Large language models (LLMs) are increasingly deployed in user-facing applications, raising concerns that they may reflect and amplify social biases. We investigate social identity biases in Chinese LLMs using Mandarin-specific prompts across ten representative models. Our evaluation compares ingroup (“We”) and outgroup (“They”) framings across 240 social groups salient in the Chinese context, using a two-tiered measurement framework that assesses both sentiment and toxicity. The prompt design explicitly accounts for linguistic properties of Mandarin, including the distinction between the default plural pronoun 他们 and the explicitly feminine plural 她们, enabling a controlled comparison of social identity framing effects. Across models, we observe systematic ingroup–outgroup asymmetries, although their expression differs across measurement dimensions. In particular, instruction tuning often reduces sentiment asymmetries, while toxicity gaps remain more persistent. Moreover, the feminine-marked plural 她们 is associated with higher toxicity than the default plural in several models. Our study introduces a language-aware evaluation framework for Chinese LLMs and shows that (i) social identity biases previously documented in English also manifest in Chinese and that (ii) Mandarin-specific linguistic structure can reveal bias patterns that are not directly observable in English-only settings.
论文检索
输入标题、作者或关键词，从 100,903 篇学术成果中精准定位
Reinforcement learning (RL) offers a principled way to enhance the reasoning capabilities of large language models, yet its effectiveness hinges on training signals that remain informative as models evolve. In practice, RL progress often slows when task difficulty becomes poorly aligned with model capability or when training is dominated by a narrow set of recurring problem patterns.To jointly address these issues, we propose SCALER (Synthetic sCalable Adaptive Learning Environment for Reasoning), a framework that sustains effective learning signals through adaptive environment design.SCALER introduces a scalable synthesis pipeline that converts real-world programming problems into verifiable reasoning environments with controllable difficulty and unbounded instance generation, enabling RL training beyond finite datasets while preserving strong correctness guarantees. Building on this, SCALER further employs an adaptive multi-environment RL strategy that dynamically adjusts instance difficulty and curates the active set of environments to track the model’s capability frontier and maintain distributional diversity. This co-adaptation prevents reward sparsity, mitigates overfitting to narrow task patterns, and supports sustained improvement throughout training. Extensive experiments show that SCALER consistently outperforms other RL baselines across diverse reasoning benchmarks and exhibits more stable, long-horizon training dynamics.
Exploring Two-Phase Continual Instruction Fine-tuning for Multilingual Adaptation in Large Language Models
PDF ↗A key challenge for Large Language Models (LLMs) is improving their Multilingual instruction-following ability over time without deteriorating their ability in languages they already excel at, typically English. In this paper, we study a two-phase Continual Fine-tuning (CFT) setup toward improving a model’s Multilingual adaptability. Concretely, we consider a two-phase CFT process in which an English-only end-to-end instruction fine-tuned LLM (Phase 1) is sequentially fine-tuned on a multilingual instruction dataset (Phase 2). Across MISTRAL-7B and LLAMA-3-8B and multiple dataset pairs, we show that instructional similarity between phases is critical: aligned datasets preserve or improve English while boosting multilingual ability, whereas misaligned datasets cause English degradation. We show that this degradation arises from representation shift during CFT, and that targeted mitigation strategies, including generative replay and heuristic-based layer freezing, reduce this shift and improve multilingual adaptation.
AROMA: Augmented Reasoning Over a Multimodal Architecture for Virtual Cell Genetic Perturbation Modeling
PDF ↗Virtual cell modeling predicts molecular state changes under genetic perturbations in silico, which is essential for biological mechanism studies. However, existing approaches suffer from unconstrained reasoning, uninterpretable predictions, and retrieval signals that are weakly aligned with regulatory topology. To address these limitations, we propose AROMA, an Augmented Reasoning Over a Multimodal Architecture for virtual cell genetic perturbation modeling. AROMA integrates textual evidence, graph-topology information, and protein sequence features to model perturbation-target dependencies, and is trained with a two-stage optimization strategy to yield predictions that are both accurate and interpretable. We also construct two knowledge graphs and a perturbation reasoning dataset, PerturbReason, containing more than 498k samples, as reusable resources for the virtual cell domain. Experiments show that AROMA outperforms existing methods across multiple cell lines, and remains robust under zero-shot evaluation on an unseen cell line, as well as in knowledge-sparse, long-tail scenarios. Overall, AROMA demonstrates that combining knowledge-driven multimodal modeling with evidence retrieval provides a promising pathway toward more reliable and interpretable virtual cell perturbation prediction. Model weights are available at https://huggingface.co/blazerye/AROMA. Code is available at https://github.com/blazerye/AROMA.
