{ "schema_version": "paper_public_manifest_v1", "paper_id": "cd-lam_2026_07", "slug": "cd-lam", "title": "World models misbehave because latent actions are contaminated", "authors": [], "source": { "arxiv_id": "2607.09185", "pdf_url": "https://arxiv.org/pdf/2607.09185", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/abs/2607.09185" }, "site": { "post_url": "/posts/cd-lam", "canonical_url": "https://haiguangboy.com/posts/cd-lam", "cover_image": "https://static.haiguangboy.com/papers/cd-lam/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "world_model", "tasks": [ "embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "Embodied intelligence", "World models", "Robot learning" ], "related_topics": [ { "paper_id": "an_open_foundation_model_towards_2026_07", "title": "An_Open_Foundation_Model_Towards", "url": "https://haiguangboy.com/posts/an_open_foundation_model_towards", "relation": "same_track", "summary": "Efficiency and data stratification: over 12x fewer adaptation updates; 1 hour of debiased data already yields 80% of the benefit, validating Ψ0's extreme data efficiency: 800h+30h surpasses a 10x larger data baseline by 40%+", "strength": "strong" }, { "paper_id": "latent_action_pretraining_through_world_modeling_2026_07", "title": "LAWM: Why action labels become a burden", "url": "https://haiguangboy.com/posts/latent_action_pretraining_through_world_modeling", "relation": "same_track", "summary": "LAWM: Why action labels become a burden", "strength": "strong" }, { "paper_id": "robointer15_a_holistic_intermediate_representation_suite_for_embodied_world_mode_2026_07", "title": "It's not \"whether to have a world model,\" but \"whether what it outputs has structure\"", "url": "https://haiguangboy.com/posts/robointer15_a_holistic_intermediate_representation_suite_for_embodied_world_mode", "relation": "same_track", "summary": "It's not \"whether to have a world model,\" but \"whether what it outputs has structure\"", "strength": "strong" }, { "paper_id": "t_rex_tactile_reactive_dexterous_manipulation_2026_07", "title": "T-Rex: Why touch needs separate modeling", "url": "https://haiguangboy.com/posts/t_rex_tactile_reactive_dexterous_manipulation", "relation": "same_track", "summary": "T-Rex: Why touch needs separate modeling", "strength": "medium" }, { "paper_id": "orca_2026_07", "title": "π0.5 trembles in place when failing to grab a spoon, yet Orca goes further with physical intuition learned from watching videos", "url": "https://haiguangboy.com/posts/orca", "relation": "same_track", "summary": "The key to a world model is readable states", "strength": "medium" }, { "paper_id": "wx_elsewhere别处发生_20260722_2026_07", "title": "Transcript of Liang Wenfeng's four-hour investor meeting", "url": "https://haiguangboy.com/posts/liangwenfeng-world-model", "relation": "same_track", "summary": "Three-stage pipeline: first debias LAM (1k steps) → then debias ACWM (2k steps) → finally connect an executable action bridge, validating ★★core judgment: the implicit premise of \"world models are irrelevant to intelligence ceilings\" is that \"the data pipeline is clear\"—for LLMs it's clear, so it's a detour; for robots it's not, so it's a bridge", "strength": "medium" } ] }, "analyst_take": { "type": "author_opinion", "text": "Reading the three papers side by side reveals three layers of the same issue. The outermost layer is an existing judgment: whether world models are useful is the wrong question—conditional signals and output structure determine whether they help or hinder. The middle layer is another paper's empirical finding: unstructured latent variables fed to a VLA cause performance drops, while structured intermediate representations yield stable gains. This paper fills in the innermost layer—why the drop happens and how to fix it.\n\nMore noteworthy is the counterintuitive finding: as the baseline scales from 2B to 14B, pixel metrics improve while action following degrades. Stacking parameters amplifies visual capability, not controllability. This parallels \"stricter segmentation beats looser segmentation\"—alignment quality determines whether data becomes capability.\n\nThere's also a cross-domain easter egg: the camera-conditioning paper found that backgrounds leak camera pose, making apparent viewpoint invariance actually a shortcut. The same mechanism strikes twice in unrelated areas—whenever a network is left to infer some quantity from pixels on its own, it's worth checking whether it's just copying the background." }, "ruling": { "importance_score": 3.0, "one_sentence": "World models misbehave because latent actions are contaminated" }, "asset_base_url": "https://static.haiguangboy.com/papers/cd-lam", "assets": [ { "type": "cover_image", "object_key": "papers/cd-lam/cover.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/cd-lam/cover.webp", "role": "post_cover", "size_bytes": 78214 }, { "type": "public_brief", "object_key": "papers/cd-lam/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/cd-lam/public_brief.md", "role": "public_brief", "size_bytes": 4318 }, { "type": "public_manifest", "object_key": "papers/cd-lam/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/cd-lam/public_manifest.json", "role": "public_manifest", "size_bytes": 5944 } ], "published_at": "2026-07-29T20:37:27+08:00", "created_at": "2026-07-29T20:37:27+08:00", "updated_at": "2026-09-02T10:55:06+08:00" }