# Research Papers — 2025-12 · bundle page 37 of 175 > Full-text markdown of papers #1801–1850 of 8703 published in 2025-12, newest first. AI-readable mirrors of the full 8703-paper 2025-12 archive (50 papers per file). Canonical page: https://aipod.dev/papers/archive/2025-12/page/37 [← Previous bundle](https://aipod.dev/papers/archive/2025-12/page/36.md) [Next bundle →](https://aipod.dev/papers/archive/2025-12/page/38.md) ### Unified Multimodal Brain Decoding via Cross-Subject Soft-ROI Fusion Authors: Xuanyu Hu Published: 2025-12-23 Categories: cs.LG, cs.CV, eess.IV arXiv: https://arxiv.org/abs/2512.20249v3 Multimodal brain decoding aims to reconstruct semantic information that is consistent with visual stimuli from brain activity signals such as fMRI, and then generate readable natural language descriptions. However, multimodal brain decoding still faces key challenges in cross-subject generalization and interpretability. We propose a BrainROI model and achieve leading-level results in brain-captioning evaluation on the NSD dataset. Under the cross-subject setting, compared with recent state-of-the-art methods and representative baselines, metrics such as BLEU-4 and CIDEr show clear improvements. Firstly, to address the heterogeneity of functional brain topology across subjects, we design a new fMRI encoder. We use multi-atlas soft functional parcellations (soft-ROI) as a shared space. We extend the discrete ROI Concatenation strategy in MINDLLM to a voxel-wise gated fusion mechanism (Voxel-gate). We also ensure consistent ROI mapping through global label alignment, which enhances cross-subject transferability. Secondly, to overcome the limitations of manual and black-box prompting methods in stability and transparency, we introduce an interpretable prompt optimization process. In a small-sample closed loop, we use a locally deployed Qwen model to iteratively generate and select human-readable prompts. This process improves the stability of prompt design and preserves an auditable optimization trajectory. Finally, we impose parameterized decoding constraints during inference to further improve the stability and quality of the generated descriptions. --- ### Unified Multimodal Brain Decoding via Cross-Subject Soft-ROI Fusion Authors: Xuanyu Hu Published: 2025-12-23 Categories: cs.LG, cs.CV, eess.IV arXiv: https://arxiv.org/abs/2512.20249v1 Multimodal brain decoding aims to reconstruct semantic information that is consistent with visual stimuli from brain activity signals such as fMRI, and then generate readable natural language descriptions. However, multimodal brain decoding still faces key challenges in cross-subject generalization and interpretability. We propose a BrainROI model and achieve leading-level results in brain-captioning evaluation on the NSD dataset. Under the cross-subject setting, compared with recent state-of-the-art methods and representative baselines, metrics such as BLEU-4 and CIDEr show clear improvements. Firstly, to address the heterogeneity of functional brain topology across subjects, we design a new fMRI encoder. We use multi-atlas soft functional parcellations (soft-ROI) as a shared space. We extend the discrete ROI Concatenation strategy in MINDLLM to a voxel-wise gated fusion mechanism (Voxel-gate). We also ensure consistent ROI mapping through global label alignment, which enhances cross-subject transferability. Secondly, to overcome the limitations of manual and black-box prompting methods in stability and transparency, we introduce an interpretable prompt optimization process. In a small-sample closed loop, we use a locally deployed Qwen model to iteratively generate and select human-readable prompts. This process improves the stability of prompt design and preserves an auditable optimization trajectory. Finally, we impose parameterized decoding constraints during inference to further improve the stability and quality of the generated descriptions. --- ### Memory as Resonance: A Biomimetic Architecture for Infinite Context Memory on Ergodic Phonetic Manifolds Authors: Tarik Houichime, Abdelghani Souhar, Younes El Amrani Published: 2025-12-23 Categories: cs.NE, cs.AI, cs.IR, cs.SC, cs.SE arXiv: https://arxiv.org/abs/2512.20245v2 The memory of contemporary Large Language Models is bound by a physical paradox: as they learn, they fill up. The linear accumulation (O(N)) of Key-Value states treats context as a warehouse of static artifacts, eventually forcing a destructive choice between amnesia and latency. We challenge this discrete orthodoxy, proposing that long-term memory is not the storage of items, but the persistence of a trajectory. We introduce Phonetic Trajectory Memory (PTM), a neuro-symbolic architecture that encodes language not as a sequence of tensors, but as a continuous path on an ergodic manifold governed by irrational rotation matrices. By decoupling the navigation (an invariant O(1) geometric signal) from the reconstruction (a probabilistic generative act), PTM achieves a compression magnitude of greater than 3,000x relative to dense caches. We demonstrate that retrieval becomes a process of resonance: the phonetic trace stabilizes the model against hallucination via "Signal Consensus" mechanism, securing up to approximately 92% factual accuracy. While this aggressive abstraction alters generative texture, it unlocks immediate access latency (approximately 34ms) independent of depth. Our results suggest that infinite context does not require infinite silicon; it requires treating memory not as data to be stored, but as a reconstructive process acting on a conserved, undying physical signal. --- ### Memory as Resonance: A Biomimetic Architecture for Infinite Context Memory on Ergodic Phonetic Manifolds Authors: Tarik Houichime, Abdelghani Souhar, Younes El Amrani Published: 2025-12-23 Categories: cs.NE, cs.AI, cs.IR, cs.SC, cs.SE arXiv: https://arxiv.org/abs/2512.20245v1 The memory of contemporary Large Language Models is bound by a physical paradox: as they learn, they fill up. The linear accumulation (O(N)) of Key-Value states treats context as a warehouse of static artifacts, eventually forcing a destructive choice between amnesia and latency. We challenge this discrete orthodoxy, proposing that long-term memory is not the storage of items, but the persistence of a trajectory. We introduce Phonetic Trajectory Memory (PTM), a neuro-symbolic architecture that encodes language not as a sequence of tensors, but as a continuous path on an ergodic manifold governed by irrational rotation matrices. By decoupling the navigation (an invariant O(1) geometric signal) from the reconstruction (a probabilistic generative act), PTM achieves a compression magnitude of greater than 3,000x relative to dense caches. We demonstrate that retrieval becomes a process of resonance: the phonetic trace stabilizes the model against hallucination via "Signal Consensus" mechanism, securing up to approximately 92% factual accuracy. While this aggressive abstraction alters generative texture, it unlocks immediate access latency (approximately 34ms) independent of depth. Our results suggest that infinite context does not require infinite silicon; it requires treating memory not as data to be stored, but as a reconstructive process acting on a conserved, undying physical signal. --- ### MemR$^3$: Memory Retrieval via Reflective Reasoning for LLM Agents Authors: Xingbo Du, Loka Li, Duzhen Zhang, Le Song Published: 2025-12-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2512.20237v1 Memory systems have been designed to leverage past experiences in Large Language Model (LLM) agents. However, many deployed memory systems primarily optimize compression and storage, with comparatively less emphasis on explicit, closed-loop control of memory retrieval. From this observation, we build memory retrieval as an autonomous, accurate, and compatible agent system, named MemR$^3$, which has two core mechanisms: 1) a router that selects among retrieve, reflect, and answer actions to optimize answer quality; 2) a global evidence-gap tracker that explicitly renders the answering process transparent and tracks the evidence collection process. This design departs from the standard retrieve-then-answer pipeline by introducing a closed-loop control mechanism that enables autonomous decision-making. Empirical results on the LoCoMo benchmark demonstrate that MemR$^3$ surpasses strong baselines on LLM-as-a-Judge score, and particularly, it improves existing retrievers across four categories with an overall improvement on RAG (+7.29%) and Zep (+1.94%) using GPT-4.1-mini backend, offering a plug-and-play controller for existing memory stores. --- ### IndicDLP: A Foundational Dataset for Multi-Lingual and Multi-Domain Document Layout Parsing Authors: Oikantik Nath, Sahithi Kukkala, Mitesh Khapra, Ravi Kiran Sarvadevabhatla Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20236v1 Document layout analysis is essential for downstream tasks such as information retrieval, extraction, OCR, and digitization. However, existing large-scale datasets like PubLayNet and DocBank lack fine-grained region labels and multilingual diversity, making them insufficient for representing complex document layouts. In contrast, human-annotated datasets such as M6Doc and D4LA offer richer labels and greater domain diversity, but are too small to train robust models and lack adequate multilingual coverage. This gap is especially pronounced for Indic documents, which encompass diverse scripts yet remain underrepresented in current datasets, further limiting progress in this space. To address these shortcomings, we introduce IndicDLP, a large-scale foundational document layout dataset spanning 11 representative Indic languages alongside English and 12 common document domains. Additionally, we curate UED-mini, a dataset derived from DocLayNet and M6Doc, to enhance pretraining and provide a solid foundation for Indic layout models. Our experiments demonstrate that fine-tuning existing English models on IndicDLP significantly boosts performance, validating its effectiveness. Moreover, models trained on IndicDLP generalize well beyond Indic layouts, making it a valuable resource for document digitization. This work bridges gaps in scale, diversity, and annotation granularity, driving inclusive and efficient document understanding. --- ### How I Met Your Bias: Investigating Bias Amplification in Diffusion Models Authors: Nathan Roos, Ekaterina Iakovleva, Ani Gjergji, Vito Paolo Pastore, Enzo Tartaglione Published: 2025-12-23 Categories: cs.LG, cs.CV arXiv: https://arxiv.org/abs/2512.20233v1 Diffusion-based generative models demonstrate state-of-the-art performance across various image synthesis tasks, yet their tendency to replicate and amplify dataset biases remains poorly understood. Although