvement, reinforcement learning, etc. However, these methods do not involve different categories of hallucinations which is important on hallucination analysis, and make detailed investigation for the internal state of LLMs which indicates the direction on hallucination occurrence. Therefore, in our research, we introduce an attribution framework to trace the origins of hallucinations based on the internal signals of LLMs. To support this framework, we develop a new benchmark named RelQA-Cate, which includes eight categories of hallucinations for the answers generated by LLMs. After that, we present a novel Differential Penalty Decoding (DPD) strategy for reducing hallucinations through adjusting post-probabilities of each answer. We conduct a series of experiments and the performance on answer reliability has significant improvement, achieving 28.25% at most, which demonstrates the effectiveness of our proposed DPD and its generalization in mitigating hallucination in LLMs.

Large Language Models (LLMs) exhibit remarkable capabilities across various downstream tasks, including empathetic dialogues. However, a non-trivial question arises: Do they possess high-context empathy and can they generate emotional interactions with humans? High-context empathy, which tends to be more indirect and concise like Chinese-style empathy, differs from the current empathy capabilities of LLMs. These capabilities are predominantly low-context empathy, which is often direct and lengthy, resembling English-style empathy. In this paper, We first construct a comprehensive Chinese High-context Empathy Dialogue dataset (HED), which consists of emotional, role-based emotional, personality-based emotional, and role-personality-based emotional dialogues. Next, we explore whether LLMs have high-context empathy in conversations. After that, we propose an innovative High-context Empathy Network (HEN) to improve LLMs' capabilities in generating high-context empathetic responses. Our empirical study demonstrates that there is much room for LLMs in generating high-context empathetic responses, and the proposed HEN can not only significantly improve LLMs' capabilities in generating high-context empathetic responses, but also has positive effects for LLMs in solving similar sentiment-related tasks.

Large Language Models (LLMs), such as OpenAI’s o1 and DeepSeek’s R1, excel at advanced reasoning tasks like math and coding via Reinforcement Learning with Verifiable Rewards (RLVR), but still struggle with puzzles solvable by humans without domain knowledge. We introduce ENIGMATA, the first comprehensive suite tailored for improving LLMs with puzzle reasoning skills. It includes 36 tasks across 7 categories, each with: 1) a generator that produces unlimited examples with controllable difficulty, and 2) a rule-based verifier for automatic evaluation. This generator-verifier design supports scalable, multi-task RL training, fine-grained analysis, and seamless RLVR integration. We further propose ENIGMATA-Eval, a rigorous benchmark, and develop optimized multi-task RLVR strategies. Our trained model, Qwen2.5-32B-ENIGMATA, consistently surpasses o3-mini-high and o1 on the puzzle reasoning benchmarks like ENIGMATA-Eval, ARC-AGI (32.8%), and ARC-AGI 2 (0.6%). It also generalizes well to out-of-domain puzzle benchmarks and mathematical reasoning, with little multi-tasking trade-off. When trained on larger models like Seed1.5-Thinking (20B activated parameters and 200B total parameters), puzzle data from ENIGMATA further boosts SoTA performance on advanced math and STEM reasoning tasks such as AIME (2024-2025), BeyondAIME and GPQA (Diamond), showing nice generalization benefits of ENIGMATA. This work offers a unified, controllable framework for advancing logical reasoning in LLMs. Project page: https://seed-enigmata.github.io.

Inference scaling empowers LLMs with unprecedented reasoning ability, with reinforcement learning as the core technique to elicit complex reasoning. However, key technical details of state-of-the-art reasoning LLMs are concealed (such as in OpenAI o1 blog and DeepSeek R1 technical report), thus the community still struggles to reproduce their RL training results. We propose the **D**ecoupled Clip and **D**ynamic s**A**mpling **P**olicy **O**ptimization (**DAPO**) algorithm, and fully open-source a state-of-the-art large-scale RL system that achieves 50 points on AIME 2024 using Qwen2.5-32B base model. Unlike previous works that withhold training details, we introduce four key techniques of our algorithm that make large-scale LLM RL a success. In addition, we open-source our training code, which is built on the verl framework, along with a carefully curated and processed dataset. These components of our open-source system enhance reproducibility and support future research in large-scale LLM RL.