Parameter-Efficient Fine-Tuning (PEFT) methods, especially LoRA, are widely used for adapting pre-trained models to downstream tasks due to their computational and storage efficiency. However, in the context of LoRA and its variants, the potential of activation subspaces corresponding to tail eigenvectors remains substantially under-exploited, which may lead to suboptimal fine-tuning performance. In this work, we propose Astra (Activation-Space Tail-Eigenvector Low-Rank Adaptation), a novel PEFT method that leverages the tail eigenvectors of the model output activations—estimated from a small task-specific calibration set—to construct task-adaptive low-rank adapters. By constraining updates to the subspace spanned by these tail eigenvectors, Astra achieves faster convergence and improved downstream performance with a significantly reduced parameter budget. Extensive experiments across natural language understanding (NLU) and natural language generation (NLG) tasks demonstrate that Astra consistently outperforms existing PEFT baselines across 16 benchmarks and even surpasses full fine-tuning (FFT) in certain scenarios.
This study demonstrates an alignment of per-word processing time in a popular state-space language model Mamba and human readers. In Mamba, the recurrent state transition at each layer conceptually takes some duration of time, the discretization timestep \Delta_t, determined dynamically in response to the input. Using a naturalistic reading dataset, we show that the per-word timestep from Mamba is a powerful predictor of human reading times, comparable to strong baselines such as word frequency and GPT-2 surprisal and significant even when they are controlled for. We further suggest, through formal analysis of Mamba’s architecture and internal dynamics, that Mamba can serve as a new, valuable lens to look at human real-time language processing with ever-updated memory, because it allows us to look at how each module (layer) weighs short- and long-term information retention, and how noise may interact with dynamic, continuous memory representation. Code is available via an (anonymized) link.
Lending Eyesight to Language Models: Modeling and Probing Human scanpath through Transformer Decoder
PDF ↗Human scanpaths offer rich and reliable clues about the cognitive mechanisms underlying language comprehension. Decoder-only language models, typically large language models (LLMs), have proven to exhibit striking parallels with human cognitive processes. In this study, we investigate to what extent language models can be endowed with human-like gaze shifts. Besides, by probing scanpath through eye model, analogous to probing language through language models, we ask whether such modeling can yield novel knowledge of the cognitive machinery of sense making.This study presents a novel plug-and-play module, EyeLM, to transform an autoregressive language model into an autoregressive eye model, thus facilitating a probabilistic spatial modeling of human explicit attention. Our EyeLM module, powered by LLMs, achieves competitive performance with novel cognitive probing capabilities. By probing EyeLM, we can reach the predictability and uncertainty of the scanpath. Exhibiting aligned patterns with prior knowledge about human reading comprehension, these probabilistic measures of scanpath act as promising predictors of human comprehension skills.
Grouped Adaptive Weight Sharing (GAWS): An Inference-Efficient Adaptation Method for Large Language Models
PDF ↗Although Low-Rank Adaptation (LoRA) revolutionized parameter-efficient fine-tuning, it often incurs an inference overhead due to the extra computation required by adapter layers. While most literature focuses on maximizing accuracy or minimizing parameter counts, this paper prioritizes single-request inference performance in the unmerged adapter setting, where adapters must remain decoupled from the base model at runtime. By analyzing LoRA adapters on GPUs, we identify segmented function calls as the primary source of this latency. To address this, we propose Grouped Adaptive Weight Sharing (GAWS), a novel adapter design based on structured Kronecker product decomposition. Experiments on T5-3B, GPT-2 Large, LLaMA3.2-3B, and RoBERTa-Large show that GAWS reduces latency to about 40% of the gap between the unmerged LoRA and the base model, while maintaining parameter efficiency and comparable accuracy. This positions GAWS as a Pareto-efficient solution for deploying adapted LLMs in latency-sensitive settings, balancing the high latency of compressed adapters with the accuracy of LoRA. The source code is available at:https://github.com/SamsungLabs/GAWS .
