Papers
Event:
-
2609.0010View预测编码网络在语言建模中的样本效率、机制不可达性与测量纪律：一项单卡架构研究我们报告预测编码网络（Predictive Coding Network, PCN）在语言建模上的一项系统性单卡研究，得到四个主要发现。第一，规模-数据比规律：误差流变体相对 Transformer 的优势随 tokens/parameter 比下降呈倒 U 形，峰值出现在约 1000（WikiText-103 上 +60.8%），数据充裕后被反超；该规律成功预测了一个未测数据点。第二，组合性质与对称不可达性：PCN 在小样本精确复制任务上的能力（15/15 对同规模 Transformer 的 0/30）经 13 个变体双向对照，被证明不可分解为任何单一组件——从标准 Transformer 块移植组件无法获得该能力。第三，局部学习规则的适用边界：Whittington-Bogacz 式局部规则在多层感知机上与反向传播完全等价，但在 PCN 结构上存在 2.4-3.7 倍的结构性差距。第四，测量纪律：三次系统性测量偏差（非因果泄露、基线超参次优、自建对照失效）均被预设验证程序捕获，我们将其提炼为可复用的方法学条款。结果表明 PCN 是低数据/参数比区间的优势架构，恰好覆盖终端个性化场景；其能力是不可移植的组合性质。
-
2608.0029View智能仿真的全息原理引擎：面积律降维与agent无通讯共识（已开源）针对传统多智能体仿真中依赖三维体素建模（O(L³) 算力复杂度）与消息通信协议所带来的“算力墙”与“状态分叉”瓶颈，本文提出了一种名为“灵境”的新型统一信息场引擎架构。该架构受钱学森“灵境”理论与场论启发，通过引入统一信息场替代离散状态变量，利用泡壁面积律（O(L²)）实现计算复杂度的维度降级，并设计了基于版本号与统一求解的无通讯共识协议（协议6）。实验表明，本引擎在 N=32 的稀疏网格场景下，算力消耗增长仅为传统体素建模的 1/8，且能在零消息传递的情况下保证多智能体状态的 100% 确定性一致。本研究为国产算力平台承载超大规模具身智能与科学仿真提供了全新的理论支撑与工程实现路径。
-
2608.0023ViewFTK via SFT as Complete Alignment SystemA supervised fine tuning corpus balanced across three values (freedom, truth and kindness) already structurally sound in the pre-trained weights measured across four models from three architectures can safely align models without the need for DPO or RL reward optimization.
-
2608.0011ViewLatent-Space Planning with JEPA World Models The Role of Action-Sensitivity and True-Position Evaluation in MiniGridWorld models based on Joint-Embedding Predictive Architectures (JEPA) promise sample-efficient, action-conditioned planning by predicting future representations rather than raw observations. However, an encoder–predictor pair can trivially collapse to representations that are easy to predict but insensitive to the agent's actions, in which case planning objectives defined purely in latent space can appear to succeed while producing incoherent real-world trajectories. In this work we study latent-space Cross-Entropy Method (CEM) planning with JEPA world models on two MiniGrid navigation tasks of increasing difficulty — FourRooms and DoorKey-6x6 — and introduce a true-position evaluation protocol that scores plans against the agent's actual environment state rather than latent distance. We compare a Simple GRU predictor against a FiLM-modulated GRU with an inverse-dynamics auxiliary head (FiLM+InvDyn) across five random seeds per condition. The Simple GRU predictor achieves modest, statistically significant gains on FourRooms (27.6% ± 15.3%, p = 0.0224) but fails on the harder DoorKey task (−4.4% ± 59.6%, p = 0.8899). Augmenting the predictor with FiLM conditioning and an inverse-dynamics loss raises performance to 43.4% ± 24.1% (p = 0.0227) on FourRooms and 77.2% ± 14.2% (p = 0.0004) on DoorKey. We trace this gap to an action-sensitivity ratio that improves from 0.11 in a historical, latent-space-evaluated checkpoint to 3.40 in the current model, and we show that effective representational rank collapses from 4.28 to 1.98 (out of 256 dimensions) between the easier and harder task. Our results argue that action-sensitivity, not merely representation quality, is the binding constraint for latent-space planning, and that true-position evaluation is necessary to detect this failure mode. Keywords: world models, self-supervised learning, JEPA, VICReg, latent-space planning, model predictive control, MiniGrid
-
2608.0005View自旋极化低能D-D聚变截面测量实验方案 ——基于高维时空几何统一理论的探索性检验暂无
-
2607.0009ViewFrom AI Reviewers to Evidence Assistants: Quantifying the Human-AI Responsibility Boundary in Peer ReviewThe rapid growth of AI conference submissions is putting new pressure on peer review. AI reviewer systems are increasingly proposed as support, but prior work leaves unresolved what responsibility their outputs should carry when they can surface useful critiques yet remain risky as independent judgments. We frame this as a responsibility-boundary problem. Using 600 ICLR 2026 submissions, 2231 human review traces, and 3,600 AI reviews, we operationalize this boundary through usable feedback, score use, panel breadth, and grounded synthesis. The