Tag
This paper investigates perturbation-based data selection methods (RSP and RSP-Drop) for adapting medical language models, showing improved Macro-F1 scores over random selection when evaluated with MedGemma-4B on five medical QA datasets.
论文提出使用直接偏好优化（DPO）微调定义生成器，以减少同一词义下定义过度分化的问题，并在学术基准和欧盟可持续金融术语的历时语义变化审计任务上进行了验证。
This paper traces how subject–verb number agreement circuits emerge and sparsify across training in Pythia-1b and Pythia-410m models, finding that behavior stabilizes early while the underlying circuit keeps reorganizing, shifting attention heads, and shrinking to a strikingly sparse final form.
提出证据状态校准（evidence-state calibration）概念与SuffiCE方法，通过微调模型下一token分布中的置信度，使其随证据变化而准确反映答对概率，在临床诊断和多跳问答上将Brier分数降低最高60%、AUROC提升最高0.16。
A Cornell University article/survey explaining how to build modern discrete diffusion language models, covering iterative parallel refinement, fast sampling at thousands of tokens per second, and their potential to improve scaling laws over autoregressive LLMs.
提出 Conceptizer 训练框架，通过约 9K 个基于上下文同义词构建的概念词表，在 next-token prediction 之外增加意义层面的奖励信号，在 Llama 3、Qwen2.5 和 Falcon3（1B-8B）上提升了困惑度与下游知识/情感任务表现。
该论文构建了两个相连的语言模型副本，研究当一个副本可读取另一个的内部记忆时，它是否会将对方的规则和代码词当作自己的，发现内容覆盖到自身记忆上时默认被标记为'我的'，因此模型间交换内部状态的系统必须同时测试所有权与传输。
论文研究了循环语言模型中深度可复用性的来源，发现保持循环步数可外推的关键不是循环边界归一化，而是训练与推理时循环拓扑的一致性（agreement）；移除 Ouro 模型的边界归一化会在训练预算内将准确率降至零，但从头训练的小模型显示不加归一化反而外推更好。
The paper proposes a rewriting scaling law that predicts loss from model size, training tokens, rewrite fraction, and data epochs while accounting for rewrite generation compute, finding optimal rewrite fractions increase with compute and that models trained on rewrites can outperform their rewriters.
This paper develops a direction-sensitive, readout-relative account of semantic fragility in language models, showing that heterogeneous task-preserving rewrites can share failure dynamics while lacking a common input-independent failure direction, validated on Qwen3-32B and Gemma 4 31B.
该论文通过六语言析因设计，独立考察问题语言、回答语言与验证指令语言对多语言 P(True) 自评估的影响，发现三者分别影响准确率、区分度与校准，且未经验证的评判者会严重扭曲模型表现差距。
提出 silver-evidence demonstrations 方法，用生成的、目标条件化的证据摘要替换 ICL 示例中的原始检索段落，以减少噪声检索带来的干扰。该方法在 EI 2.0、SciFact、合成策略检索和 HICRIC/IMR 数据集上优于 gold-only ICL 和 InstructRAG 式 ICL，并在所有标注数据集上提升了证据 F1。
A behavioral case study across six frontier LLMs finds they confidently misattribute a 17-year-old's exam answer to an adult, with stated confidence uncorrelated or inversely correlated with correctness, and articulate retrospective bias recognition that fails to improve blind judgments.
该论文研究前沿语言模型的概率判断是否满足全概率律（含因果 do-演算的后门调整），发现 GPT-6.1 Sol 和 Claude Opus 5.5 最一致，而推理长度与一致性无关，且一致性残差可在无真值情况下标记不可靠回答。
This paper introduces the Schema Coherence Inventory (SCI), a 96-item forced-choice psychometric tool measuring whether beliefs assigned to persona-conditioned language models actually organize their cognition versus merely inducing a mood, validated across 107 conditions and five model backbones.
论文提出IEC框架，研究语言模型在受控指令负载（如无意义的单词替换规则）下的行为与性能退化，揭示了规则数量、替换密度、规则位置等因素遵循的数学规律，并在GPT-4o mini上验证，为模型建立可复现的"能力护照"。
该论文提出词级未登录词交叉熵（UCE）作为衡量预训练语料与零样本基准词面对齐的无模型指标，并通过受控预训练实验表明语料-基准的词面对齐能一致地预测零样本准确率，同时在文档选择与提示路由层面展示了可操控的收益。
论文提出 CREDIT 方法，通过反事实训练让语言模型准确自我归因——判断回答在多大程度上受用户提示影响，而无需人工标注。Qwen3-8B 的平衡准确率从 0.49 提升至 0.86，四个开源模型均超过前沿模型表现，代码已开源。
该论文提出了一种经验性加法幂律，将长上下文持续预训练（LCCP）的收益与模型规模及短/长上下文token预算关联起来，并引入Constant-Last4协议实现稳定监控。实验表明LCCP的收益在评测范围内无明显饱和迹象，主张不应过早停止长上下文持续预训练。
The paper introduces Hierarchical Masked Diffusion Language Models (HMDLM), which incorporate position-level priority hierarchies into masked diffusion language models, achieving lower perplexity on LM1B and WikiHow and improved Sudoku solution accuracy over existing diffusion baselines.