Zixiang Di

Jiahao Yuan, Zixiang Di, Zhiqing Cui et al.

Empathetic response generation necessitates the integration of emotional and intentional dynamics to foster meaningful interactions. Existing research either neglects the intricate interplay between emotion and intent, leading to suboptimal controllability of empathy, or resorts to large language models (LLMs), which incur significant computational overhead. In this paper, we introduce ReflectDiffu, a lightweight and comprehensive framework for empathetic response generation. This framework incorporates emotion contagion to augment emotional expressiveness and employs an emotion-reasoning mask to pinpoint critical emotional elements. Additionally, it integrates intent mimicry within reinforcement learning for refinement during diffusion. By harnessing an intent twice reflect mechanism of Exploring-Sampling-Correcting, ReflectDiffu adeptly translates emotional decision-making into precise intent actions, thereby addressing empathetic response misalignments stemming from emotional misrecognition. Through reflection, the framework maps emotional states to intents, markedly enhancing both response empathy and flexibility. Comprehensive experiments reveal that ReflectDiffu outperforms existing models regarding relevance, controllability, and informativeness, achieving state-of-the-art results in both automatic and human evaluations.

Jinyi Han, Zixiang Di, Zishang Jiang et al.

Large language models (LLMs) achieve strong performance by generating long chains of thought, but longer traces always introduce redundant or ineffective reasoning steps. One typical behavior is that they often perform unnecessary verification and revisions even if they have reached the correct answers. This limitation stems from the unstructured nature of reasoning trajectories and the lack of targeted supervision for critical reasoning abilities. To address this, we propose Structured Reasoning (SCR), a framework that decouples reasoning trajectories into explicit, evaluable, and trainable components. We mainly implement SCR using a Generate-Verify-Revise paradigm. Specifically, we construct structured training data and apply Dynamic Termination Supervision to guide the model in deciding when to terminate reasoning. To avoid interference between learning signals for different reasoning abilities, we adopt a progressive two-stage reinforcement learning strategy: the first stage targets initial generation and self-verification, and the second stage focuses on revision. Extensive experiments on three backbone models show that SCR substantially improves reasoning efficiency and self-verification. Besides, compared with existing reasoning paradigms, it reduces output token length by up to 50%.

Jiahao Yuan, Xingzhe Sun, Xing Yu et al.

The LLMSR@XLLM25 formulates a low-resource structural reasoning task that challenges LLMs to generate interpretable, step-by-step rationales with minimal labeled data. We present Less is More, the third-place winning approach in the LLMSR@XLLM25, which focuses on structured reasoning from only 24 labeled examples. Our approach leverages a multi-agent framework with reverse-prompt induction, retrieval-augmented reasoning synthesis via GPT-4o, and dual-stage reward-guided filtering to distill high-quality supervision across three subtasks: question parsing, CoT parsing, and step-level verification. All modules are fine-tuned from Meta-Llama-3-8B-Instruct under a unified LoRA+ setup. By combining structure validation with reward filtering across few-shot and zero-shot prompts, our pipeline consistently improves structure reasoning quality. These results underscore the value of controllable data distillation in enhancing structured inference under low-resource constraints. Our code is available at https://github.com/JhCircle/Less-is-More.

Houping Yue, Zixiang Di, Mei Jiang et al.

While Large Language Models (LLMs) are reshaping the paradigm of AI for Social Science (AI4SS), rigorously evaluating their capabilities in scholarly writing remains a major challenge. Existing benchmarks largely emphasize single-shot, monolithic generation and thus lack the fine-grained assessments required to reflect complex academic research workflows. To fill this gap, we introduce EduResearchBench, the first comprehensive evaluation platform dedicated to educational academic writing. EduResearchBench is built upon our Hierarchical Atomic Task Decomposition (HATD) framework, which decomposes an end-to-end research workflow into six specialized research modules (e.g., Quantitative Analysis, Qualitative Research, and Policy Research) spanning 24 fine-grained atomic tasks. This taxonomy enables an automated evaluation pipeline that mitigates a key limitation of holistic scoring, where aggregate scores often obscure specific capability bottlenecks, and instead provides fine-grained, diagnostic feedback on concrete deficiencies. Moreover, recognizing the high cognitive load inherent in scholarly writing, we propose a curriculum learning strategy that progressively builds competence from foundational skills to complex methodological reasoning and argumentation. Leveraging 55K raw academic samples, we curate 11K high-quality instruction pairs to train EduWrite, a specialized educational scholarly writing model. Experiments show that EduWrite (30B) substantially outperforms larger general-purpose models (72B) on multiple core metrics, demonstrating that in vertical domains, data quality density and hierarchically staged training curricula are more decisive than parameter scale.

Zixiang Di, Jinyi Han, Shuo Zhang et al.

Learning from negative samples holds great promise for improving Large Language Model (LLM) reasoning capability, yet existing methods treat all incorrect responses as equally informative, overlooking the crucial role of sample quality. To address this, we propose Plausible Negative Samples (PNS), a method that synthesizes high-quality negative samples exhibiting expected format and structural coherence while ultimately yielding incorrect answers. PNS trains a dedicated model via reverse reinforcement learning (RL) guided by a composite reward combining format compliance, accuracy inversion, reward model assessment, and chain-of-thought evaluation, generating responses nearly indistinguishable from correct solutions. We further validate PNS as a plug-and-play data source for preference optimization across three backbone models on seven mathematical reasoning benchmarks. Results demonstrate that PNS consistently outperforms other negative sample synthesis methods, achieving an average improvement of 2.03% over RL-trained models.

Jiahao Yuan, Zixiang Di, Shangzixin Zhao et al.