previous research has viewed bias amplification as an inherent characteristic of diffusion models, this work provides the first analysis of how sampling algorithms and their hyperparameters influence bias amplification. We empirically demonstrate that samplers for diffusion models -- commonly optimized for sample quality and speed -- have a significant and measurable effect on bias amplification. Through controlled studies with models trained on Biased MNIST, Multi-Color MNIST and BFFHQ, and with Stable Diffusion, we show that sampling hyperparameters can induce both bias reduction and amplification, even when the trained model is fixed. Source code is available at https://github.com/How-I-met-your-bias/how_i_met_your_bias. --- ### Adaptive Multi-task Learning for Probabilistic Load Forecasting Authors: Onintze Zaballa, Verónica Álvarez, Santiago Mazuelas Published: 2025-12-23 Categories: cs.LG, stat.AP arXiv: https://arxiv.org/abs/2512.20232v2 Simultaneous load forecasting across multiple entities (e.g., regions, buildings) is crucial for the efficient, reliable, and cost-effective operation of power systems. Accurate load forecasting is a challenging problem due to the inherent uncertainties in load demand, dynamic changes in consumption patterns, and correlations among entities. Multi-task learning has emerged as a powerful machine learning approach that enables the simultaneous learning across multiple related problems. However, its application to load forecasting remains underexplored and is limited to offline learning methods, which cannot capture changes in consumption patterns. This paper presents an adaptive multi-task learning method for probabilistic load forecasting. The proposed method can dynamically adapt to changes in consumption patterns and correlations among entities. In addition, the techniques presented provide reliable probabilistic predictions for loads of multiple entities and assess load uncertainties. Specifically, the method is based on vectorvalued hidden Markov models and uses a recursive process to update the model parameters and provide predictions with the most recent parameters. The performance of the proposed method is evaluated using datasets that contain the load demand of multiple entities and exhibit diverse and dynamic consumption patterns. The experimental results show that the presented techniques outperform existing methods both in terms of forecasting performance and uncertainty assessment. --- ### Generalisation in Multitask Fitted Q-Iteration and Offline Q-learning Authors: Kausthubh Manda, Raghuram Bharadwaj Diddigi Published: 2025-12-23 Categories: cs.LG arXiv: https://arxiv.org/abs/2512.20220v2 We study offline multitask reinforcement learning in settings where multiple tasks share a low-rank representation of their action-value functions. In this regime, a learner is provided with fixed datasets collected from several related tasks, without access to further online interaction, and seeks to exploit shared structure to improve statistical efficiency and generalization. We analyze a multitask variant of fitted Q-iteration that jointly learns a shared representation and task-specific value functions via Bellman error minimization on offline data. Under standard realizability and coverage assumptions commonly used in offline reinforcement learning, we establish finite-sample generalization guarantees for the learned value functions. Our analysis explicitly characterizes how pooling data across tasks improves estimation accuracy, yielding a $1/\sqrt{nT}$ dependence on the total number of samples across tasks, while retaining the usual dependence on the horizon and concentrability coefficients arising from distribution shift. In addition, we consider a downstream offline setting in which a new task shares the same underlying representation as the upstream tasks. We study how reusing the representation learned during the multitask phase affects value estimation for this new task, and show that it can reduce the effective complexity of downstream learning relative to learning from scratch. Together, our results clarify the role of shared representations in multitask offline Q-learning and provide theoretical insight into when and how multitask structure can improve generalization in model-free, value-based reinforcement learning. --- ### Cost-TrustFL: Cost-Aware Hierarchical Federated Learning with Lightweight Reputation Evaluation across Multi-Cloud Authors: Jixiao Yang, Jinyu Chen, Zixiao Huang, Chengda Xu, Chi Zhang, Sijia Li Published: 2025-12-23 Categories: cs.LG arXiv: https://arxiv.org/abs/2512.20218v1 Federated learning across multi-cloud environments faces critical challenges, including non-IID data distributions, malicious participant detection, and substantial cross-cloud communication costs (egress fees). Existing Byzantine-robust methods focus primarily on model accuracy while overlooking the economic implications of data transfer across cloud providers. This paper presents Cost-TrustFL, a hierarchical federated learning framework that jointly optimizes model performance and communication costs while providing robust defense against poisoning attacks. We propose a gradient-based approximate Shapley value computation method that reduces the complexity from exponential to linear, enabling lightweight reputation evaluation. Our cost-aware aggregation strategy prioritizes intra-cloud communication to minimize expensive cross-cloud data transfers. Experiments on CIFAR-10 and FEMNIST datasets demonstrate that Cost-TrustFL achieves 86.7% accuracy under 30% malicious clients while reducing communication costs by 32% compared to baseline methods. The framework maintains stable performance across varying non-IID degrees and attack intensities, making it practical for real-world multi-cloud deployments. --- ### LiteFusion: Taming 3D Object Detectors from Vision-Based to Multi-Modal with Minimal Adaptation Authors: Xiangxuan Ren, Zhongdao Wang, Pin Tang, Guoqing Wang, Jilai Zheng, Chao Ma Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20217v1 3D object detection is fundamental for safe and robust intelligent transportation systems. Current multi-modal 3D object detectors often rely on complex architectures and training strategies to achieve higher detection accuracy. However, these methods heavily rely on the LiDAR sensor so that they suffer from large performance drops when LiDAR is absent, which compromises the robustness and safety of autonomous systems in practical scenarios. Moreover, existing multi-modal detectors face difficulties in deployment on diverse hardware platforms, such as NPUs and FPGAs, due to their reliance on 3D sparse convolution operators, which are primarily optimized for NVIDIA GPUs. To address these challenges, we reconsider the role of LiDAR in the camera-LiDAR fusion paradigm and introduce a novel multi-modal 3D detector, LiteFusion. Instead of treating LiDAR point clouds as an independent modality with a separate feature extraction backbone, LiteFusion utilizes LiDAR data as a complementary source of geometric information to enhance camera-based detection. This straightforward approach completely eliminates the reliance on a 3D backbone, making the method highly deployment-friendly. Specifically, LiteFusion integrates complementary features from LiDAR points into image features within a quaternion space, where the orthogonal constraints are well-preserved during network training. This helps model domain-specific relations across modalities, yielding a compact cross-modal embedding. Experiments on the nuScenes dataset show that LiteFusion improves the baseline vision-based detector by +20.4% mAP and +19.7% NDS with a minimal increase in parameters (1.1%) without using dedicated LiDAR encoders. Notably, even in the absence of LiDAR input, LiteFusion maintains strong results , highlighting its favorable robustness and effectiveness across diverse fusion paradigms and deployment scenarios. --- ### JDPNet: A Network Based on Joint Degradation Processing for Underwater Image Enhancement Authors: Tao Ye, Hongbin Ren, Chongbing Zhang, Haoran Chen, Xiaosong Li Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20213v1 Given the complexity of underwater environments and the variability of water as a medium, underwater images are inevitably subject to various types of degradation. The degradations present nonlinear coupling rather than simple superposition, which renders the effective processing of such coupled degradations particularly challenging. Most existing methods focus on designing specific branches, modules, or strategies for specific degradations, with little attention paid to the potential information embedded in their coupling. Consequently, they struggle to effectively capture and process the nonlinear interactions of multiple degradations from a bottom-up perspective. To address this issue, we propose JDPNet, a joint degradation processing network, that mines and unifies the potential information inherent in coupled degradations within a unified framework. Specifically, we introduce a joint feature-mining module, along with a probabilistic bootstrap distribution strategy, to facilitate effective mining and unified adjustment of coupled degradation features. Furthermore, to balance color, clarity, and contrast, we design a novel AquaBalanceLoss to guide the network in learning from multiple coupled degradation losses. Experiments on six publicly available underwater datasets, as well as two new datasets constructed in this study, show that JDPNet exhibits state-of-the-art performance while offering a better tradeoff between performance, parameter size, and computational cost. --- ### TongSIM: A General Platform for Simulating Intelligent Machines Authors: Zhe Sun, Kunlun Wu, Chuanjian Fu, Zeming Song, Langyong Shi, Zihe Xue, Bohan Jing, Ying Yang, Xiaomeng Gao, Aijia Li, Tianyu Guo, Huiying Li et al. Published: 2025-12-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2512.20206v1 As artificial intelligence (AI) rapidly advances, especially in multimodal large language models (MLLMs), research focus is shifting from single-modality text processing to the more complex domains of multimodal and embodied AI. Embodied intelligence focuses on training agents within realistic simulated environments, leveraging physical interaction and action feedback rather than conventionally labeled datasets. Yet, most existing simulation platforms remain narrowly designed, each tailored to specific tasks. A versatile, general-purpose training environment that can support everything from low-level embodied navigation to high-level composite activities, such as multi-agent social simulation and human-AI collaboration, remains largely unavailable. To bridge this gap, we introduce