论文检索

Tiehua Mei, Minxuan Lv, Leiyu Pan, Zhenpeng Su, Hongru Hou, Hengrui Chen, Ao Xu, Deqing Yang

Reinforcement Learning with Verifiable Rewards (RLVR) improves reasoning in large language models but treats all correct solutions equally, potentially reinforcing flawed traces that arrive at correct answers by chance. We observe that better reasoning makes better demonstrations: high-quality solutions serve as more effective in-context examples than low-quality ones. We term this teaching ability Demonstration Utility, and show that the policy model’s own in-context learning ability provides an efficient way to measure it, yielding a quality signal termed Evidence Gain. To leverage this signal during training, we introduce In-Context RLVR, which prepends demonstrations before each rollout. Theoretically, we prove that this simple input modification implicitly reweights rewards by a factor approximately proportional to Evidence Gain, assigning higher weights to high-quality traces without requiring costly computation. Experiments on mathematical reasoning benchmarks demonstrate consistent improvements in both accuracy and reasoning quality over standard RLVR baselines. Our codes and datasets are available at https://github.com/Mithas-114/IC-DAPO.

Cheng Yang, Xuemeng Yang, Licheng Wen, Daocheng Fu, Jianbiao Mei, Rong Wu, Pinlong Cai, Yufan Shen, Nianchen Deng, Jia Xu 等

Large Language Models often struggle with complex, multi-step operational tasks because they remain static during inference and cannot learn from past experience. To address this, we propose MUSE, a framework that enables iterative self-improvement through a hierarchical Memory Module. MUSE organizes cross-domain insights to facilitate the orchestration of long-horizon workflows. The core of our approach is an autonomous post-execution critique mechanism: after completing each sub-task, the system analyzes its operational logs and distills raw execution data into structured, reusable knowledge. This allows the agent to evolve dynamically rather than relying on fixed parameters. Evaluated on the rigorous TAC productivity benchmark, MUSE achieves new state-of-the-art results, significantly outperforming previous methods using only the streamlined Gemini-2.5 Flash model. Our analysis demonstrates that MUSE’s performance scales with the accumulation of insights and exhibits strong cross-task transferability, marking a key step toward autonomous systems capable of lifelong learning in professional environments. Demo videos can be found in our supplementary materials.

Seohee Yoon, Yong Suk Choi

Mixture-of-Experts (MoE) architectures have emerged as an effective approach for scaling Large Language Models (LLMs) by activating only a subset of experts during inference. Despite their computational efficiency, MoE models incur a substantial memory bottleneck from maintaining all expert parameters during inference. To address this challenge, numerous MoE pruning methods have been proposed. However, most existing methods adopt uniform pruning across layers, which fails to capture layer-wise variations in expert importance and redundancy. In this paper, we propose COmpensated MoE Pruning with Expert-Layer distribution (COMPEL). COMPEL performs layer-adaptive expert pruning by estimating expert importance using Fisher information and deriving layer importance from layer-wise outlier distributions, enabling pruning decisions that capture layer-wise heterogeneity. Furthermore, to mitigate performance degradation resulting from expert pruning, we propose a Fisher information guided expert weight compensation method. Experimental results on the Qwen1.5-MoE-A2.7B achieve near lossless performance at 25% expert pruning and maintains performance within a 4% margin even at 50% pruning. Moreover, COMPEL consistently outperforms existing pruning methods while substantially reducing inference latency and peak GPU memory usage.

Yufeng Shi, Weilin Luo, Yuxiang Zhang, Zongmeng Zhang, Haoyang Liu, Yubing Wang, Bin Wang, Wengang Zhou, Houqiang Li

While excelling at solving complex problems, Large Reasoning Models (LRMs) are still constrained by the overthinking issue. Most current studies rely on reward shaping in Reinforcement Learning (RL) to shorten the Chain-of-Thought (CoT) of LRMs, remaining sample-inefficient and non-robust due to the absence of guided exploration and prioritized exploitation. To address these issues, we propose a novel policy optimization framework with **S**elf-**I**mitation and self-**G**uidance **M**ech**A**nisms (SIGMA), which reshapes the exploration and exploitation through two core components: (i) **self-imitation exploitation**, which enables the prioritized exploitation of high-value prompts and rollouts by introducing a self-imitated loss and a dynamic sampling strategy based on compression rate; (ii) **self-guidance exploration**, which provides a preference-aware exploration guidance through diverse and pluggable self-rewriting strategies. Experiments across various datasets indicate that our method achieves superior reasoning efficiency without compromising, and even facilitating, the overall accuracy. Furthermore, ablation studies show that the proposed mechanisms can provide flexible control interfaces for the tradeoff between the reasoning accuracy and efficiency of LRMs.