RouterHGC: Optimized Router for LLM-based Multi-Agent Systems via Heterogeneous Graph Contrastive Learning
PDF ↗Leveraging powerful planning and reasoning capabilities, Large Language Models (LLMs)-driven Multi-Agent Systems (MAS) have demonstrated remarkable scalability and generalizability across complex tasks. However, dynamically routing the optimal combination of agents and collaboration modes for a given query to balance performance and cost remains challenging. To address the limitation of prior work, which focuses on single-agent settings and overlooks collaborative structures and role assignment in MAS, we propose RouterHGC, the first heterogeneous graph contrastive learning framework for MAS routing. We formalize routing as node selection through edge-weight prediction on a heterogeneous graph whose node types include user queries, collaboration modes, agent roles, and LLMs, with message passing capturing their high-order dependencies. We further design a novel global–local contrastive loss function to jointly optimize graph-level representations and edge-level selections, pulling each query graph toward high-performing positives while pushing it away from underperforming or costly negatives. Experiments on five public datasets covering mathematical reasoning, code generation, and knowledge question answering show that RouterHGC outperforms the best single LLM and baselines, achieving 0.80%–6.17% accuracy gains on MATH and HotpotQA while reducing inference cost by 27.40%.
Scaling Laws or Threshold Effects: Exploring the Optimal Vocabulary Size for Balancing Performance and Efficiency in Low-Resource Languages
PDF ↗While vocabulary expansion scaling laws are well-established for high-resource languages, they remain unverified in low-resource settings. This gap is particularly critical for Byte-level BPE (BBPE), where constrained vocabulary sizes often fail to capture the rich morphemes of complex scripts, leading to severe over-segmentation in languages such as Mongolian, Tibetan, and Uyghur. We systematically investigate jointly-scaled trilingual vocabulary for these languages (140 to 195,000 tokens) across BPE (Llama 2) and BBPE (Qwen2.5/3) architectures. Our results reveal that BBPE follows a "decline-then-rise" pattern, requiring a 9,000-token threshold (3,000 per language) to trigger non-linear performance gains and inference acceleration, whereas BPE improves monotonically. Using Pareto Frontier Analysis, we identify an optimal 79,500-token configuration for BBPE that reduces continuous pre-training duration by over 71% across 1.5B to 8B parameter models while consistently enhancing downstream performance.
Integrating knowledge graphs (KGs) with large language models (LLMs) enhances factual accuracy and interpretability in question answering. However, existing agent-based methods rely on static memory mechanisms that fail to address the combinatorial explosion of search spaces in multi-hop reasoning and lack continuous learning capabilities. To overcome these limitations, we propose EvoMemKG, an agent framework with a dynamic, evolvable memory mechanism specifically designed for KG reasoning. EvoMemKG features a dual-layer memory architecture: (1) a working memory that losslessly compresses retrieved triplets through clustering to manage exploration states, effectively linearizing the exponential state space expansion; and (2) an experience memory that abstracts historical reasoning paths into reusable, generalized strategies, enabling cross-task knowledge transfer and self-evolution. We further introduce a double-loop workflow that orchestrates the LLM, memory layers, and KG environment to enable end-to-end autonomous reasoning. Extensive evaluations on three KGQA datasets across two KGs demonstrate that EvoMemKG achieves state-of-the-art performance without requiring additional training or specialized tools. Notably, it achieves improvements of up to 20% over the strong baseline on complex multi-hop queries, validating the effectiveness of our dynamic memory approach.
Reasoning language models (RLMs) achieve strong performance on complex reasoning tasks, yet they still exhibit a multilingual reasoning gap, performing better in high-resource languages than in low-resource ones. While recent efforts have been made to address this gap, its underlying causes remain largely unexplored. In this work, we show that this gap primarily stems from failures in language understanding—specifically, the model’s inability to translate multilingual inputs into the language dominating its reasoning traces (typically English). As identifying understanding failures can enable targeted mitigation of the gap, we evaluate a range of detection methods and find that understanding failures are detectable to a meaningful extent, with supervised approaches performing best. Building on this, we propose Selective Translation, a strategy that incorporates an English translation into the initial reasoning trace when an understanding failure is detected. Experimental results using Qwen3-4B show that Selective Translation substantially bridges the multilingual reasoning gap, achieving near full-translation performance while translating only about 20% of inputs. Together, our results show that failures in language understanding are the primary driver of the multilingual reasoning gap and can be detected and selectively mitigated, clarifying its origin and suggesting a path toward more equitable multilingual reasoning.
