{ "schema_version": "paper_public_manifest_v1", "paper_id": "decoding_task_progress_from_vla_representations_2026_09", "slug": "decoding_task_progress_from_vla_representations", "title": "打开VLA黑箱:进度信号", "authors": [], "source": { "arxiv_id": "2608.13474", "pdf_url": "https://arxiv.org/pdf/2608.13474", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/abs/2608.13474" }, "site": { "post_url": "/posts/decoding_task_progress_from_vla_representations", "canonical_url": "https://haiguangboy.com/posts/decoding_task_progress_from_vla_representations", "cover_image": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "vla", "tasks": [ "embodied_ai", "vla", "action_generation", "robotics", "state_prediction", "具身智能", "可解释性" ], "related_topics": [ { "paper_id": "gemini_robotics_2_safety_2026_07", "title": "Gemini_Robotics_2_Safety", "url": "https://haiguangboy.com/posts/gemini_robotics_2_safety", "relation": "same_track", "summary": "核心发现:任务进度(轨迹剩余时间的归一化占比)在π0.5的残差流里既弱可解码又强可解码,但不可操控——沿探针方向注入'已经推进了0.2进度'的偏移量,模型后续层的进度读数没有相应改变,注入动作和真实反事实动作的距离依然很远 validates 实测发现:模型「懂约束」和「按约束行动」之间存在明显落差", "strength": "strong" }, { "paper_id": "latent_action_pretraining_through_world_modeling_2026_07", "title": "LAWM：为什么动作标签会成为负担", "url": "https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling", "relation": "same_track", "summary": "LAWM：为什么动作标签会成为负担", "strength": "strong" }, { "paper_id": "pave_predictive_alignment_and_value_guided_evolution_for_world_action_policies_2026_09", "title": "PAVE:失败数据也能喂饱策略", "url": "https://haiguangboy.com/posts/pave_predictive_alignment_and_value_guided_evolution_for_world_action_policies", "relation": "same_track", "summary": "PAVE:失败数据也能喂饱策略", "strength": "strong" }, { "paper_id": "pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation_2026_08", "title": "PointWorld:点流统一状态动作", "url": "https://haiguangboy.com/posts/pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation", "relation": "same_track", "summary": "PointWorld:点流统一状态动作", "strength": "strong" }, { "paper_id": "causally_debiased_latent_action_model_for_embodied_action_conditioned_world_mode_2026_07", "title": "世界模型不听话，是潜在动作被污染了", "url": "https://haiguangboy.com/posts/cd-lam", "relation": "same_track", "summary": "朴素训练出的进度探针域内拟合优度高达0.95(看起来很成功),但语言反事实测试(把指令里的目标物体换掉)下输出几乎不变——暴露出这个探针其实完全依赖视觉线索、根本没在用语言输入,域内高准确率这个指标本身完全没能发现这个问题 validates 像素指标排不出潜在动作空间的好坏——这是整篇最可迁移的方法论主张", "strength": "strong" }, { "paper_id": "blind_dexterity_whole_body_humanoid_manipulation_via_pure_proprioception_2026_09", "title": "别人堆传感器,这篇全拆了", "url": "https://haiguangboy.com/posts/blind_dexterity_whole_body_humanoid_manipulation_via_pure_proprioception", "relation": "same_track", "summary": "别人堆传感器,这篇全拆了", "strength": "strong" } ] }, "analyst_take": { "type": "author_opinion", "text": "\"读得出来不等于真被用来做决策\"这个判断，和已经解读过的几篇工作反复印证：What/Where/How发现视频模型能把相机运动组织成连贯的几何轨迹，却学不到真正的物理推理；Blind Dexterity发现把状态估计器的预测接回观测对控制没有稳定收益，但估计器本身的误差确实在接触后系统性下降，证明信息真实存在。三篇论文从三个不同角度撞上了同一堵墙：一个信号能被稳定读出，不代表它是模型用来行动的那个变量。" }, "ruling": { "importance_score": 3.0, "one_sentence": "打开VLA黑箱:进度信号" }, "asset_base_url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations", "assets": [ { "type": "pdf_screenshot", "object_key": "papers/decoding_task_progress_from_vla_representations/page_01.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/page_01.webp", "role": "paper_first_page", "size_bytes": 170164 }, { "type": "pdf_screenshot", "object_key": "papers/decoding_task_progress_from_vla_representations/key_figure.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/key_figure.webp", "role": "method_figure", "size_bytes": 169524 }, { "type": "cover_image", "object_key": "papers/decoding_task_progress_from_vla_representations/cover.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/cover.webp", "role": "post_cover", "size_bytes": 88302 }, { "type": "public_brief", "object_key": "papers/decoding_task_progress_from_vla_representations/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/public_brief.md", "role": "public_brief", "size_bytes": 5204 }, { "type": "public_brief", "object_key": "papers/decoding_task_progress_from_vla_representations/public_brief.en.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/public_brief.en.md", "role": "public_brief_en", "size_bytes": 6510 }, { "type": "public_manifest", "object_key": "papers/decoding_task_progress_from_vla_representations/public_manifest.en.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/public_manifest.en.json", "role": "public_manifest_en", "size_bytes": 7556 }, { "type": "public_manifest", "object_key": "papers/decoding_task_progress_from_vla_representations/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/public_manifest.json", "role": "public_manifest", "size_bytes": 8139 } ], "published_at": "2026-09-14T08:33:17+08:00", "created_at": "2026-09-14T08:33:17+08:00", "updated_at": "2026-09-14T08:33:29+08:00" }