Aoqiang Zhu, Min Hu, Yan Xing

Most existing Multimodal Sentiment Analysis (MSA) methods rely on holistic fusion, treating all modalities and temporal segments equally. Such strategies often introduce redundant information and obscure the decision process, limiting both robustness and interpretability. Inspired by dual-process theory, we propose FRSR (Fast Retrieval and Slow Reasoning), an interpretable framework that decomposes multimodal sentiment modeling into two cooperative pathways. The Fast Pathway acts as a lightweight evidence selector, using context-aware convolution and auxiliary supervision to retrieve a sparse set of Top-K sentiment-relevant cues from noisy multimodal inputs. Based on these cues, the Slow Pathway performs deeper cross-modal reasoning through learnable reasoning tokens, enabling hierarchical sentiment inference. By separating salient evidence retrieval from multimodal reasoning, FRSR improves interpretability while reducing computational cost. Experiments on three benchmark datasets show that FRSR achieves competitive performance, higher efficiency, stronger robustness to noise, and clearer decision transparency than existing holistic fusion methods.

Yangqin Jiang, Chao Huang

With the advancement of multimodal large language models (MLLMs), building GUI agent systems has become an increasingly promising direction—especially for mobile platforms, given their rich app ecosystems and intuitive touch interactions. Yet mobile GUI agents face a critical dilemma: truly on-device models (4B or smaller) lack sufficient performance, while capable models (starting from 7B) are either too large for mobile deployment or prohibitively costly (e.g., cloud-only closed-source MLLMs). To resolve this, we propose OpenPhone, a mobile GUI agent system that leverages device-cloud collaboration to tap the cost-efficiency of on-device models and the high capability of cloud models, while avoiding their drawbacks. Specifically, OpenPhone enhances Qwen2.5-VL-3B via two-stage SFT→GRPO training on synthetic GUI data for strong decision-making, integrates an efficient long-reasoning mechanism to utilize historical interactions under tight resources, and defaults to on-device execution—only escalating challenging subtasks to the cloud via real-time complexity assessment. Experiments on the online AndroidLab benchmark and diverse apps show OpenPhone matches or nears larger models, with a significant reduction in cloud costs.

Shihao Zou, Wei Wei, Yongshuo Zhang

Multimodal Sentiment Analysis (MSA) often suffers from performance degradation due to missing modalities in practical applications. Existing methods typically focus on feature completion but neglect semantic shifts caused by distribution gaps and decision risks under high uncertainty. In this paper, we propose a Distributional Error-Aware Reliability (DEAR) estimation framework for robust MSA. Specifically, we design a Hierarchical Distribution-Constrained Reconstruction (HDCR) module to mitigate semantic shifts by explicitly aligning reconstructed features with the original distributional manifold. Meanwhile, a reliability evaluation module (SURE) is introduced to quantitatively measure reconstruction fidelity. By perceiving inherent uncertainty, SURE provides a reliability-driven gating mechanism for the Synergistic-Robust Dual-Stream (SR-DS) architecture. This mechanism enables the model to dynamically adjust contribution weights: strengthening cross-modal synergistic effects when data fidelity is high, while shifting focus toward robust paths under high-risk missingness to safeguard performance. Extensive experiments on MOSI, MOSEI, and SIMS datasets validate the effectiveness and decision reliability of DEAR.

