GEWU
EN

搜索科普文

输入关键词搜索已发表科普文。

    Disciplines / Research questions

    情报学

    10 个研究主题 · 31 个问题

    持续学习中的灾难性遗忘与可塑性—稳定性矛盾4 个问题

    1. Applicability of Frobenius-Isometry Reinitialization (FIRE) under Restricted Past-Data Access and at Large-Language-Model Scale

    2. Completing the scaling-limit (DMFT) theory of continual learning: finite-width convergence, the learning-rate-driven lazy–rich transition, depth, and mitigation scaling

    3. Extending the readout-decomposition of activation change (RDAC) framework beyond one-hidden-layer linear, task-incremental continual learning

    4. Predictive theory of the architecture-level stability-plasticity trade-off in continual learning: when depth versus width favors stability or plasticity at matched parameter budgets

    多智能体协作中的信息级联与群体极化2 个问题

    1. Formalizing infostorms: the exact relations between information technologies and amplified pluralistic ignorance, informational cascades, and belief polarization, and design principles to prevent them

    2. Causal mechanism, generality, and ecological validity of polarization in collectives of autonomous generative-AI agents under echo chambers

    机器遗忘的有效删除与效用保持权衡3 个问题

    1. A Principled Termination Criterion for the Machine-Unlearning Process

    2. Minimal alignment requirements and deletion guarantees for residual feature alignment unlearning of pre-trained models

    3. Representation geometry and unlearning fidelity: characterizing and closing the utility gap of label-free, representation-space approximate machine unlearning

    奖励模型过优化导致的性能反转3 个问题

    1. Cost-efficient and reliable reward-model evaluation via the degree of overoptimization

    2. Overoptimisation dynamics and stability of iterated RLHF beyond controlled small-scale settings

    3. Ground-Truth-Free Identification of Reward-Model Overoptimization Thresholds in Composite-Reward RLHF

    离线强化学习中的分布外动作价值高估3 个问题

    1. Safe out-of-distribution value generalization in offline reinforcement learning beyond the convex hull and its neighborhood

    2. Convergence and action-value gap guarantees for the Imagination-Limited Bellman operator under function approximation

    3. Finite-sample control of the segmentation error probability in uncertainty-penalized Q-learning with a consistency model (QDQ)

    神经符号系统组合泛化失败1 个问题

    1. Closing the procedural composition gap on ClassicLogic: learning systems that compose known puzzle-solving strategies into novel multi-step plans

    世界模型长期滚动中的误差累积临界点4 个问题

    1. Domain of validity of the phase-drift mechanism for autoregressive rollout error in latent-space reduced-order models of bluff-body wakes

    2. Empirical validity of the GEAF proxy for the joint node-edge error operator in Graph World Models beyond synthetic simulators

    3. Closing the residual cache-maintenance gap for open-ended autoregressive video diffusion, and extending drift mitigation to multi-shot generation

    4. Does the Coverage Account of World-Model Hallucination Transfer to Billion-Parameter Models and Real Robot Data?

    思维链可解释性与真实因果过程不一致4 个问题

    1. Ground-truth-validated chain-of-thought faithfulness measurement at both the step and CoT level

    2. Causal Origins of Implicit Post-Hoc Rationalization: Are Answer-Driving Biases Encoded Before Chain-of-Thought Generation?

    3. Generalizing and validating CoT-induced feature sparsity, interpretability, and causal transfer in language models beyond a single token position and small Pythia models

    4. Causal Origins and Generality of the Dissociation Between Chain-of-Thought Influence and Verbalised Faithfulness Across Post-Training Recipes

    智能体工具使用能力随模型规模出现的非单调变化3 个问题

    1. Dynamic thinking policies and the generality of orchestration-over-scaling gains in small multi-agent LLM systems

    2. Why does inverse scaling become U-shaped? The causal mechanism and predictive boundaries of scaling-trend reversals in language models

    3. Execution-faithful evaluation of programmatic tool calling: certifying genuine tool execution beyond echo-return benchmarks

    智能体基准中记忆污染与真实规划能力区分4 个问题

    1. Contamination-resistant evaluation of LLM software-engineering agents: validated benchmark construction, temporal-freshness protocols, and black-box contamination detection

    2. Open problems in search-time contamination: comprehensive measurement, temporal dynamics, mitigation trade-offs, and evaluation-aware agent behavior

    3. Contamination detection for large reasoning models robust to RL-induced concealment and chain-of-thought contamination

    4. Belief-sufficient agent memory for interdependent multi-session tasks: representation, joint training, belief-fidelity evaluation, and the latency trade-off

    返回开放问题