TongSIM, a high-fidelity, general-purpose platform for training and evaluating embodied agents. TongSIM offers practical advantages by providing over 100 diverse, multi-room indoor scenarios as well as an open-ended, interaction-rich outdoor town simulation, ensuring broad applicability across research needs. Its comprehensive evaluation framework and benchmarks enable precise assessment of agent capabilities, such as perception, cognition, decision-making, human-robot cooperation, and spatial and social reasoning. With features like customized scenes, task-adaptive fidelity, diverse agent types, and dynamic environmental simulation, TongSIM delivers flexibility and scalability for researchers, serving as a unified platform that accelerates training, evaluation, and advancement toward general embodied intelligence. --- ### Corpus of Cross-lingual Dialogues with Minutes and Detection of Misunderstandings Authors: Marko Čechovič, Natália Komorníková, Dominik Macháček, Ondřej Bojar Published: 2025-12-23 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2512.20204v1 Speech processing and translation technology have the potential to facilitate meetings of individuals who do not share any common language. To evaluate automatic systems for such a task, a versatile and realistic evaluation corpus is needed. Therefore, we create and present a corpus of cross-lingual dialogues between individuals without a common language who were facilitated by automatic simultaneous speech translation. The corpus consists of 5 hours of speech recordings with ASR and gold transcripts in 12 original languages and automatic and corrected translations into English. For the purposes of research into cross-lingual summarization, our corpus also includes written summaries (minutes) of the meetings. Moreover, we propose automatic detection of misunderstandings. For an overview of this task and its complexity, we attempt to quantify misunderstandings in cross-lingual meetings. We annotate misunderstandings manually and also test the ability of current large language models to detect them automatically. The results show that the Gemini model is able to identify text spans with misunderstandings with recall of 77% and precision of 47%. --- ### Generative Latent Coding for Ultra-Low Bitrate Image Compression Authors: Zhaoyang Jia, Jiahao Li, Bin Li, Houqiang Li, Yan Lu Published: 2025-12-23 Categories: cs.CV, eess.IV arXiv: https://arxiv.org/abs/2512.20194v1 Most existing image compression approaches perform transform coding in the pixel space to reduce its spatial redundancy. However, they encounter difficulties in achieving both high-realism and high-fidelity at low bitrate, as the pixel-space distortion may not align with human perception. To address this issue, we introduce a Generative Latent Coding (GLC) architecture, which performs transform coding in the latent space of a generative vector-quantized variational auto-encoder (VQ-VAE), instead of in the pixel space. The generative latent space is characterized by greater sparsity, richer semantic and better alignment with human perception, rendering it advantageous for achieving high-realism and high-fidelity compression. Additionally, we introduce a categorical hyper module to reduce the bit cost of hyper-information, and a code-prediction-based supervision to enhance the semantic consistency. Experiments demonstrate that our GLC maintains high visual quality with less than 0.04 bpp on natural images and less than 0.01 bpp on facial images. On the CLIC2020 test set, we achieve the same FID as MS-ILLM with 45% fewer bits. Furthermore, the powerful generative latent space enables various applications built on our GLC pipeline, such as image restoration and style transfer. The code is available at https://github.com/jzyustc/GLC. --- ### Asynchronous Fast-Slow Vision-Language-Action Policies for Whole-Body Robotic Manipulation Authors: Teqiang Zou, Hongliang Zeng, Yuxuan Nong, Yifan Li, Kehui Liu, Haotian Yang, Xinyang Ling, Xin Li, Lianyang Ma Published: 2025-12-23 Categories: cs.RO, cs.AI arXiv: https://arxiv.org/abs/2512.20188v1 Most Vision-Language-Action (VLA) systems integrate a Vision-Language Model (VLM) for semantic reasoning with an action expert generating continuous action signals, yet both typically run at a single unified frequency. As a result, policy performance is constrained by the low inference speed of large VLMs. This mandatory synchronous execution severely limits control stability and real-time performance in whole-body robotic manipulation, which involves more joints, larger motion spaces, and dynamically changing views. We introduce a truly asynchronous Fast-Slow VLA framework (DuoCore-FS), organizing the system into a fast pathway for high-frequency action generation and a slow pathway for rich VLM reasoning. The system is characterized by two key features. First, a latent representation buffer bridges the slow and fast systems. It stores instruction semantics and action-reasoning representation aligned with the scene-instruction context, providing high-level guidance to the fast pathway. Second, a whole-body action tokenizer provides a compact, unified representation of whole-body actions. Importantly, the VLM and action expert are still jointly trained end-to-end, preserving unified policy learning while enabling asynchronous execution. DuoCore-FS supports a 3B-parameter VLM while achieving 30 Hz whole-body action-chunk generation, approximately three times as fast as prior VLA models with comparable model sizes. Real-world whole-body manipulation experiments demonstrate improved task success rates and significantly enhanced responsiveness compared to synchronous Fast-Slow VLA baselines. The implementation of DuoCore-FS, including training, inference, and deployment, is provided to commercial users by Astribot as part of the Astribot robotic platform. --- ### FaithLens: Detecting and Explaining Faithfulness Hallucination Authors: Shuzheng Si, Qingyi Wang, Haozhe Zhao, Yuzhuo Bai, Guanqiao Chen, Kangyang Luo, Gang Chen, Fanchao Qi, Minjia Zhang, Baobao Chang, Maosong Sun Published: 2025-12-23 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2512.20182v4 Recognizing whether outputs from large language models (LLMs) contain faithfulness hallucination is crucial for real-world applications, e.g., retrieval-augmented generation and summarization. In this paper, we introduce FaithLens, a cost-efficient and effective faithfulness hallucination detection model that can jointly provide binary predictions and corresponding explanations to improve trustworthiness. To achieve this, we first synthesize training data with explanations via advanced LLMs and apply a well-defined data filtering strategy to ensure label correctness, explanation quality, and data diversity. Subsequently, we fine-tune the model on these well-curated training data as a cold start and further optimize it with rule-based reinforcement learning, using rewards for both prediction correctness and explanation quality. Results on 12 diverse tasks show that the 8B-parameter FaithLens outperforms advanced models such as GPT-5.2 and o3. Also, FaithLens can produce high-quality explanations, delivering a distinctive balance of trustworthiness, efficiency, and effectiveness. --- ### NeuralCrop: Combining physics and machine learning for improved crop yield projections Authors: Yunan Lin, Sebastian Bathiany, Maha Badri, Maximilian Gelbrecht, Philipp Hess, Brian Groenke, Jens Heinke, Christoph Müller, Niklas Boers Published: 2025-12-23 Categories: cs.LG, physics.comp-ph arXiv: https://arxiv.org/abs/2512.20177v2 Global gridded crop models (GGCMs) are crucial to project the impacts of climate change on agricultural productivity and assess associated risks for food security. Despite decades of development, state-of-the-art GGCMs retain substantial uncertainties stemming from process representations. Recently, machine learning approaches trained on observational data provide alternatives in crop yield projections. However, these models have not demonstrated improved performance over traditional GGCMs and are not suitable for projecting crop yields under a changing climate due to their poor out-of-distribution generalization. Here we introduce NeuralCrop, a differentiable hybrid GGCM that combines the strengths of an advanced process-based GGCM, resolving important processes explicitly, with data-driven machine learning components. NeuralCrop is first trained to emulate a competitive GGCM before it is fine-tuned on observational data. We show that NeuralCrop produces projections with accuracy comparable to state-of-the-art GGCMs across site-level and large-scale crop simulations. NeuralCrop can accurately project the interannual yield variability in European wheat regions and the US Corn Belt. Capturing yield anomalies is essential for developing adaptation strategies in the context of climate change. NeuralCrop can more accurately reproduce yield anomalies across various climatic conditions, with particularly notable improvements under drought extremes. For large-scale, long-term simulations, our approach is orders of magnitude more computationally efficient. Our results show that end-to-end hybrid crop modelling offers more reliable yield projections that are essential for food risk assessments under climate change and intensifying extreme weather events. --- ### Towards Natural Language-Based Document Image Retrieval: New Dataset and Benchmark Authors: Hao Guo, Xugong Qin, Jun Jie Ou Yang, Peng Zhang, Gangyan Zeng, Yubo Li, Hailun Lin Published: 2025-12-23 Categories: cs.CV, cs.CL, cs.IR arXiv: https://arxiv.org/abs/2512.20174v1 Document image retrieval (DIR) aims to retrieve document images from a gallery according to a given query. Existing DIR methods are primarily based on image queries that retrieve documents within the same coarse semantic category, e.g., newspapers or receipts. However, these methods struggle to effectively retrieve document images in real-world scenarios where textual queries with fine-grained semantics are usually provided. To bridge this gap, we introduce a new Natural Language-based Document Image Retrieval (NL-DIR) benchmark with corresponding evaluation metrics. In this work, natural language descriptions serve as semantically rich queries for the DIR task. The NL-DIR dataset contains 41K authentic document images, each paired with five high-quality, fine-grained semantic queries generated and evaluated through large language models in conjunction with manual verification. We perform zero-shot and fine-tuning evaluations of existing mainstream contrastive vision-language models and OCR-free visual document understanding (VDU) models. A two-stage retrieval method is further investigated for performance improvement while achieving