Liwei Zheng, Xuemin Liu, Jie Liu

Retrieval-Augmented Generation (RAG) systems have become a standard approach for grounding large language models in external knowledge. However, they are constrained by a decoupled architecture: retrieval and reasoning operate as separate stages, with retrieved text merely prepended as passive context. This prevents deep integration of knowledge into the model’s parametric reasoning, leading to fragmented responses for complex queries requiring multi-document synthesis or conflict resolution. To bridge this gap, we propose NeuRAG, an end-to-end Neuralized RAG framework that unifies knowledge retrieval and fusion through Hyper-Neurons—parameterized modules encoding entire documents directly into the model’s parameter space. In NeuRAG, each document is encoded as a lightweight LoRA module, conceptualized as a knowledge neuron. These neurons collectively form a document-adaptive Hyper-Layer, which dynamically activates and fuses knowledge neurons via an attention mechanism conditioned on the input hidden-state query. This enables the model to jointly retrieve and reason within a single forward pass, seamlessly integrating external knowledge into its inference pathway. Extensive experiments across multiple datasets and LLMs demonstrate NeuRAG’s strong and consistent performance as a promising novel RAG paradigm.

Yingxu Li, Jingjie Zeng, Zekun Wang, Hongfei Lin, Liang Yang

Conventional Euclidean geometries lead to structural distortion and entangle core pharmacophoric identities with peripheral groups. Existing molecule-language models, relying on linear or uniform encodings, often obscure the hierarchical organization of chemical semantics. To address this, we propose Geometric-Language Alignment (GLA), a framework integrating intrinsic molecular topology into large language models. GLA employs a mixed-curvature encoder that adaptively learns geometric representations through a gating mechanism. These representations are aligned with text via a dual-view contrastive objective and injected into a frozen language model. Experiments on cross-modal retrieval, captioning, and property prediction benchmarks show GLA consistently improves performance over baselines, suggesting that modeling geometric heterogeneity enhances the grounding between molecular structure and chemical language.

Mukul Ranjan, Prince Jha, Khushboo Kumari, Zhiqiang Shen

Vision-Language Models (VLMs) are increasingly applied to cultural heritage materials, from digital archives to educational platforms. This work identifies a fundamental issue in how these models interpret historical artifacts. We define this phenomenon as cultural anachronism, the tendency to misinterpret historical objects using temporally inappropriate concepts, materials, or cultural frameworks. To quantify this phenomenon, we introduce the Temporal Anachronism Benchmark for Vision-Language Models TAB-VLM, a dataset of 600 questions across six categories, designed to evaluate temporal reasoning on 1,600 Indian cultural artifacts spanning prehistoric to modern periods. Systematic evaluations of ten state-of-the-art models reveal significant deficiencies on our benchmark, and even the best model (GPT-5.2) achieves only 58.7% overall accuracy. The performance gap persists across varying architectures and scales, suggesting that cultural anachronism represents a significant limitation in visual AI systems, regardless of model size. These findings highlight the disparity between current VLM capabilities and the requirements for accurately interpreting cultural heritage materials, particularly for non-Western visual cultures underrepresented in training data. Our benchmark provides a foundation for enhancing temporal cognition in multimodal AI systems that interact with historical artifacts. The dataset and code are available in the supplementary material.

Jaeyoung Kim, Jongho Kim, Seung-won Hwang, Seoho Song, Young-In Song

Retrieval-augmented generation systems often suffer from a gap between optimizing retrieval relevance and generative utility. With such a gap, retrieved documents may be topically relevant but still lack the content needed for effective reasoning during generation. While existing bridge modules attempt to rewrite the retrieved text for better generation, we show how they fail by not capturing "document utility". In this work, we propose R2U, with a key distinction of approximating true utility through joint observation of rewriting and answering in the reasoning process. To distill this observation reliably, R2U scales such supervision to enhance reliability in distillation. We further construct utility-improvement supervision by measuring the generator’s gain of the answer under the rewritten context, yielding signals for fine-tuning and preference optimization. We evaluate our method across multiple open-domain question-answering benchmarks. The empirical results demonstrate consistent improvements over strong bridging baselines.

Sunghwan Steve Cho, Yunseok Han, Jaeyoung Do

Longitudinal chest X-ray (CXR) interpretation requires reasoning over disease evolution across multiple patient visits, yet most existing medical VQA benchmarks focus on single images or short-horizon image pairs. We introduce **MI-CXR**, a benchmark for standardized evaluation of **M**ulti-**I**nterval longitudinal reasoning over multi-visit **CXR** sequences, without requiring free-form report generation or additional clinical context. MI-CXR comprises five-way multiple-choice questions over five-visit patient timelines and instantiates three complementary task families: Temporal Event Localization, Interval-wise Change Reasoning, and Global Trajectory Summarization, which assess clinically grounded visual reasoning over time. Evaluating 14 state-of-the-art vision–language models (VLMs) shows low overall performance (29.3% accuracy), only modestly above random guessing. Using stage-wise diagnostic probing, we find that models often produce locally plausible interval descriptions but fail to enforce temporal constraints or compose evidence into globally consistent decisions over the full timeline. These findings reveal key limitations of current VLMs and establish MI-CXR as a principled benchmark for longitudinal medical reasoning. The benchmark is available at: https://github.com/AIDASLab/MI-CXR