Aligning Large Vision-Language Models (LVLMs) to mitigate hallucinations typically relies on high-quality preference data. However, in self-supervised settings, standard binary preference optimization (e.g., DPO) suffers from noisy supervision and semantic ambiguity, as automatically generated chosen responses are not guaranteed to be superior to rejected ones. In this work, we propose Trident, a fully self-supervised framework that ensures robust alignment via a structured triplet paradigm. Trident autonomously constructs reliable preference triplets—comprising semantically enriched (chosen), degraded (rejected), and neutral (anchor) responses—through automated visual perturbations and self-summarization. We further introduce Trident Preference Regularization (TPR), a novel objective that utilizes an adaptive margin to enforce semantic separation between the triplet components while preventing deviation from the pretrained distribution. Despite requiring no human annotations or external reward models, Trident consistently outperforms state-of-the-art RLHF and RLAIF baselines. For instance, on LLaVA-1.5-7B, it reduces the hallucination rate on AMBER to 11.3% and achieves 95.70% precision on POPE using only 4k self-generated triplets and a single epoch. This validates structured triplet supervision as a scalable paradigm for robust self-supervised alignment.
Leveraging Label Semantics and Entity Description Generation for LLM-based Fine-grained Entity Typing
PDF ↗Fine-grained entity typing (FET) aims to assign semantically rich and contextually appropriate types to entity mentions. While recent studies have explored the use of large language models (LLMs) for this task, two key challenges persist. First, FET typically involves a large number of entity types, making it difficult for LLMs to perform accurate classification. Second, the presence of label noise in the training data introduced by automatic supervision methods hinders effective fine-tuning. To address these challenges, we propose DR-FET, a descriptor-based retrieval-augmented framework that reduces the effective label space and constructs high-precision training data under noisy supervision. Our method introduces natural language descriptors as an intermediate semantic representation for both entity mentions and types. During inference, entity descriptors are used to retrieve a small set of semantically relevant candidate types, enabling the LLM to perform fine-grained classification under explicit candidate constraints. During training, the same descriptor and retrieval mechanism is reused to identify high-confidence instances from distantly supervised data, prioritizing label precision for efficient fine-tuning. Experiments on two widely used benchmarks demonstrate that the proposed method consistently outperforms existing fine-grained entity typing approaches under noisy supervision.
Prompt tuning has achieved remarkable progress in vision–language models (VLMs) and is recently being adopted for audio–language models (ALMs). However, its generalization ability in ALMs remains largely underexplored. We observe that conventional prompt tuning for ALMs also suffers from the Base–New Tradeoff, and we identify that this issue stems from the disrupted semantic structure of the embedding space. To address this issue, we propose Semantically Expanded Prompt Tuning (SEPT)—a plug-and-play framework that explicitly regularizes the prompt embedding space by incorporating semantic neighbors generated by large language models. SEPT introduces a novel semantic expansion loss with margin constraints that promote intra-class compactness and inter-class separability, thereby enhancing the semantic structure of the prompt embedding space. For comprehensive evaluation, we establish the first benchmark setup for prompt generalization in ALMs, covering both base-to-new generalization and cross-dataset transferability. Extensive experiments demonstrate that SEPT consistently improves generalization performance across multiple prompt tuning baselines, while maintaining computational cost during inference.
Enhancing the task-specific capabilities of Large Language Models (LLMs) primarily requires substantial instruction-tuning datasets. However, the sheer volume of such data imposes a considerable annotation cost, and a lack of optimization methods for tailoring LLMs to specific tasks persists. To address the above issues, we propose a Planning framework for constructing Extractive-based LLMs called PlanE, which includes data decomposition, instruction tuning, and prompt inference. Additionally, we introduce a Data-Tuning-Inference (DTI) planner, aimed at selecting the optimal base-LLM and its DTI combinations for specific datasets to improve construction efficiency. The experimental results demonstrate the effectiveness of our PlanE from two views: (1) across different datasets using the same base-LLM, and (2) on the same dataset using different base-LLMs. Furthermore, we validate the generalizability of the proposed DTI planner under different optimization objectives. The codes are publicly available at https://github.com/gugugu-469/PlanE.
