苏瑞涵Ruihan Su

MoPE-JEPA

通用 VLM 对视频中的物理规律（动力学、属性、空间关系）理解有限，需要专门的物理视觉编码器来增强感知与推理。

General-purpose VLMs have limited understanding of physical laws in video (dynamics, properties, spatial relations), calling for a specialized physics-aware visual encoder.

在 VideoMAEv2 (ViT-B) 顶部 4 层引入 4 专家 top-2 MoE + JEPA 自监督预训练；8 层全解冻，双塔注入 Qwen3-VL-2B 后 SFT + GRPO。

Introduced 4-expert top-2 MoE in the top 4 layers of VideoMAEv2 (ViT-B) with JEPA self-supervised pre-training; all 8 layers unfrozen, dual-tower injection into Qwen3-VL-2B followed by SFT + GRPO.

一次路由、四层复用的高效稀疏策略；SIGReg 正则化防止表征坍塌；在 PhysBench 上评测动力学、属性、场景、空间关系四类物理推理。

Efficient one-shot routing reused across four layers; SIGReg regularization to prevent representation collapse; evaluated on PhysBench across dynamics, properties, scene, and spatial reasoning.

Reading, Not Reasoning

CoT SFT 将多步视觉推理压缩为单次 VLM 前向，准确率提升被解读为"内化了推理"——但 emitted chain 真的承载了计算吗？

CoT SFT compresses multi-step visual reasoning into a single VLM forward pass. The accuracy gain is read as internalized reasoning — but does the emitted chain actually carry the computation?

因果干预：篡改 CoT 中数值中间量（v→2v+7），以 shuffle/local-num 为无目标对照，force-continue 后追踪 snap vs follow 答案来源。

Causal intervention: corrupt numeric intermediates in the CoT (v→2v+7), use shuffle/local-num as no-target controls, then force-continue and track snap vs. follow answer sources.

CoT SFT 稳定提升准确率（ChartQA +6.5pp），但 F≤0、snap≈97%、follow≈1%——思维链不承载因果，模型在读图而非推理。

CoT SFT reliably improves accuracy (ChartQA +6.5pp), but F≤0, snap≈97%, follow≈1% — the chain of thought does not carry causality: the model is reading charts rather than reasoning.

Moji-Video

Flow matching 采样是确定性 ODE，与 GRPO 要求的多次采样随机性存在结构性冲突；通过 ODE-to-SDE 转换引入可控噪声，保持边际分布不变的同时满足组内相对优势估计的前提条件。

Flow matching sampling is a deterministic ODE, structurally conflicting with GRPO's requirement for stochastic multi-sampling; ODE-to-SDE conversion introduces controllable noise, preserving the marginal distribution while enabling group relative advantage estimation.

设计基于 Qwen3-VL-8B 转写 + Levenshtein 编辑距离的 OCR verifiable reward；独立搭建 FSDP + vLLM-Omni 异步 rollout 多卡训练链路；应用 Denoising Reduction（训练 10 步 / 验证 50 步）控制 rollout 成本。

Designed OCR verifiable reward based on Qwen3-VL-8B transcription + Levenshtein edit distance; independently built FSDP + vLLM-Omni async rollout multi-GPU training pipeline; applied Denoising Reduction (10-step training / 50-step validation) to control rollout cost.

视觉理解算法Visual Understanding Algorithm

Unified Early-Fusion Transformer 共享图文表示；Global Presence Head 判定 concept 存在性后初始化 N 个并行 Slot，每个 Slot 内部依次自回归预测位置→尺寸→掩码。

Unified Early-Fusion Transformer for shared vision-language representation; Global Presence Head determines concept existence and initializes N parallel Slots, each autoregressively predicting coord → size → mask.

FlexAttention 稀疏块实现 Slot 隔离注意力掩码；Hungarian Matching 集合级监督解决 Slot 与 GT 实例的对应关系；Gated Cross-Slot Refinement 选择性跨 Slot 交互。

FlexAttention sparse blocks for Slot isolation attention masks; Hungarian Matching set-level supervision for Slot–GT instance assignment; Gated Cross-Slot Refinement for selective inter-Slot interaction.

模型已投入大规模预训练；架构原生支持扩展至 Slot-level Pass@K 与 GRPO/RL 后训练等生成式增强策略。

Model deployed for large-scale pre-training; architecture natively supports extension to Slot-level Pass@K and GRPO/RL post-training strategies.