Haowen Li, Yoichi Ishibashi, Masafumi Oyamada

Peer review is an essential process in scientific research, yet the growing workload has made its automation increasingly necessary. In this study, we analyze how different types of reviewer guidelines, such as official conference guidelines and reviewer-imitating ones distilled from high-quality human reviews, affect automated peer review. Our experiments show that official conference guidelines produce review results most consistent with human judgments, suggesting that evaluation criteria refined through conference practice serve as effective guidance for automated reviewing as well. In contrast, reviewer-imitating guidelines, especially those enforcing strict rubric-style scoring, consistently degraded automated review performance, highlighting the importance of allowing subjective and holistic scoring.

Jiechao Gao, Rohan Kumar Yadav, Yuangang Li, Yuandong Pan, Jie Wang, Ying Liu, Michael Lepech

Pretrained language models (PLMs) like BERT provide strong semantic representations but are costly and opaque, while symbolic models such as the Tsetlin Machine (TM) offer transparency but lack semantic generalization. We propose a semantic bootstrapping framework that transfers LLM knowledge into symbolic form, combining interpretability with semantic capacity. Given a class label, an LLM generates sub-intents that guide synthetic data creation through a three-stage curriculum (seed, core, enriched), expanding semantic diversity. A Non-Negated TM (NTM) learns from these examples to extract high-confidence literals as interpretable semantic cues. Injecting these cues into real data enables a TM to align clause logic with LLM-inferred semantics. Our method requires no embeddings or runtime LLM calls, yet equips symbolic models with pretrained semantic priors. Across multiple text classification tasks, it improves interpretability and accuracy over vanilla TM, achieving performance comparable to BERT while remaining fully symbolic and efficient.

Dikshant Kukreja, Kshitij Sah, Gautam Gupta, Avinash Anand, Rajiv Ratn Shah, Zhengkui Wang, Aik Beng Ng, Erik Cambria

Larger language models become simultaneously better and worse at handling contextual information—better at ignoring false claims, worse at ignoring irrelevant tokens. We formalize this apparent paradox through the first scaling laws for contextual entrainment, the tendency of models to favor tokens that appeared in context regardless of relevance. Analyzing the Cerebras-GPT (111M–13B) and Pythia (14M–12B) model families, we find entrainment follows predictable power-law scaling, but with opposite trends depending on context type: semantic contexts show decreasing entrainment with scale, while non-semantic contexts show increasing entrainment. Concretely, the largest models are four times more resistant to counterfactual misinformation than the smallest, yet simultaneously twice as prone to copying arbitrary tokens. These diverging trends, which replicate across model families, suggest that semantic filtering and mechanical copying are functionally distinct behaviors that scale in opposition. These opposing trends suggest that scaling alone does not resolve context sensitivity—it reshapes it.

Haozhe Zhao, Zefan Cai, Shuzheng Si, Liang Chen, Jiuxiang Gu, Wen Xiao, Minjia Zhang, Junjie Hu

Recent text-to-image models achieve impressive visual quality but still face challenges in precise controllability, balancing multimodal inputs, and high training cost for multimodal image generation. To address these limitations, we propose MENTOR, an autoregressive (AR) framework with a two-stage training paradigm for controllable multimodal image generation: (1) a multimodal alignment stage that establishes robust pixel and semantic-level alignment between inputs and generated tokens, followed by (2) a multimodal instruction tuning stage that balance model’s integration of multimodal inputs and enhance generation controllability. Extensive experiments on DreamBench++ and DreamBench demonstrate that, despite modest model size and training resources, achieves a strong balance between textual and visual guidance for controllable image generation, delivering competitive performance at significantly lower computational cost compared to leading baselines. Moreover, our approach attains superior image reconstruction fidelity, broad adaptability across different tasks, and training efficiency.