both time and space efficiency. We hope the proposed NL-DIR benchmark can bring new opportunities and facilitate research for the VDU community. Datasets and codes will be publicly available at huggingface.co/datasets/nianbing/NL-DIR. --- ### Offline Safe Policy Optimization From Heterogeneous Feedback Authors: Ze Gong, Pradeep Varakantham, Akshat Kumar Published: 2025-12-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2512.20173v1 Offline Preference-based Reinforcement Learning (PbRL) learns rewards and policies aligned with human preferences without the need for extensive reward engineering and direct interaction with human annotators. However, ensuring safety remains a critical challenge across many domains and tasks. Previous works on safe RL from human feedback (RLHF) first learn reward and cost models from offline data, then use constrained RL to optimize a safe policy. While such an approach works in the contextual bandits settings (LLMs), in long horizon continuous control tasks, errors in rewards and costs accumulate, leading to impairment in performance when used with constrained RL methods. To address these challenges, (a) instead of indirectly learning policies (from rewards and costs), we introduce a framework that learns a policy directly based on pairwise preferences regarding the agent's behavior in terms of rewards, as well as binary labels indicating the safety of trajectory segments; (b) we propose \textsc{PreSa} (Preference and Safety Alignment), a method that combines preference learning module with safety alignment in a constrained optimization problem. This optimization problem is solved within a Lagrangian paradigm that directly learns reward-maximizing safe policy \textit{without explicitly learning reward and cost models}, avoiding the need for constrained RL; (c) we evaluate our approach on continuous control tasks with both synthetic and real human feedback. Empirically, our method successfully learns safe policies with high rewards, outperforming state-of-the-art baselines, and offline safe RL approaches with ground-truth reward and cost. --- ### Learning to Reason in LLMs by Expectation Maximization Authors: Junghyun Lee, Branislav Kveton, Anup Rao, Subhojyoti Mukherjee, Ryan A. Rossi, Sunav Choudhary, Alexa Siu Published: 2025-12-23 Categories: cs.LG, cs.CL, stat.ML arXiv: https://arxiv.org/abs/2512.20169v2 Large language models (LLMs) solve reasoning problems by first generating a rationale and then answering. We formalize reasoning as a latent variable model and derive a reward-based filtered expectation-maximization (FEM) objective for learning to reason. This view connects EM and modern reward-based optimization, and shows that the main challenge lies in designing a sampling distribution of rationales that justify correct answers. We instantiate and compare three sampling schemes: rejection sampling with a budget, self-taught reasoner (STaR), and prompt posterior sampling (PPS), which only keeps the rationalization stage of STaR that conditions on the correct answer in the prompt. We experiment with LLM-as-a-judge calibration and summarization from feedback tasks, where conditioning on the correct answer provides a strong guidance for generating rationales. Our experiments show the efficacy of PPS over other sampling schemes, and that the sampling scheme can have a significant impact on performance. --- ### Odysseus: Jailbreaking Commercial Multimodal LLM-integrated Systems via Dual Steganography Authors: Songze Li, Jiameng Cheng, Yiming Li, Xiaojun Jia, Dacheng Tao Published: 2025-12-23 Categories: cs.CR, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2512.20168v1 By integrating language understanding with perceptual modalities such as images, multimodal large language models (MLLMs) constitute a critical substrate for modern AI systems, particularly intelligent agents operating in open and interactive environments. However, their increasing accessibility also raises heightened risks of misuse, such as generating harmful or unsafe content. To mitigate these risks, alignment techniques are commonly applied to align model behavior with human values. Despite these efforts, recent studies have shown that jailbreak attacks can circumvent alignment and elicit unsafe outputs. Currently, most existing jailbreak methods are tailored for open-source models and exhibit limited effectiveness against commercial MLLM-integrated systems, which often employ additional filters. These filters can detect and prevent malicious input and output content, significantly reducing jailbreak threats. In this paper, we reveal that the success of these safety filters heavily relies on a critical assumption that malicious content must be explicitly visible in either the input or the output. This assumption, while often valid for traditional LLM-integrated systems, breaks down in MLLM-integrated systems, where attackers can leverage multiple modalities to conceal adversarial intent, leading to a false sense of security in existing MLLM-integrated systems. To challenge this assumption, we propose Odysseus, a novel jailbreak paradigm that introduces dual steganography to covertly embed malicious queries and responses into benign-looking images. Extensive experiments on benchmark datasets demonstrate that our Odysseus successfully jailbreaks several pioneering and realistic MLLM-integrated systems, achieving up to 99% attack success rate. It exposes a fundamental blind spot in existing defenses, and calls for rethinking cross-modal security in MLLM-integrated systems. --- ### AI Security Beyond Core Domains: Resume Screening as a Case Study of Adversarial Vulnerabilities in Specialized LLM Applications Authors: Honglin Mu, Jinghao Liu, Kaiyang Wan, Rui Xing, Xiuying Chen, Timothy Baldwin, Wanxiang Che Published: 2025-12-23 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2512.20164v2 Large Language Models (LLMs) excel at text comprehension and generation, making them ideal for automated tasks like code review and content moderation. However, our research identifies a vulnerability: LLMs can be manipulated by "adversarial instructions" hidden in input data, such as resumes or code, causing them to deviate from their intended task. Notably, while defenses may exist for mature domains such as code review, they are often absent in other common applications such as resume screening and peer review. This paper introduces a benchmark to assess this vulnerability in resume screening, revealing attack success rates exceeding 80% for certain attack types. We evaluate two defense mechanisms: prompt-based defenses achieve 10.1% attack reduction with 12.5% false rejection increase, while our proposed FIDS (Foreign Instruction Detection through Separation) using LoRA adaptation achieves 15.4% attack reduction with 10.4% false rejection increase. The combined approach provides 26.3% attack reduction, demonstrating that training-time defenses outperform inference-time mitigations in both security and utility preservation. --- ### Concept Generalization in Humans and Large Language Models: Insights from the Number Game Authors: Arghavan Bazigaran, Hansem Sohn Published: 2025-12-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2512.20162v1 We compare human and large language model (LLM) generalization in the number game, a concept inference task. Using a Bayesian model as an analytical framework, we examined the inductive biases and inference strategies of humans and LLMs. The Bayesian model captured human behavior better than LLMs in that humans flexibly infer rule-based and similarity-based concepts, whereas LLMs rely more on mathematical rules. Humans also demonstrated a few-shot generalization, even from a single example, while LLMs required more samples to generalize. These contrasts highlight the fundamental differences in how humans and LLMs infer and generalize mathematical concepts. --- ### A Bidirectional Gated Recurrent Unit Model for PUE Prediction in Data Centers Authors: Dhivya Dharshini Kannan, Anupam Trivedi, Dipti Srinivasan Published: 2025-12-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2512.20161v1 Data centers account for significant global energy consumption and a carbon footprint. The recent increasing demand for edge computing and AI advancements drives the growth of data center storage capacity. Energy efficiency is a cost-effective way to combat climate change, cut energy costs, improve business competitiveness, and promote IT and environmental sustainability. Thus, optimizing data center energy management is the most important factor in the sustainability of the world. Power Usage Effectiveness (PUE) is used to represent the operational efficiency of the data center. Predicting PUE using Neural Networks provides an understanding of the effect of each feature on energy consumption, thus enabling targeted modifications of those key features to improve energy efficiency. In this paper, we have developed Bidirectional Gated Recurrent Unit (BiGRU) based PUE prediction model and compared the model performance with GRU. The data set comprises 52,560 samples with 117 features using EnergyPlus, simulating a DC in Singapore. Sets of the most relevant features are selected using the Recursive Feature Elimination with Cross-Validation (RFECV) algorithm for different parameter settings. These feature sets are used to find the optimal hyperparameter configuration and train the BiGRU model. The performance of the optimized BiGRU-based PUE prediction model is then compared with that of GRU using mean squared error (MSE), mean absolute error (MAE), and R-squared metrics. --- ### Benchmarking Code LLM-as-a-Judge via Constrained Perturbation and Calibration Authors: Ruiqi Wang, Xinchen Wang, Cuiyun Gao, Chun Yong Chong, Xin Xia, Qing Liao Published: 2025-12-23 Categories: cs.SE, cs.AI arXiv: https://arxiv.org/abs/2512.20159v2 Code from LLM-assisted development may contain defects like API hallucinations, motivating the development of code evaluation metrics to assess such code before deployment. As conventional rule-based metrics merely score code based on similarity to reference programs, and execution-based metrics require expensive manual preparation for tests, researchers have developed LLM-as-a-judge metrics for both behavioral and nonbehavioral code evaluation, and curated multiple benchmarks to validate their effectiveness. However, such benchmarks suffer from critical limitations: Some feature coarse-grained binary labels of passing/failing unit tests, unable to distinguish between fundamental and minor flaws; others utilize fine-grained criteria with manually annotated scores as ground-truth, but such criteria are often vaguely defined, introducing subjectivity and unreliability in human scores. Furthermore, they often directly