contrastive-learning

A laptop-scale replication of Balestriero & LeCun's "good views beat next-token prediction" claim using only human COCO captions and 11M-param text encoders confirms the ordering (good views 0.320 > bad views 0.192 > next-token 0.148 on mutual 10-NN with DINOv2), while noting smaller retrieval gaps and scale limitations.

A paper proposing a unified training objective combining focal loss and supervised contrastive learning for automatic CEFR level assessment, achieving state-of-the-art results across sentence- and document-level monolingual and multilingual benchmarks.

A new paper by Randall Balestriero and Yann LeCun argues that paired-view training (e.g., image-caption pairs), not sheer scale or next-word prediction, is what produces unified world representations—showing a 109M-parameter model outperforming a 72B language model in representational alignment.

A survey and position paper arguing that embedding research across modalities is converging via contrastive learning and Transformers, proposing a framework of coverage, comparability, and compositionality with transformation, alignment, and joint embedding as paths toward omnimodal unification.

The paper turns the Platonic Representation Hypothesis into a theorem, proving that JEPAs trained separately on multiple modalities converge to shared representations up to a single rotation, and shows that next-token prediction is essentially a JEPA with 'bad views' — making it far less efficient for cross-modal alignment (good views need 42x fewer samples).

Meta-CLIP introduces a metadata-conditioned contrastive learning framework that reweights cross-modal negatives using DICOM acquisition fields, improving zero-shot disease classification on CT-RATE and several CXR benchmarks over CT-CLIP.

该论文提出一个理论框架，分析监督对比学习目标在长尾分布下是否能形成正则单纯形几何，并提出一种完全平衡的对比目标，可证明在最优时形成正则单纯形，从而在长尾识别和迁移学习任务上取得更好的性能。

论文提出 ANT-A 多标注者基准（跨7个数据集、每样本5个人工标注）以及多智能体标注框架 ANT-OR，通过从正负反馈中提炼对比决策规则优化提示，在 Fleiss kappa 一致性上显著超越基线。

A paper reports a systematic gap in modern text embedders: sentiment is highly decodable via linear probes but nearly absent from their cosine-similarity metric, and proposes SISE, a tiny supervised-contrastive head that fixes this on frozen encoders.

TRACE is a multimodal framework that organizes daily clinical events (EHR, chest x-rays, radiology reports, discharge summaries) into temporal trajectories for inpatient risk prediction, achieving the highest AUROC across nine dataset-task settings on MIMIC-IV, eICU, and MIMIC-III cohorts.

MIT researchers published a Nature Communications paper on a multimodal AI model that combines 3D atomic structures with natural language to better predict and design crystalline materials, outperforming text-only models in identifying promising candidates like solid-state lithium-ion battery electrolytes.

该论文提出一个对比学习框架，将EEG信号与冻结的预训练文本编码器的嵌入空间对齐，用于解码想象言语，并在Chisco数据集上验证了其在固定意图解码、相似度检索和语义结构探针方面的能力。

CLUE提出了一种对比学习框架，用于无监督的事件类型归纳、命名与标注，通过联合嵌入表示捕捉动词语义与多义性，并用对比损失引导的聚类替代人工事件本体构建。

该论文在统一可控的基准框架下系统评估了多模态嵌入学习中的各类对比训练策略，发现大多数策略相比标准对齐基线仅有边际收益甚至带来负面影响，并提出结合难负样本挖掘、假负样本缓解与难批次构建的协同训练方案，在MMEB上以Qwen2-VL-2B平均提升6.0%。

Meta AI introduces MetaEncoder, which fine-tunes a pre-trained Muse-Glimmer 30B decoder into an instruction-following decision-making encoder using unidirectional contrastive learning, evaluating it across 11 benchmarks and 190 tasks for multimodal closed-set and open-set decision making.

ReCast 提出一种面向归因的步表示学习方法，将冻结 LLM 的隐藏状态转换为可更好区分根因步骤的表示，并发布 ReCast-2K 训练数据集，在四个基准上取得最佳 Hit@1 成绩。

该论文提出将 GUI 世界模型的预测目标从全屏像素重建转移到潜空间中的动作诱导未来信息，通过共享的多模态模型表示配合动作头与转移头，并引入动作感知对比目标来缓解身份捷径问题，相比像素空间世界模型最高实现 31.9% 和 40.6% 的成功率相对提升，并将推理延迟降低最多 98.9%。

This paper introduces Anchor Divergences, a method that leverages the interplay between contrastive learning, exponential families, and information geometry to define context-specific semantic geometries on fixed CLIP-like representations, allowing similarity to reflect task-relevant attributes (e.g., dog breed vs. background) instead of a single fixed cosine geometry. Retrieval experiments show the approach is both effective and efficient for specifying context-dependent semantic similarity.

该论文系统研究了混合模态检索器的模态偏好问题，发现无关文本比无关图像对检索性能的干扰更严重（称为 "Chaos in the Text"），并提出 Trident 方法，通过 Multi-Positive View InfoNCE 联合优化文本、图像和融合视图来缓解该偏差。

The paper investigates whether text conditioning actually steers neural PDE surrogates using OperatorCLIP, a CLIP-style contrastively aligned FNO with FiLM control, finding that constant-prompt controls often match or beat task-text conditioning and that single-description training makes text alignment unidentifiable.