Large language models (LLMs) face challenges in aligning with diverse cultural values despite their remarkable performance in generation, which stems from inherent monocultural biases and difficulties in capturing nuanced cultural semantics. Existing methods struggle to adapt to unknown culture after fine-tuning. Inspired by cultural geography across five continents, we propose Cultural Palette, a multi-agent framework that redefines cultural alignment as an adaptive "color-blending" process for country-specific adaptation. Our approach harnesses cultural geography across five continents (Africa, America, Asia, Europe, Oceania) through three key steps: First, we synthesize the Pentachromatic Cultural Palette Dataset using GPT-4o, refining continental-level dialogues with Hofstede's cultural dimensions to establish foundational cultural representations. Second, five continent-level alignment agents form specialized cultural communities that generate region-specific draft responses. Third, a Meta Agent employs Cultural MoErges to dynamically blend these cultural "colors" through attention-gated parameter merging, akin to mixing pigments on a palette, resolving conflicts while preserving cultural nuances to produce the final culturally-aligned response. Extensive experiments across various countries demonstrate that Cultural Palette surpasses existing baselines in cultural alignment.

Jiahao Yuan, Dehui Du, Hao Zhang et al.

Large language models (LLMs) have shown remarkable performance in reasoning tasks but face limitations in mathematical and complex logical reasoning. Existing methods to improve LLMs' logical capabilities either involve traceable or verifiable logical sequences that generate more reliable responses by constructing logical structures yet increase computational costs, or introduces rigid logic template rules, reducing flexibility. In this paper, we propose Reversal of Thought (RoT), a plug-and-play and cost-effective reasoning framework designed to enhance the logical reasoning abilities of LLMs during the warm-up phase prior to batch inference. RoT utilizes a Preference-Guided Reverse Reasoning warm-up strategy, which integrates logical symbols for pseudocode planning through meta-cognitive mechanisms and pairwise preference self-evaluation to generate task-specific prompts solely through demonstrations, aligning with LLMs' cognitive preferences shaped by RLHF. Through reverse reasoning, we utilize a Cognitive Preference Manager to assess knowledge boundaries and further expand LLMs' reasoning capabilities by aggregating solution logic for known tasks and stylistic templates for unknown tasks. Experiments across various tasks demonstrate that RoT surpasses existing baselines in both reasoning accuracy and efficiency.

Bingdong Li, Zixiang Di, Yongfan Lu et al.

Multi-objective Bayesian optimization (MOBO) has shown promising performance on various expensive multi-objective optimization problems (EMOPs). However, effectively modeling complex distributions of the Pareto optimal solutions is difficult with limited function evaluations. Existing Pareto set learning algorithms may exhibit considerable instability in such expensive scenarios, leading to significant deviations between the obtained solution set and the Pareto set (PS). In this paper, we propose a novel Composite Diffusion Model based Pareto Set Learning algorithm, namely CDM-PSL, for expensive MOBO. CDM-PSL includes both unconditional and conditional diffusion model for generating high-quality samples. Besides, we introduce an information entropy based weighting method to balance different objectives of EMOPs. This method is integrated with the guiding strategy, ensuring that all the objectives are appropriately balanced and given due consideration during the optimization process; Extensive experimental results on both synthetic benchmarks and real-world problems demonstrates that our proposed algorithm attains superior performance compared with various state-of-the-art MOBO algorithms.

Yongfan Lu, Zixiang Di, Bingdong Li et al.

Multi-objective combinatorial optimization (MOCO) problems are prevalent in various real-world applications. Most existing neural MOCO methods rely on problem decomposition to transform an MOCO problem into a series of singe-objective combinatorial optimization (SOCO) problems and train attention models based on a single-step and deterministic greedy rollout. However, inappropriate decomposition and undesirable short-sighted behaviors of previous methods tend to induce a decline in diversity. To address the above limitation, we design a Context-aware Diversity Enhancement algorithm named CDE, which casts the neural MOCO problems as conditional sequence modeling via autoregression (node-level context awareness) and establishes a direct relationship between the mapping of preferences and diversity indicator of reward based on hypervolume expectation maximization (solution-level context awareness). Based on the solution-level context awareness, we further propose a hypervolume residual update strategy to enable the Pareto attention model to capture both local and non-local information of the Pareto set/front. The proposed CDE can effectively and efficiently grasp the context information, resulting in diversity enhancement. Experimental results on three classic MOCO problems demonstrate that our CDE outperforms several state-of-the-art baselines.

Bingdong Li, Zixiang Di, Yanting Yang et al.

In this paper, we introduce a novel approach for addressing the multi-objective optimization problem in large language model merging via black-box multi-objective optimization algorithms. The goal of model merging is to combine multiple models, each excelling in different tasks, into a single model that outperforms any of the individual source models. However, model merging faces two significant challenges: First, existing methods rely heavily on human knowledge or intuition. Second, it's difficult to obtain the great model merging configuration in limited evaluations. To address these challenges, we formalize model merging as a multi-objective optimization problem and propose an automated optimization approach named MM-MO. This method leverages multi-objective optimization algorithms to autonomously search for optimal merging configurations across various tasks, alleviating the need for human intervention. In MM-MO, a weak-to-strong method is employed to enhance the acquisition function, allowing previously evaluated superior configurations to guide the search for new ones. Meanwhile, Fisher information is applied to screen these configurations, increasing the possibility of identifying high-quality merging configuration. Additionally, we designed a sparsity metric as an additional optimization objective to enhance the model's generalization performance across different tasks. We conducted comprehensive experiments with other mainstream model merging methods, demonstrating that the proposed MM-MO algorithm is competitive and effective in achieving high-quality model merging.