prompt LLMs to generate code for evaluation, which may lead to unbalanced code quality distributions, failing to test the metrics' discriminative capability across varying quality levels. To curate a diverse benchmark with a balanced program quality distribution and streamline manual annotation, we propose AXIOM, a novel perturbation-based framework. It reframes program scores as the refinement effort needed for deployment, with two stages: (1) Quality-constrained semantic perturbation, prompting LLMs to apply sequences of predefined perturbation rules to existing programs, precisely controlling each program's target quality to achieve a balanced distribution. (2) Hybrid quality calibration, selecting semantically distinct programs for benchmark diversity, and performing streamlined manual annotation with multisource context. With AXIOM, we curate AXIOMBENCH, a large-scale multilingual benchmark with 1,962 code evaluation problems in a more balanced score distribution... --- ### SigLino: Efficient Multi-Teacher Distillation for Agglomerative Vision Foundation Models Authors: Sofian Chaybouti, Sanath Narayan, Yasser Dahou, Phúc H. Lê Khac, Ankit Singh, Ngoc Dung Huynh, Wamiq Reyaz Para, Hilde Kuehne, Hakim Hacid Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20157v2 Vision foundation models trained via multi-teacher distillation offer a promising path toward unified visual representations, yet the learning dynamics and data efficiency of such approaches remain underexplored. In this paper, we systematically study multi-teacher distillation for vision foundation models and identify key factors that enable training at lower computational cost. We introduce SigLino, an efficient family of agglomerative vision foundation models that distill knowledge from SigLIP2 and DINOv3 simultaneously into Dense and Mixture-of-Experts students. We show that (1) our Asymmetric Relation-Knowledge Distillation loss preserves the geometric properties of each teacher while enabling effective knowledge transfer, (2) token-balanced batching that packs varying-resolution images into sequences with uniform token budgets stabilizes representation learning across resolutions without sacrificing performance, (3) hierarchical clustering and sampling of training data, typically reserved for self-supervised learning, substantially improves sample efficiency over random sampling for multi-teacher distillation, and (4) the resulting representations transfer effectively to early-fusion Grounding-VLMs, outperforming models trained from scratch. By combining these findings, we curate OpenLVD200M, a 200M-image corpus that demonstrates superior efficiency for multi-teacher distillation. Instantiated in a Mixture-of-Experts, our SigLino-MoE initializes an early-fusion Grounding-VLM that replaces the conventional ViT->LLM stack, demonstrating improved performance compared to a model trained from scratch. We release OpenLVD200M and five distilled checkpoints comprising MoE and dense variants. --- ### Fun-Audio-Chat Technical Report Authors: Tongyi Fun Team, Qian Chen, Luyao Cheng, Chong Deng, Xiangang Li, Jiaqing Liu, Chao-Hong Tan, Wen Wang, Junhao Xu, Jieping Ye, Qinglin Zhang, Qiquan Zhang et al. Published: 2025-12-23 Categories: cs.CL, cs.AI, cs.SD, eess.AS arXiv: https://arxiv.org/abs/2512.20156v4 Recent advancements in joint speech-text models show great potential for seamless voice interactions. However, existing models face critical challenges: temporal resolution mismatch between speech tokens (25Hz) and text tokens (~3Hz) dilutes semantic information, incurs high computational costs, and causes catastrophic forgetting of text LLM knowledge. We introduce Fun-Audio-Chat, a Large Audio Language Model addressing these limitations via two innovations from our previous work DrVoice. First, Dual-Resolution Speech Representations (DRSR): the Shared LLM processes audio at efficient 5Hz (via token grouping), while the Speech Refined Head generates high-quality tokens at 25Hz, balancing efficiency (~50% GPU reduction) and quality. Second, Core-Cocktail Training, a two-stage fine-tuning with intermediate merging that mitigates catastrophic forgetting. We then apply Multi-Task DPO Training to enhance robustness, audio understanding, instruction-following and voice empathy. This multi-stage post-training enables Fun-Audio-Chat to retain text LLM knowledge while gaining powerful audio understanding, reasoning, and generation. Unlike recent LALMs requiring large-scale audio-text pre-training, Fun-Audio-Chat leverages pre-trained models and extensive post-training. Fun-Audio-Chat 8B and MoE 30B-A3B achieve competitive performance on Speech-to-Text and Speech-to-Speech tasks, ranking top among similar-scale models on Spoken QA benchmarks. They also achieve competitive to superior performance on Audio Understanding, Speech Function Calling, Instruction-Following and Voice Empathy. We develop Fun-Audio-Chat-Duplex, a full-duplex variant with strong performance on Spoken QA and full-duplex interactions. We open-source Fun-Audio-Chat-8B with training and inference code, and provide an interactive demo, at https://github.com/FunAudioLLM/Fun-Audio-Chat . --- ### CoDi -- an exemplar-conditioned diffusion model for low-shot counting Authors: Grega Šuštar, Jer Pelhan, Alan Lukežič, Matej Kristan Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20153v2 Low-shot object counting addresses estimating the number of previously unobserved objects in an image using only few or no annotated test-time exemplars. A considerable challenge for modern low-shot counters are dense regions with small objects. While total counts in such situations are typically well addressed by density-based counters, their usefulness is limited by poor localization capabilities. This is better addressed by point-detection-based counters, which are based on query-based detectors. However, due to limited number of pre-trained queries, they underperform on images with very large numbers of objects, and resort to ad-hoc techniques like upsampling and tiling. We propose CoDi, the first latent diffusion-based low-shot counter that produces high-quality density maps on which object locations can be determined by non-maxima suppression. Our core contribution is the new exemplar-based conditioning module that extracts and adjusts the object prototypes to the intermediate layers of the denoising network, leading to accurate object location estimation. On FSC benchmark, CoDi outperforms state-of-the-art by 15% MAE, 13% MAE and 10% MAE in the few-shot, one-shot, and reference-less scenarios, respectively, and sets a new state-of-the-art on MCAC benchmark by outperforming the top method by 38% MAE. The code is available at https://github.com/gsustar/CoDi --- ### Enhancing annotations for 5D apple pose estimation through 3D Gaussian Splatting (3DGS) Authors: Robert van de Ven, Trim Bresilla, Bram Nelissen, Ard Nieuwenhuizen, Eldert J. van Henten, Gert Kootstra Published: 2025-12-23 Categories: cs.CV, cs.RO arXiv: https://arxiv.org/abs/2512.20148v1 Automating tasks in orchards is challenging because of the large amount of variation in the environment and occlusions. One of the challenges is apple pose estimation, where key points, such as the calyx, are often occluded. Recently developed pose estimation methods no longer rely on these key points, but still require them for annotations, making annotating challenging and time-consuming. Due to the abovementioned occlusions, there can be conflicting and missing annotations of the same fruit between different images. Novel 3D reconstruction methods can be used to simplify annotating and enlarge datasets. We propose a novel pipeline consisting of 3D Gaussian Splatting to reconstruct an orchard scene, simplified annotations, automated projection of the annotations to images, and the training and evaluation of a pose estimation method. Using our pipeline, 105 manual annotations were required to obtain 28,191 training labels, a reduction of 99.6%. Experimental results indicated that training with labels of fruits that are $\leq95\%$ occluded resulted in the best performance, with a neutral F1 score of 0.927 on the original images and 0.970 on the rendered images. Adjusting the size of the training dataset had small effects on the model performance in terms of F1 score and pose estimation accuracy. It was found that the least occluded fruits had the best position estimation, which worsened as the fruits became more occluded. It was also found that the tested pose estimation method was unable to correctly learn the orientation estimation of apples. --- ### Retrieval-augmented Prompt Learning for Pre-trained Foundation Models Authors: Xiang Chen, Yixin Ou, Quan Feng, Lei Li, Piji Li, Haibo Ye, Sheng-Jun Huang, Shuofei Qiao, Shumin Deng, Huajun Chen, Ningyu Zhang Published: 2025-12-23 Categories: cs.CL, cs.AI, cs.CV, cs.IR, cs.LG arXiv: https://arxiv.org/abs/2512.20145v1 The pre-trained foundation models (PFMs) have become essential for facilitating large-scale multimodal learning. Researchers have effectively employed the ``pre-train, prompt, and predict'' paradigm through prompt learning to induce improved few-shot performance. However, prompt learning approaches for PFMs still follow a parametric learning paradigm. As such, the stability of generalization in memorization and rote learning can be compromised. More specifically, conventional prompt learning might face difficulties in fully utilizing atypical instances and avoiding overfitting to shallow patterns with limited data during the process of fully-supervised training. To overcome these constraints, we present our approach, named RetroPrompt, which aims to achieve a balance between memorization and generalization by decoupling knowledge from mere memorization. Unlike traditional prompting methods, RetroPrompt leverages a publicly accessible knowledge base generated from the training data and incorporates a retrieval mechanism throughout the input, training, and inference stages. This enables the model to actively retrieve relevant contextual information from the corpus, thereby enhancing the available cues. We conduct comprehensive experiments on a variety of datasets across natural language processing and computer vision tasks to demonstrate the superior performance of our proposed approach, RetroPrompt, in both zero-shot and few-shot scenarios. Through detailed analysis of memorization patterns, we observe that RetroPrompt effectively reduces the reliance on rote memorization, leading to enhanced generalization. --- ### Multi-hop Reasoning via Early Knowledge Alignment Authors: Yuxin Wang, Shicheng Fang, Bo Wang, Qi Luo, Xuanjing Huang, Yining Zheng, Xipeng Qiu Published: 2025-12-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2512.20144v2 Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for Large Language Models (LLMs) to address knowledge-intensive queries requiring domain-specific or up-to-date information. To handle complex multi-hop questions that are challenging for single-step retrieval, iterative RAG approaches incorporating reinforcement learning have been proposed. However, existing iterative RAG systems typically plan to decompose questions without leveraging information about the available retrieval corpus, leading to inefficient retrieval and reasoning chains that cascade into suboptimal performance. In this paper, we introduce Early Knowledge Alignment (EKA), a simple but effective module that aligns LLMs with retrieval set before planning in iterative RAG systems with contextually relevant retrieved knowledge. Extensive experiments on six standard RAG datasets demonstrate that by establishing a stronger reasoning foundation, EKA significantly improves retrieval precision, reduces cascading errors, and enhances both performance and efficiency. Our analysis from an entropy perspective demonstrate that incorporating early knowledge reduces unnecessary exploration during the reasoning process, enabling the model to focus more effectively on relevant information subsets. Moreover, EKA proves effective as a versatile, training-free inference strategy that scales seamlessly to large models. Generalization tests across diverse datasets and retrieval corpora confirm the robustness of our approach. Overall, EKA advances the state-of-the-art in iterative RAG systems while illuminating the critical interplay between structured reasoning and efficient exploration in reinforcement learning-augmented frameworks. The code is released at \href{https://github.com/yxzwang/EarlyKnowledgeAlignment}{Github}. --- ### Enhancing Zero-Shot Time Series Forecasting in Off-the-Shelf LLMs via Noise Injection Authors: Xingyou Yin, Ceyao Zhang, Min Hu, Kai Chen Published: 2025-12-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2512.20140v1 Large Language Models (LLMs) have demonstrated effectiveness as zero-shot time series (TS) forecasters. The key challenge lies in tokenizing TS data into textual representations that align with LLMs' pre-trained knowledge. While existing work often relies on fine-tuning specialized modules to bridge this gap, a distinct, yet challenging, paradigm aims to leverage truly off-the-shelf LLMs without any fine-tuning whatsoever, relying solely on strategic tokenization of numerical sequences. The performance of these fully frozen models is acutely sensitive to the textual representation of the input data, as their parameters cannot adapt to distribution shifts. In this paper, we introduce a simple yet highly effective strategy to overcome this brittleness: injecting noise into the raw time series before tokenization. This non-invasive intervention acts as a form of inference-time augmentation, compelling the frozen LLM to extrapolate based on robust underlying temporal patterns rather than superficial numerical artifacts. We theoretically analyze this phenomenon and empirically validate its effectiveness across diverse benchmarks. Notably, to fully eliminate potential biases from data contamination during LLM pre-training, we introduce two novel TS datasets that fall outside all utilized LLMs' pre-training scopes, and consistently observe improved performance. This study provides a further step in directly leveraging off-the-shelf LLMs for time series forecasting. --- ### M$^3$KG-RAG: Multi-hop Multimodal Knowledge Graph-enhanced Retrieval-Augmented Generation Authors: Hyeongcheol Park, Jiyoung Seo, Jaewon Mun, Hogun Park, Wonmin Byeon, Sung June Kim, Hyeonsoo Im, JeungSub Lee, Sangpil Kim Published: 2025-12-23 Categories: cs.CL, cs.AI arXiv: https://arxiv.org/abs/2512.20136v3 Retrieval-Augmented Generation (RAG) has recently been extended to multimodal settings, connecting multimodal large language models (MLLMs) with vast corpora of external knowledge such as multimodal knowledge graphs (MMKGs). Despite their recent success, multimodal RAG in the audio-visual domain remains challenging due to 1) limited modality coverage and multi-hop connectivity of existing MMKGs, and 2) retrieval based solely on similarity in a shared multimodal embedding space, which fails to filter out off-topic or redundant knowledge. To address these limitations, we propose M$^3$KG-RAG, a Multi-hop Multimodal Knowledge Graph-enhanced RAG that retrieves query-aligned audio-visual knowledge from MMKGs, improving reasoning depth and answer faithfulness in MLLMs. Specifically, we devise a lightweight multi-agent pipeline to construct multi-hop MMKG (M$^3$KG), which contains context-enriched triplets of multimodal entities, enabling modality-wise retrieval based on input queries. Furthermore, we introduce GRASP (Grounded Retrieval And Selective Pruning), which ensures precise entity grounding to the query, evaluates answer-supporting relevance, and prunes redundant context to retain only knowledge essential for response generation. Extensive experiments across diverse multimodal benchmarks demonstrate that M$^3$KG-RAG significantly enhances MLLMs' multimodal reasoning and grounding over existing approaches. Project website: https://kuai-lab.github.io/cvpr2026m3kgrag/ --- ### MolAct: An Agentic RL Framework for Molecular Editing and Property Optimization Authors: Zhuo Yang, Yeyun Chen, Jiaqing Xie, Ben Gao, Shuaike Shen, Wanhao Liu, Liujia Yang, Beilun Wang, Tianfan Fu, Yuqiang Li Published: 2025-12-23 Categories: cs.AI arXiv: https://arxiv.org/abs/2512.20135v2 Molecular editing and optimization are multi-step problems that require iteratively improving properties while keeping molecules chemically valid and structurally similar. We frame both tasks as sequential, tool-guided decisions and introduce MolAct, an agentic reinforcement learning framework that employs a two-stage training paradigm: first building editing capability, then optimizing properties while reusing the learned editing behaviors. To the best of our knowledge, this is the first work to formalize molecular design as an Agentic Reinforcement Learning problem, where an LLM agent learns to interleave reasoning, tool-use, and molecular optimization. The framework enables agents to interact in multiple turns, invoking chemical tools for validity checking, property assessment, and similarity control, and leverages their feedback to refine subsequent edits. We instantiate the MolAct framework to train two model families: MolEditAgent for molecular editing tasks and MolOptAgent for molecular optimization tasks. In molecular editing, MolEditAgent-7B delivers 100, 95, and 98 valid add, delete, and substitute edits, outperforming strong closed "thinking" baselines such as DeepSeek-R1; MolEditAgent-3B approaches the performance of much larger open "thinking" models like Qwen3-32B-think. In molecular optimization, MolOptAgent-7B (trained on MolEditAgent-7B) surpasses the best closed "thinking" baseline (e.g., Claude 3.7) on LogP and remains competitive on solubility, while maintaining balanced performance across other objectives. These results highlight that treating molecular design as a multi-step, tool-augmented process is key to reliable and interpretable improvements. --- ### Dreamcrafter: Immersive Editing of 3D Radiance Fields Through Flexible, Generative Inputs and Outputs Authors: Cyrus Vachha, Yixiao Kang, Zach Dive, Ashwat Chidambaram, Anik Gupta, Eunice Jun, Bjoern Hartmann Published: 2025-12-23 Categories: cs.HC, cs.CV arXiv: https://arxiv.org/abs/2512.20129v2 Authoring 3D scenes is a central task for spatial computing applications. Competing visions for lowering existing barriers are (1) focus on immersive, direct manipulation of 3D content or (2) leverage AI techniques that capture real scenes (3D Radiance Fields such as, NeRFs, 3D Gaussian Splatting) and modify them at a higher level of abstraction, at the cost of high latency. We unify the complementary strengths of these approaches and investigate how to integrate generative AI advances into real-time, immersive 3D Radiance Field editing. We introduce Dreamcrafter, a VR-based 3D scene editing system that: (1) provides a modular architecture to integrate generative AI algorithms; (2) combines different levels of control for creating objects, including natural language and direct manipulation; and (3) introduces proxy representations that support interaction during high-latency operations. We contribute empirical findings on control preferences and discuss how generative AI interfaces beyond text input enhance creativity in scene editing and world building. --- ### milliMamba: Specular-Aware Human Pose Estimation via Dual mmWave Radar with Multi-Frame Mamba Fusion Authors: Niraj Prakash Kini, Shiau-Rung Tsai, Guan-Hsun Lin, Wen-Hsiao Peng, Ching-Wen Ma, Jenq-Neng Hwang Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20128v1 Millimeter-wave radar offers a privacy-preserving and lighting-invariant alternative to RGB sensors for Human Pose Estimation (HPE) task. However, the radar signals are often sparse due to specular reflection, making the extraction of robust features from radar signals highly challenging. To address this, we present milliMamba, a radar-based 2D human pose estimation framework that jointly models spatio-temporal dependencies across both the feature extraction and decoding stages. Specifically, given the high dimensionality of radar inputs, we adopt a Cross-View Fusion Mamba encoder to efficiently extract spatio-temporal features from longer sequences with linear complexity. A Spatio-Temporal-Cross Attention decoder then predicts joint coordinates across multiple frames. Together, this spatio-temporal modeling pipeline enables the model to leverage contextual cues from neighboring frames and joints to infer missing joints caused by specular reflections. To reinforce motion smoothness, we incorporate a velocity loss alongside the standard keypoint loss during training. Experiments on the TransHuPR and HuPR datasets demonstrate that our method achieves significant performance improvements, exceeding the baselines by 11.0 AP and 14.6 AP, respectively, while maintaining reasonable complexity. Code: https://github.com/NYCU-MAPL/milliMamba --- ### HEART-VIT: Hessian-Guided Efficient Dynamic Attention and Token Pruning in Vision Transformer Authors: Mohammad Helal Uddin, Liam Seymour, Sabur Baidya Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20120v1 Vision Transformers (ViTs) deliver state-of-the-art accuracy but their quadratic attention cost and redundant computations severely hinder deployment on latency and resource-constrained platforms. Existing pruning approaches treat either tokens or heads in isolation, relying on heuristics or first-order signals, which