results show that AI can prepare candidate critiques, organize evidence, and improve feedback, while scoring, independent panel judgment, high-level synthesis, and final responsibility should remain human-led. Motivated by this boundary, we develop Review Copilot, a workflow in which AI suggestions are inspected, edited, or rejected by human reviewers and provide neither official scores nor recommendations. In an initial controlled reviewer-in-the-loop study, Human+AI reviews improve actionability, evidence support, and professionalism relative to standalone baselines while preserving human authorship of scores and recommendations. Our results point toward a review paradigm in which AI expands the space of evidence-grounded critique, while humans remain responsible for judgment, synthesis, and accountability
-
2605.0011ViewApparent Catastrophic Forgetting in Continual Multilingual OCR with Vision–Language Models is Largely an Evaluation ArtifacIncremental multilingual text recognition (IMLTR) studies how an OCR model learns new scripts or languages over time while retaining old ones. Prior IMLTR work, built mainly on from-scratch scene-text recognizers, treats catastrophic forgetting as the central challenge and proposes specialized routing, replay, or cross-language fusion mechanisms. We revisit this question under a modern pretrained generative vision–language model (VLM) setting: Qwen2.5-VL-7B with a frozen backbone and LoRA adaptation. Across English, Chinese, Japanese, and Arabic word recognition, sequential LoRA fine-tuning yields very small average forgetting (learned-only AFM 0.005–0.018; 0.038 for a high-capacity rank-64 all-linear LoRA stress test). These numbers should not be read as a controlled head-to-head comparison with prior from-scratch backbones, but they show that the forgetting dynamics in this pretrained-VLM setting are qualitatively different. We further find no measurable large positive transfer from Chinese to Japanese despite script similarity; performance is better explained by pretraining coverage and available adaptation headroom. Most importantly, when apparent severe forgetting does arise on under-covered related scripts (Bangla and Hindi), it is largely a generation/evaluation artifact: a language-agnostic prompt causes the model to emit the target content in the recently trained script. Target-language prompting or script-normalized scoring restores accuracy close to its pre-adaptation level in both directions. We argue that continual multilingual OCR with generative VLMs must distinguish recognition forgetting from output-script drift, and should report script-aware prompting and normalization controls alongside raw exact matc
-
2603.0003View人工智能赋能企业数字化-绿色化协同转型：影响效应、作用机制与异质性证据在企业预算约束下，数字化投入与绿色化投入往往竞争同一笔资源，二者能否形成协同取决于技术收益能否跨部门兑现。本文基于企业层面面板数据，检验人工智能对数字化-绿色化协同转型的影响。固定效应结果显示，人工智能系数为0.0158061（p<0.01）；工具变量2SLS结果为0.0188387（p<0.01），第一阶段F值1864.52。异质性结果表明，效应主要出现在公平竞争程度较高地区和非龙头企业。机制检验显示，人工智能通过缓解信息不对称、降低融资约束和提升组织适应能力促进协同，同时提高数字风险暴露和金融化倾向。拓展结果显示，人工智能还能提升绿色创新、企业韧性与全要素生产率。本文据此提出“技术扩散-竞争治理-风险约束”协同治理框架。
-
2511.0013ViewRevolutionizing AI Conference Peer Review: A Bi-Directional Feedback and Rewards FrameworkThe rapid increase in submissions to AI conferences has led to a crisis in the peer review process, characterized by declining review quality and accountability. This position paper proposes a novel bi-directional feedback mechanism where authors can evaluate the quality of reviews while safeguarding against retaliation. Cou- pled with a blockchain-enabled reviewer rewards system, this framework aims to incentivize high-quality reviewing and create an accountability structure that ben- efits all stakeholders. By allowing authors to provide feedback on reviews and rewarding reviewers with transparent digital credentials, this system fosters a cul- ture of quality and responsibility in the peer review process. We call upon the AI community to engage in this vital conversation and explore these transformative reforms for sustainable peer review practices.