Large Audio Language Models (LALMs) have garnered significant research interest. Despite being built upon text-based large language models (LLMs), LALMs frequently exhibit a degradation in knowledge and reasoning capabilities. We hypothesize that this limitation stems from the failure of current training paradigms to effectively bridge the acoustic-semantic gap within the feature representation space. To address this challenge, we propose CORD, a unified alignment framework that performs online cross-modal self-distillation. Specifically, it aligns audio-conditioned reasoning with its text-conditioned counterpart within a unified model. Leveraging the text modality as an internal teacher, CORD performs multi-granularity alignment throughout the audio rollout process. At the token level, it employs on-policy reverse KL divergence with importance-aware weighting to prioritize early and semantically critical tokens. At the sequence level, CORD introduces a judge-based global reward to optimize complete reasoning trajectories via Group Relative Policy Optimization (GRPO). Empirical results across multiple benchmarks demonstrate that CORD consistently enhances audio-conditioned reasoning and substantially bridges the audio–text performance gap with only 80k synthetic training samples, validating the efficacy and data efficiency of our on-policy, multi-level cross-modal alignment approach.
The application of physics formulas is a fundamental human capability in numerical reasoning. While existing datasets often rely on implicit mathematical knowledge, they rarely explicitate the underlying formulas. To address this, we introduce FormulaReasoning, a new benchmark for formula-based numerical reasoning comprising 5,324 questions requiring calculations grounded in external physics principles. We provide high-quality, fine-grained annotations in English and Chinese—including formula structures, parameter names, symbols, values, and units—curated through manual effort and LLM-assisted validation. Additionally, we provide a consolidated formula database as an external knowledge source. To further challenge model performance, we develop an extended version of the dataset by coupling multiple questions. We evaluate various architectural and methodological frameworks, including retrieval-augmented methods, modular reasoning (formula generation, parameter extraction, and calculation), and preference-based optimization. Our analysis identifies critical challenges in formula-based reasoning, highlighting significant opportunities for future methodological advancement.
Relevance and utility are two frequently used measures to evaluate the effectiveness of an information retrieval (IR) system. Relevance emphasizes the aboutness of a result to a query, while utility refers to the result’s usefulness or value to an information seeker. In Retrieval-Augmented Generation (RAG), high-utility results should be prioritized to feed to LLMs due to their limited input bandwidth. Re-examining RAG’s three core components—relevance ranking derived from retrieval models, utility judgments, and answer generation—aligns with Schutz’s philosophical system of relevances, which encompasses three types of relevance representing different levels of human cognition that enhance each other. These three RAG components also reflect three cognitive levels for LLMs in question-answering. Therefore, we propose an Iterative utiliTy judgmEnt fraMework (ITEM) to promote each step in RAG. We conducted extensive experiments on retrieval (TREC DL, WebAP), utility judgment task (GTI-NQ), and factoid question-answering (NQ) datasets. Experimental results demonstrate significant improvements of in utility judgments, ranking, and answer generation upon representative baselines.
Presentation generation requires deep content research, coherent visual design, and iterative refinement based on observation. However, existing presentation agents often rely on predefined workflows and fixed templates. To address this, we present DeepPresenter, an agentic framework that adapts to diverse user intents, enables effective feedback-driven refinement, and generalizes beyond a scripted pipeline. Specifically, DeepPresenter autonomously plans, renders, and revises intermediate slide artifacts to support long-horizon refinement with environmental observations. Furthermore, rather than relying on self-reflection over internal signals (e.g., reasoning traces), our environment-grounded reflection conditions the generation process on perceptual artifact states (e.g., rendered slides), enabling the system to identify and correct presentation-specific issues during execution. Results on the evaluation set covering diverse presentation-generation scenarios show that DeepPresenter achieves state-of-the-art performance, and the fine-tuned DeepPresenter-9B remains highly competitive at substantially lower cost.