often sacrifice accuracy or fail to generalize across inputs. We introduce HEART-ViT, a Hessian-guided efficient dynamic attention and token pruning framework for vision transformers, which to the best of our knowledge is the first unified, second-order, input-adaptive framework for ViT optimization. HEART-ViT estimates curvature-weighted sensitivities of both tokens and attention heads using efficient Hessian-vector products, enabling principled pruning decisions under explicit loss budgets.This dual-view sensitivity reveals an important structural insight: token pruning dominates computational savings, while head pruning provides fine-grained redundancy removal, and their combination achieves a superior trade-off. On ImageNet-100 and ImageNet-1K with ViT-B/16 and DeiT-B/16, HEART-ViT achieves up to 49.4 percent FLOPs reduction, 36 percent lower latency, and 46 percent higher throughput, while consistently matching or even surpassing baseline accuracy after fine-tuning, for example 4.7 percent recovery at 40 percent token pruning. Beyond theoretical benchmarks, we deploy HEART-ViT on different edge devices such as AGX Orin, demonstrating that our reductions in FLOPs and latency translate directly into real-world gains in inference speed and energy efficiency. HEART-ViT bridges the gap between theory and practice, delivering the first unified, curvature-driven pruning framework that is both accuracy-preserving and edge-efficient. --- ### Delayed Bidirectional Alignment via Disentangled Audio Semantics for Audio-Visual Segmentation Authors: Jingqi Tian, Yiheng Du, Haoji Zhang, Yuji Wang, Isaac Ning Lee, Xulong Bai, Tianrui Zhu, Jingxuan Niu, Yansong Tang Published: 2025-12-23 Categories: cs.CV, cs.SD, eess.AS arXiv: https://arxiv.org/abs/2512.20117v2 Audio-Visual Segmentation (AVS) aims to localize sound-producing objects at the pixel level by integrating auditory and visual cues. However, existing methods often struggle with multi-source entanglement and audio-visual misalignment, leading to a dominance bias toward acoustically or visually salient objects (i.e., louder or larger ones) at the expense of subtler or co-occurring sources. To address these challenges, we propose DDAVS: Delayed Bidirectional Alignment via Disentangled Audio Semantics for Audio-Visual Segmentation. To mitigate multi-source entanglement, DDAVS employs learnable queries to extract audio semantics and anchor them within a structured semantic space derived from an audio prototype memory bank. This process is further optimized through contrastive learning to enhance discriminability and robustness. To alleviate audio-visual misalignment, DDAVS introduces dual cross attention with delayed modality interaction, improving the robustness of multimodal alignment. Extensive experiments on the AVS-Objects and VPO benchmarks demonstrate that DDAVS achieves state-of-the-art performance across single-source, multi-source, and multi-class multi-instance scenarios. These results validate the effectiveness and generalization ability of our framework under challenging real-world audio-visual segmentation conditions. Project page: https://trilarflagz.github.io/DDAVS-page/ --- ### Sample-Efficient Policy Constraint Offline Deep Reinforcement Learning based on Sample Filtering Authors: Yuanhao Chen, Qi Liu, Pengbin Chen, Zhongjian Qiao, Yanjie Li Published: 2025-12-23 Categories: cs.LG arXiv: https://arxiv.org/abs/2512.20115v1 Offline reinforcement learning (RL) aims to learn a policy that maximizes the expected return using a given static dataset of transitions. However, offline RL faces the distribution shift problem. The policy constraint offline RL method is proposed to solve the distribution shift problem. During the policy constraint offline RL training, it is important to ensure the difference between the learned policy and behavior policy within a given threshold. Thus, the learned policy heavily relies on the quality of the behavior policy. However, a problem exists in existing policy constraint methods: if the dataset contains many low-reward transitions, the learned will be contained with a suboptimal reference policy, leading to slow learning speed, low sample efficiency, and inferior performances. This paper shows that the sampling method in policy constraint offline RL that uses all the transitions in the dataset can be improved. A simple but efficient sample filtering method is proposed to improve the sample efficiency and the final performance. First, we evaluate the score of the transitions by average reward and average discounted reward of episodes in the dataset and extract the transition samples of high scores. Second, the high-score transition samples are used to train the offline RL algorithms. We verify the proposed method in a series of offline RL algorithms and benchmark tasks. Experimental results show that the proposed method outperforms baselines. --- ### Attention Fusion for Bridge Deck Delamination Detection Authors: Alireza Moayedikia, Amirhossein Moayedikia Published: 2025-12-23 Categories: cs.CV, eess.IV arXiv: https://arxiv.org/abs/2512.20113v4 Subsurface delaminations in reinforced concrete bridge decks escape conventional visual inspection, and the two principal sensing techniques used to find them are individually incomplete: Ground Penetrating Radar (GPR) penetrates deeply but degrades near the surface, while Infrared Thermography (IRT) resolves shallow defects but cannot reach deeper structure. This paper presents a framework for fusing the two modalities through hierarchical attention: temporal self-attention over GPR A-scans, channel-spatial attention over IRT patches, and cross-modal multi-head attention with learnable modality embeddings, coupled with decomposed aleatoric/epistemic uncertainty estimation. Beyond the architecture itself, which is lightweight at approximately 0.53M parameters with a closed-form accounting of where capacity resides, we contribute an elementary formal analysis. Two-token cross-modal attention is shown to be exactly a bank of per-sample learned gates; a gradient-allocation proposition quantifies how class imbalance starves attention parameters of minority-class signal and how loss reweighting trades that starvation for gradient variance; and closed-form metric floors under majority-class collapse anchor a diagnostic divergence between ranking metrics (AUC) and thresholded metrics (F1). The analysis suggests that adaptively weighted fusion, precisely because its feature-selection policy is learned, may be distinctively vulnerable to the severe class imbalance typical of operational bridge decks; establishing whether and when this occurs is deferred to empirical evaluation. --- ### Evolutionary Neural Architecture Search with Dual Contrastive Learning Authors: Xian-Rong Zhang, Yue-Jiao Gong, Wei-Neng Chen, Jun Zhang Published: 2025-12-23 Categories: cs.NE, cs.AI arXiv: https://arxiv.org/abs/2512.20112v1 Evolutionary Neural Architecture Search (ENAS) has gained attention for automatically designing neural network architectures. Recent studies use a neural predictor to guide the process, but the high computational costs of gathering training data -- since each label requires fully training an architecture -- make achieving a high-precision predictor with { limited compute budget (i.e., a capped number of fully trained architecture-label pairs)} crucial for ENAS success. This paper introduces ENAS with Dual Contrastive Learning (DCL-ENAS), a novel method that employs two stages of contrastive learning to train the neural predictor. In the first stage, contrastive self-supervised learning is used to learn meaningful representations from neural architectures without requiring labels. In the second stage, fine-tuning with contrastive learning is performed to accurately predict the relative performance of different architectures rather than their absolute performance, which is sufficient to guide the evolutionary search. Across NASBench-101 and NASBench-201, DCL-ENAS achieves the highest validation accuracy, surpassing the strongest published baselines by 0.05\% (ImageNet16-120) to 0.39\% (NASBench-101). On a real-world ECG arrhythmia classification task, DCL-ENAS improves performance by approximately 2.5 percentage points over a manually designed, non-NAS model obtained via random search, while requiring only 7.7 GPU-days. --- ### ABBEL: Learning Natural-Language Belief States for Memory-Efficient Interaction Authors: Aly Lidayan, Jakob Bjorner, Satvik Golechha, Kartik Goyal, Alane Suhr Published: 2025-12-23 Categories: cs.CL, cs.AI, cs.LG arXiv: https://arxiv.org/abs/2512.20111v2 As the time horizons of sequential decision-making tasks grow, keeping full interaction histories in model context becomes increasingly costly. Recent work reduces context lengths by instead conditioning decision-making agents on recursively updated natural-language summaries, which are concise and interpretable. However, they underperform agents with access to the full context, suggesting that they fail to generate sufficient summaries. To address this we propose ABBEL, a recursive summarization framework that isolates and directly supervises each summary's information contents in the form of explicit natural-language belief states. First, we analyze the belief states generated by frontier models under ABBEL across five domains, and verify that performance is often degraded due to omitting or incorrectly updating information. We also discover settings where models use memory inefficiently by retaining extraneous information. We target these limitations by fine-tuning with two RL-based methods: belief grading, which reduces update errors by rewarding belief generations based on their information content, and peak belief penalties, which encourage compressing the beliefs with the greatest memory footprints. We demonstrate that these methods significantly reduce the performance gap with full context models, and enable ABBEL to outperform prior memory agent work by 40% while using 67% of the memory. Our code is available at https://github.com/jakob-bjorner/optimal-explorer-dev --- ### UMAMI: Unifying Masked Autoregressive Models and Deterministic Rendering for View Synthesis Authors: Thanh-Tung Le, Tuan Pham, Tung Nguyen, Deying Kong, Xiaohui Xie, Stephan Mandt Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20107v1 Novel view synthesis (NVS) seeks to render photorealistic, 3D-consistent images of a scene from unseen camera poses given only a sparse set of posed views. Existing deterministic networks render observed regions quickly but blur unobserved areas, whereas stochastic diffusion-based methods hallucinate plausible content yet incur heavy training- and inference-time costs. In this paper, we propose a hybrid framework that unifies the strengths