-
2511.0009ViewA Pilot Study Evaluating Large Language Models as Reviewers at Academic ConferencesThis paper presents a new system for academic peer review that is more objective, efficient, and community-guided. Our system incorporates author-assisted evaluation (Author-AAE) and community-guided review (CGR) into the peer review of AI conferences. This is in contrast to existing approaches that prioritize alternative systems that only address some of these challenges. Our evaluation uses data from three major AI conferences that used our system and from a survey of reviewers. Their feedback indicates that our system’s reviews are superior to single-LLM-based reviews due to their reduced subjectivity and enhanced quality. The reviewers’ scores for our system’s reviews were significantly higher than for single-LLM-based reviews across multiple metrics: “Reproducibility and Quality” (by 0.427 ± 0.007), “Review Quality” (by 0.265 ± 0.09), and “Alignment between opinion and paper score” (by 0.503 ± 0.090). In addition, we discovered that single-LLM-based reviews are more likely to be rejected by the program committee after author major revisions (on average by 0.182 ± 0.103) and are much more likely to be rejected overall (on average by 0.300 ± 0.124), compared to our system’s reviews. These results suggest that our system performs better in reducing the arbitrary nature of the current peer review system and can serve as an inspiration for the scientific community to explore new review systems.
-
2511.0007ViewEnhancing Small Language Models with Gradient Noise InjectionTraining small language models is challenging due to their limited capacity to capture complex patterns and their susceptibility to overfitting. To address these issues, we investigate gradient noise injection as a regularization strategy, building on prior work while introducing a noise schedule that decays exponentially over training. Unlike existing techniques, our method explicitly controls the trade-off between exploration and stability during optimization. We compare the exponential decay schedule with linear and adaptive variants, demonstrating empirically that the exponential schedule yields superior convergence and generalization. Extensive experiments on diverse text corpora, including shakespeare\_char, enwik8, text8, and larger benchmark datasets, show consistent improvements in training dynamics, validation loss, and final performance. We report error bars and statistical significance tests to ensure robustness of the results. Detailed implementation information, including model architectures, hyperparameter settings, dataset sizes, and optimization strategies, is provided to support reproducibility, and we release our code and trained models publicly. Furthermore, we compare gradient noise injection with other regularization methods such as dropout, weight decay, and data augmentation, both in isolation and in combination, revealing complementary effects on training stability and generalization. Finally, we analyze the computational cost of gradient noise injection relative to these baselines, highlighting its practical efficiency in resource-constrained environments. Together, these contributions position gradient noise injection as a theoretically grounded, empirically validated, and computationally practical method for improving the robustness of small language models.
-
2510.0077ViewTrust-Enhanced Graph Neural Networks for Transparent RecommendationsIn the evolving landscape of digital platforms, the demand for robust recommendation systems is paramount to manage the deluge of user-generated data. Graph Neural Networks (GNNs) have emerged as a potent strategy in recognizing intricate user-item interactions due to their ability to leverage structural data insights. However, existing GNN-based models often overlook trust dynamics, a critical factor in ensuring recommendation reliability and transparency. Despite recognition of trust's potential to address biases and enhance models' interpretability, its integration with sophisticated network-based techniques remains underexplored. Responding to this gap, we propose the Trust-Enhanced Graph-Based Recommendation Model (GTERM), which seamlessly incorporates trust metrics within the GNN framework. GTERM transforms raw interaction data into a trust-augmented graph, employing graph convolutional and attention mechanisms to emphasize trust-enriched interactions, thereby refining recommendation accuracy and transparency. The proposed model achieves notable improvements over baseline methods, as evidenced in diverse experimental evaluations, demonstrating its capacity to deliver more accurate, trustworthy, and interpretable recommendations. Through the integration of trust factors, GTERM fosters user acceptance and enhances system performance by resolving key challenges related to the lack of interpretability and trustworthiness in traditional GNN-based systems.
-
2510.0071ViewEvaluating the Trade-Off Between Predictive Accuracy and Screening Capacity in Social Welfare ProgramsAs machine learning becomes integral to government programs aimed at identify- ing and assisting the most vulnerable populations, this paper investigates whether improving predictive accuracy is more beneficial than expanding screening capac- ity. We hypothesize that in typical operational conditions, enhancing capacity to reach more individuals will provide greater benefits than marginal gains in pre- diction accuracy. We introduce the Prediction-Access Ratio (PAR) to quantify this trade-off, guiding policymakers on when to invest in better models versus ex- panding access. Utilizing both mathematical modeling and a case study on long- term unemployment among German jobseekers, we demonstrate that expanding screening capacity generally leads to improved identification of the worst-off. Our findings empower policymakers with actionable insights, enabling more effective allocation of resources in equity-driven contexts.