Haoyang Chen, Yi Liu, Jianzhi Shao, Tao Zhang, Chengfu Huo, Wei Hu

Thinking LLMs produce reasoning traces before answering. Prior activation steering work mainly targets on shaping these traces. It remains less understood how answer tokens actually read and integrate the reasoning to produce reliable outcomes. Focusing on quantitative reasoning, we analyze the answer-to-reasoning attention and observe a benign self-reading pattern aligned with correctness, characterized by a forward drift of the reading focus along the reasoning trace and a persistent concentration on key semantic anchors, whereas incorrect solutions exhibit diffuse and irregular attention pattern. We interpret this as internal certainty during answer decoding, where the model commits to a viable solution branch and integrates key evidence. Following this, we propose a training-free steering method driven by Self-Reading Quality (SRQ) scores combining geometric metrics for process control with semantic metrics for content monitoring. SRQ selects data to build steering vectors that guide inference toward benign self-reading and away from uncertain and disorganized reading. Experiments show that our method yields consistent accuracy gains.

Jialin Li, Zhenhao Chen, Hanjun Luo, Hanan Salam

LLM-based agents can complete tasks correctly yet still frustrate users through poor interaction patterns, such as excessive confirmations, opaque reasoning, or misaligned pacing. Current benchmarks evaluate task accuracy but overlook how agents interact: whether they infer preferences from implicit cues, adapt dynamically, or maintain fine-grained interaction quality. We introduce , a configurable environment that evaluates both what agents accomplish and how they interact. Central to  is the Interaction-as-a-Tool (IaaT) paradigm, which treats interaction behaviors as structured tool calls, unifying them with existing evaluation frameworks. We define 31 preference settings across 14 attributes and formalize user experience (UX) as a core metric alongside task accuracy. A composite LLM-as-a-Judge mechanism across seven UX dimensions achieves strong aggregate reliability (ICC > 0.79), high internal consistency (\alpha = 0.943), and human correlation (\rho = 0.52-0.78). Preference-aware agents show 7.6% average UX improvement and 18.5% gain in preference alignment.

Daocheng Fu, Jianbiao Mei, Rong Wu, Xuemeng Yang, Jia Xu, Ding Wang, Pinlong Cai, Yong Liu, Licheng Wen, Botian Shi

The rapid evolution of Multi-modal Large Language Models (MLLMs) has advanced workflow automation; however, existing research mainly targets performance upper bounds in static environments, overlooking robustness for stochastic real-world deployment. We identify three key challenges: dynamic task scheduling, active exploration under uncertainty, and continuous learning from experience. To bridge this gap, we introduce TraineeBench, a dynamic evaluation environment that simulates a "trainee" agent continuously exploring a novel setting. Unlike traditional benchmarks, TraineeBench evaluates agents along three dimensions: (1) context-aware scheduling for streaming tasks with varying priorities; (2) prudent information acquisition to reduce hallucination via active exploration; and (3) continuous evolution by distilling generalized strategies from rule-based, dynamically generated tasks. Experiments show that cutting-edge agents have significant deficiencies in dynamic environments, especially in active exploration and continual learning. Our work establishes a framework for assessing agent reliability, shifting evaluation from static tests to realistic, production-oriented scenarios.

Zijian Li, Xiachong Feng, Weitao Ma, Yichong Huang, Xiaocheng Feng, Bing Qin

Merging a large number of low-rank adaptations (LoRAs) is a key technology for enhancing the integration and deployment efficiency of large language models (LLMs). However, current general model merging methods are prone to “parameter interference” problem, and this issue is especially pronounced when merging high-rank LoRAs, where parameter conflicts tend to be more severe. While the classical rotation alignment approach can enhance robustness, it is difficult to apply due to incompatibility with the LoRA structure and its high computational complexity. To address these challenges, we propose a novel two-stage parameter alignment (TSPA) framework. TSPA is designed from the perspective of the LoRA architecture, overcoming the limitations of existing methods and reducing the computational complexity from quadratic to linear. We conduct experiments on Natural Language Processing (NLP) tasks using models such as Llama-3-8B. The results show that the two-stage design of TSPA achieves a balance between task capabilities and general knowledge. It exhibits greater robustness than other methods in high-rank and high-interference scenarios, while effectively preserving fine-grained functions.