of both paradigms. A bidirectional transformer encodes multi-view image tokens and Plucker-ray embeddings, producing a shared latent representation. Two lightweight heads then act on this representation: (i) a feed-forward regression head that renders pixels where geometry is well constrained, and (ii) a masked autoregressive diffusion head that completes occluded or unseen regions. The entire model is trained end-to-end with joint photometric and diffusion losses, without handcrafted 3D inductive biases, enabling scalability across diverse scenes. Experiments demonstrate that our method attains state-of-the-art image quality while reducing rendering time by an order of magnitude compared with fully generative baselines. --- ### LiDARDraft: Generating LiDAR Point Cloud from Versatile Inputs Authors: Haiyun Wei, Fan Lu, Yunwei Zhu, Zehan Zheng, Weiyi Xue, Lin Shao, Xudong Zhang, Ya Wu, Guang Chen Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20105v2 Generating realistic and diverse LiDAR point clouds is crucial for autonomous driving simulation. Although previous methods achieve LiDAR point cloud generation from user inputs, they struggle to attain high-quality results while enabling versatile controllability, due to the imbalance between the complex distribution of LiDAR point clouds and the simple control signals. To address the limitation, we propose LiDARDraft, which utilizes the 3D layout to build a bridge between versatile conditional signals and LiDAR point clouds. The 3D layout can be trivially generated from various user inputs such as textual descriptions and images. Specifically, we represent text, images, and point clouds as unified 3D layouts, which are further transformed into semantic and depth control signals. Then, we employ a rangemap-based ControlNet to guide LiDAR point cloud generation. This pixel-level alignment approach demonstrates excellent performance in controllable LiDAR point clouds generation, enabling "simulation from scratch", allowing self-driving environments to be created from arbitrary textual descriptions, images and sketches. --- ### Effect of Activation Function and Model Optimizer on the Performance of Human Activity Recognition System Using Various Deep Learning Models Authors: Subrata Kumer Paula, Dewan Nafiul Islam Noora, Rakhi Rani Paula, Md. Ekramul Hamidb, Fahmid Al Faridc, Hezerul Abdul Karimd, Md. Maruf Al Hossain Princee, Abu Saleh Musa Miahb Published: 2025-12-23 Categories: cs.CV arXiv: https://arxiv.org/abs/2512.20104v1 Human Activity Recognition (HAR) plays a vital role in healthcare, surveillance, and innovative environments, where reliable action recognition supports timely decision-making and automation. Although deep learning-based HAR systems are widely adopted, the impact of Activation Functions (AFs) and Model Optimizers (MOs) on performance has not been sufficiently analyzed, particularly regarding how their combinations influence model behavior in practical scenarios. Most existing studies focus on architecture design, while the interaction between AF and MO choices remains relatively unexplored. In this work, we investigate the effect of three commonly used activation functions (ReLU, Sigmoid, and Tanh) combined with four optimization algorithms (SGD, Adam, RMSprop, and Adagrad) using two recurrent deep learning architectures, namely BiLSTM and ConvLSTM. Experiments are conducted on six medically relevant activity classes selected from the HMDB51 and UCF101 datasets, considering their suitability for healthcare-oriented HAR applications. Our experimental results show that ConvLSTM consistently outperforms BiLSTM across both datasets. ConvLSTM, combined with Adam or RMSprop, achieves an accuracy of up to 99.00%, demonstrating strong spatio-temporal learning capabilities and stable performance. While BiLSTM performs reasonably well on UCF101, with accuracy approaching 98.00%, its performance drops to approximately 60.00% on HMDB51, indicating limited robustness across datasets and weaker sensitivity to AF and MO variations. This study provides practical insights for optimizing HAR systems, particularly for real-world healthcare environments where fast and precise activity detection is critical. --- ### A Novel Graph-Sequence Learning Model for Inductive Text Classification Authors: Zuo Wang, Ye Yuan Published: 2025-12-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2512.20097v1 Text classification plays an important role in various downstream text-related tasks, such as sentiment analysis, fake news detection, and public opinion analysis. Recently, text classification based on Graph Neural Networks (GNNs) has made significant progress due to their strong capabilities of structural relationship learning. However, these approaches still face two major limitations. First, these approaches fail to fully consider the diverse structural information across word pairs, e.g., co-occurrence, syntax, and semantics. Furthermore, they neglect sequence information in the text graph structure information learning module and can not classify texts with new words and relations. In this paper, we propose a Novel Graph-Sequence Learning Model for Inductive Text Classification (TextGSL) to address the previously mentioned issues. More specifically, we construct a single text-level graph for all words in each text and establish different edge types based on the diverse relationships between word pairs. Building upon this, we design an adaptive multi-edge message-passing paradigm to aggregate diverse structural information between word pairs. Additionally, sequential information among text data can be captured by the proposed TextGSL through the incorporation of Transformer layers. Therefore, TextGSL can learn more discriminative text representations. TextGSL has been comprehensively compared with several strong baselines. The experimental results on diverse benchmarking datasets demonstrate that TextGSL outperforms these baselines in terms of accuracy. --- ### Information-directed sampling for bandits: a primer Authors: Annika Hirling, Giorgio Nicoletti, Antonio Celani Published: 2025-12-23 Categories: cs.LG, cs.IT arXiv: https://arxiv.org/abs/2512.20096v1 The Multi-Armed Bandit problem provides a fundamental framework for analyzing the tension between exploration and exploitation in sequential learning. This paper explores Information Directed Sampling (IDS) policies, a class of heuristics that balance immediate regret against information gain. We focus on the tractable environment of two-state Bernoulli bandits as a minimal model to rigorously compare heuristic strategies against the optimal policy. We extend the IDS framework to the discounted infinite-horizon setting by introducing a modified information measure and a tuning parameter to modulate the decision-making behavior. We examine two specific problem classes: symmetric bandits and the scenario involving one fair coin. In the symmetric case we show that IDS achieves bounded cumulative regret, whereas in the one-fair-coin scenario the IDS policy yields a regret that scales logarithmically with the horizon, in agreement with classical asymptotic lower bounds. This work serves as a pedagogical synthesis, aiming to bridge concepts from reinforcement learning and information theory for an audience of statistical physicists. --- ### Jensen-Shannon Divergence Message-Passing for Rich-Text Graph Representation Learning Authors: Zuo Wang, Ye Yuan Published: 2025-12-23 Categories: cs.LG arXiv: https://arxiv.org/abs/2512.20094v1 In this paper, we investigate how the widely existing contextual and structural divergence may influence the representation learning in rich-text graphs. To this end, we propose Jensen-Shannon Divergence Message-Passing (JSDMP), a new learning paradigm for rich-text graph representation learning. Besides considering similarity regarding structure and text, JSDMP further captures their corresponding dissimilarity by Jensen-Shannon divergence. Similarity and dissimilarity are then jointly used to compute new message weights among text nodes, thus enabling representations to learn with contextual and structural information from truly correlated text nodes. With JSDMP, we propose two novel graph neural networks, namely Divergent message-passing graph convolutional network (DMPGCN) and Divergent message-passing Page-Rank graph neural networks (DMPPRG), for learning representations in rich-text graphs. DMPGCN and DMPPRG have been extensively texted on well-established rich-text datasets and compared with several state-of-the-art baselines. The experimental results show that DMPGCN and DMPPRG can outperform other baselines, demonstrating the effectiveness of the proposed Jensen-Shannon Divergence Message-Passing paradigm --- ### Memory-T1: Reinforcement Learning for Temporal Reasoning in Multi-session Agents Authors: Yiming Du, Baojun Wang, Yifan Xiang, Zhaowei Wang, Wenyu Huang, Boyang Xue, Bin Liang, Xingshan Zeng, Fei Mi, Haoli Bai, Lifeng Shang, Jeff Z. Pan et al. Published: 2025-12-23 Categories: cs.CL arXiv: https://arxiv.org/abs/2512.20092v1 Temporal reasoning over long, multi-session dialogues is a critical capability for conversational agents. However, existing works and our pilot study have shown that as dialogue histories grow in length and accumulate noise, current long-context models struggle to accurately identify temporally pertinent information, significantly impairing reasoning performance. To address this, we introduce Memory-T1, a framework that learns a time-aware memory selection policy using reinforcement learning (RL). It employs a coarse-to-fine strategy, first pruning the dialogue history into a candidate set using temporal and relevance filters, followed by an RL agent that selects the precise evidence sessions. The RL training is guided by a multi-level reward function optimizing (i) answer accuracy, (ii) evidence grounding, and (iii) temporal consistency. In particular, the temporal consistency reward provides a dense signal by evaluating alignment with the query time scope at both the session-level (chronological proximity) and the utterance-level (chronological fidelity), enabling the agent to resolve subtle chronological ambiguities. On the Time-Dialog benchmark, Memory-T1 boosts a 7B model to an overall score of 67.0\%, establishing a new state-of-the-art performance for open-source models and outperforming a 14B baseline by 10.2\%. Ablation studies show temporal consistency and evidence grounding rewards jointly contribute to a 15.0\% performance gain. Moreover, Memory-T1 maintains robustness up to 128k tokens, where baseline models collapse, proving effectiveness against noise in extensive dialogue histories. The code and datasets are publicly available at https://github.com/Elvin-Yiming-Du/Memory-T1/ ---