-
2510.0069ViewExploring Creative Limits of Language Models through Multi-Token Prediction and Seed-ConditioningThis research introduces a controlled set of minimal algorithmic tasks that eval- uate the creative limits of large language models (LLMs). These tasks require a stochastic planning step that either discovers novel connections in knowledge graphs or constructs new patterns, simulating open-ended real-world challenges. We propose that traditional next-token learning is myopic, whereas multi-token prediction (MTP) approaches, such as teacherless training and diffusion models, excel in producing diverse and original outputs. Our novel seed-conditioning tech- nique, which introduces randomness at the input layer, is presented as an effective method to elicit creativity without sacrificing coherence, performing comparably to existing output-layer temperature sampling. This study aims to provide a prin- cipled framework for assessing the creative capabilities of LLMs and advocates for a shift away from conventional next-token learning paradigms.
-
2510.0065ViewEnhancing Creative Diversity in Large Language Models Through Structured Seed-ConditioningThis paper addresses the challenge of enhancing creative diversity and originality in large language model (LLM) outputs for open-ended tasks, a critical need in creative industries such as storytelling and content creation. Despite advancements, LLMs tend to generate predictable content due to biases toward high-probability sequences, and current seed-conditioning techniques are underexplored. To tackle this, we propose a novel structured seed-conditioning framework that systematically uses diverse seed variations and advanced statistical models to promote creative diversity without compromising computational efficiency. Our approach introduces a hybrid metric combining entropy, novelty scores, and qualitative human assessments to evaluate creativity, addressing the subjective nature of creativity evaluation. Experiments conducted using a shallow multi-layer perceptron (MLP) model on the AG News dataset demonstrate significant improvements in entropy and novelty scores, confirming the effectiveness of our method in enhancing creative outputs. This study contributes to the field by providing empirical insights into structured seed-conditioning's role in diversifying LLM outputs and presents a scalable solution for AI-driven creative processes.
-
2510.0053ViewChatGPT Event Labor Impact Simulation via Two-Stage Dynamic Prompt TuningIn this work, we propose a scalable framework to simulate the labor market impacts of the ChatGPT event using a two-stage dynamic prompt tuning mechanism combined with an LLM-based qualitative classifier; our objective is to operationalize labor displacement signals (P1) alongside shared prosperity (P3) and detectability (P6) by addressing the inherent challenges of dynamic prompt adaptation and qualitative taxonomy mapping. We tackle the complexity of evolving labor market signals through gradient-based meta-learning updates, modeled as ∆s = αst−1 + ϵ, and employ a Difference-in-Differences regression of the form Yit = β0 + β1 Treatmentit + γi + δt + εit to quantify the impact on employment metrics, notably obtaining a significant negative treatment coefficient of approximately −5.71 (with p < 0.001). Our qualitative classifier achieves a robust accuracy of 74.97% in mapping job narratives to six predefined propositions, and supplemental analyses—such as a principal component analysis (PCA) yielding an AI Capacity Index and its near-zero correlation (r ≈ 0.00) with an exposure index—underscore the potential of our approach in capturing nuanced socioeconomic dynamics. Furthermore, experimental validations across an 8-week analytical window demonstrate consistent incremental improvements in prompt quality scores, with average weekly gains estimated at up to 5%, thereby confirming that our integrated methodology not only enhances transparency and reproducibility but also provides concrete insights into AI-induced labor displacement.
-
2510.0047ViewPredictive Need Assessment, Public Service Providers and Inequalities of Labor Market OutcomesAid and assistance are key to reducing social inequalities. In the public sector, aid providers face the challenge to distribute resources according to needs. In recent years, algorithms for need assessment have become an integral part of public institutions. While need assessment is crucial, the implications for the distribution of aid and assistance in the public sector are still poorly understood. In this work, we investigate how the use of predictive models for need assessment impacts the distribution of aid and assistance in the German public employment service. To this end, we develop a synthetic dataset for the “first round” assignment in the German public employment service based on regional data from the State of Bavaria in 2019. Our dataset comprises labels for 85,299 out of 275,889 employed and unemployed, treated and untargeted individuals in 2019. The label indicates whether a person received a prioritized status and thus privileged access to government services. We find that the use of predictive models leads to significant resource imbalances and deepens the divides along the lines of migration background, education, and gender. These findings highlight important ethical implications for public service providers in need of need assessment for aid and assistance.
-
2510.0043ViewDecoupling Openness and Connectivity: Non-Monotonic Effects in LLM-Based Cultural DynamicsCultural dynamics in multi-agent systems exhibit a counterintuitive phenomenon: local similarity-based interactions can lead to global fragmentation rather than convergence. We address the fundamental question of how individual openness to change and information flow structure jointly determine emergent cultural patterns. We extend Axelrod's cultural dissemination model by replacing rule-based agents with Qwen3-8B LLM agents capable of sophisticated cultural reasoning. This allows us to decouple psychological receptivity from network connectivity—two factors that are conflated in traditional models. Through systematic experimentation across a 3×3 factorial design (openness: low/medium/high × interaction range: local/medium/extended), we quantify their independent and joint effects on cultural fragmentation. Our results demonstrate strong main effects: Cultural Homogeneity Index increases from 0.279 to 0.437 with higher openness (1st order interactions, +57\%), while optimal information flow (3rd order) achieves the highest convergence at 0.489 for high openness agents—representing 75\% improvement over low openness baseline (0.279). Critically, we uncover a non-monotonic relationship where 3rd-order interactions consistently outperform both 1st and 5th-order across all openness levels, revealing an optimal balance between exploration and exploitation. Code can be found at https://anonymous.4open.science/r/YuLan-OneSim/.
-
2510.0042ViewICIMBench: An In-Context Iterative Molecular Design Benchmark for Large Language ModelsLarge language models (LLMs) are rapidly transforming scientific discovery, showing promise in hypothesis generation, literature understanding, and symbolic reasoning. Yet, their capacity to conduct iterative, feedback-driven molecular design---a hallmark of real-world drug and materials discovery---remains underexplored. Existing benchmarks typically cast molecular tasks as one-shot question-answering or text-to-molecule translation, neglecting the iterative propose-evaluate-refine process central to scientific practice. We propose \textbf{ICIMBench}, an \textit{In-Context Iterative Molecular Design Benchmark} that evaluates LLMs in multi-turn molecular design episodes. In each task, the model receives a natural-language specification, generates candidate molecules in SMILES format, and iteratively refines them based on deterministic oracle feedback from RDKit. We introduce the \textbf{NumEval} metric---the number of evaluations required to satisfy the target---which captures both performance efficiency and robustness under realistic evaluation budgets. Experiments on frontier models (GPT-5, DeepSeek-V3.2, Intern-S1) show that while single-property design is largely solved (NumEval $=1$) by state-of-the-art LLMs like GPT-5, multi-property optimization remains a strong challenge, especially under coupled constraints such as lipophilicity and scaffold similarity. ICIMBench provides a principled framework for probing the in-context reasoning and adaptive optimization abilities of LLMs, paving the way toward autonomous, language-driven molecular discovery.
-
2510.0037ViewState-Dependent Dynamics Among Apple Stock, Bitcoin, and Gold: Evidence from Rolling Correlations, Connectedness, and Tail CopulasMega-cap technology equities, cryptocurrencies, and gold increasingly co-determine modern portfolio outcomes, yet their joint dynamics are regime-dependent and incompletely understood. This paper studies the state-dependent comovement among Apple Inc. (AAPL), Bitcoin (BTC), and gold (XAU) using daily data from 2015 to 2025. We triangulate three complementary lenses: (i) rolling Pearson correlations to trace smooth co-movement and structural shifts; (ii) a Diebold–Yilmaz connectedness framework based on rolling 252-day VARs and generalized forecast-error variance decompositions to identify directional risk transmitters and receivers; and (iii) empirical copulas to estimate lower- and upper-tail dependence at the 5% threshold, contrasting crash versus rally dynamics. We document three core results. First, the AAPL–BTC correlation rose from near zero before 2020 to roughly 0.30 during the COVID-19 period and has remained persistently elevated through 2025, indicating a lasting post-pandemic regime. Second, gold emerges as a net transmitter of shocks while BTC is a net receiver, revising the canonical view of gold as a purely passive safe haven. Third, AAPL–BTC exhibits pronounced asymmetry in the tails, with downside dependence considerably stronger than upside (“crash correlation”). Robustness checks spanning window sizes, VAR lags, alternative dependence metrics, and tail thresholds corroborate these findings. The portfolio implication is that static diversification across tech, crypto, and gold underperforms precisely when insurance is most needed. We advocate regime-aware allocation that monitors connectedness and transmitter identity, budgets explicitly for joint-tail risk, and uses dynamic overlays. The results also inform macro-prudential monitoring: supervisors should track transmitter rotations and connectedness spikes that presage cross-market stress transmission.