{ "schema_version": "paper_public_manifest_v1", "paper_id": "mt-wam_2026_09", "slug": "mt-wam", "title": "MT-WAM: How Future Supervision Serves Action", "authors": [], "source": { "arxiv_id": "2609.21474", "pdf_url": "https://arxiv.org/pdf/2609.21474", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/abs/2609.21474" }, "site": { "post_url": "/posts/mt-wam", "canonical_url": "https://haiguangboy.com/posts/mt-wam", "cover_image": "https://static.haiguangboy.com/papers/mt-wam/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "world_model", "tasks": [ "embodied_ai", "world_model", "action_generation", "robotics", "state_prediction", "Embodied Intelligence", "Robotics", "World Models", "mtwam", "fastwam", "Representation Learning" ], "related_topics": [ { "paper_id": "openwam_an_open_modular_exploration_towards_systematic_world_action_model_pretra_2026_09", "title": "OpenWAM: The Right Answer Changes with Scale", "url": "https://haiguangboy.com/posts/openwam_an_open_modular_exploration_towards_systematic_world_action_model_pretra", "relation": "contrast", "summary": "OpenWAM: The Right Answer Changes with Scale", "strength": "strong" }, { "paper_id": "orca_2026_07", "title": "π0.5 Freezes in Place When It Fails to Grab a Spoon, While Orca Takes a Step Further with Physical Intuition Gained from Watching Videos", "url": "https://haiguangboy.com/posts/orca", "relation": "contrast", "summary": "The key to a world model is a readable state", "strength": "strong" }, { "paper_id": "omega_0_a_latent_predictive_world_action_model_for_concurrent_humanoid_loco_mani_2026_08", "title": "The More Accurate the Future Prediction, the More Sluggish the Robot Becomes", "url": "https://haiguangboy.com/posts/omega-0", "relation": "contrast", "summary": "Duplicating the end of the video backbone to form a dual-stream branch, rather than just adding an output head, contradicts the core design: future prediction is not a video generation objective, but rather a compact predictive signal reconstructed independently; the action branch does not need pixel-level generation or test-time video-to-action conversion", "strength": "strong" }, { "paper_id": "slim_05b_learning_action_grounded_predictive_latents_for_robot_manipulation_2026_08", "title": "Lower Prediction Error, Yet the Representation Has Already Collapsed", "url": "https://haiguangboy.com/posts/slim-05b", "relation": "contrast", "summary": "Retaining the FAST-WAM direct action interface and adding two types of future supervision contradicts Stage 2, which removes the future latent variables entirely, keeping only the predictive slot shaped by Stage 1", "strength": "strong" }, { "paper_id": "agile_wam_an_agile_tactile_world_action_model_for_contact_rich_robot_control_2026_09", "title": "Agile-WAM: Lightweight Visuotactile Joint Prediction", "url": "https://haiguangboy.com/posts/agile-wam", "relation": "same_track", "summary": "Retaining the FAST-WAM direct action interface and adding two types of future supervision validates the authors' claim: whichever modalities are used as input should also have their futures learned", "strength": "strong" }, { "paper_id": "jepa_wam_learning_vision_language_action_policies_with_joint_embedding_world_modeling_2026_08", "title": "JEPA-WAM: An Innovative Method That Makes World Models Better Understand State Changes", "url": "https://haiguangboy.com/posts/jepa-wam", "relation": "same_track", "summary": "The authors' judgment: after eliminating future generation, priority should be given to improving the current single forward-pass representation, which validates the core claim: transition supervision must directly shape the very backbone that generates actions; neither external modules nor serving as context will work", "strength": "strong" } ] }, "analyst_take": { "type": "author_opinion", "text": "💡 Placed in the context of related work\nTacWAM uses future touch for supervision, and in nested ablations finds that gradually opening up information interaction harms control; this echoes MT-WAM's asymmetric division of labor.\n\nBut OpenWAM finds that cutting off the action's access to world features costs about 5 points. The two do not directly negate each other: the world stream and the auxiliary visual feature stream are not the same kind of signal, and MT-WAM also retains the action's read access to the motion stream. What truly should be verified item by item in engineering is: which objective improves the representation, and which information pathway improves control." }, "ruling": { "importance_score": 3.0, "one_sentence": "MT-WAM: How Future Supervision Serves Action" }, "asset_base_url": "https://static.haiguangboy.com/papers/mt-wam", "assets": [ { "type": "cover_image", "object_key": "papers/mt-wam/cover.webp", "url": "https://static.haiguangboy.com/papers/mt-wam/cover.webp", "content_type": "image/webp", "upload_status": "pending", "role": "post_cover" }, { "type": "pdf_screenshot", "object_key": "papers/mt-wam/page_01.webp", "url": "https://static.haiguangboy.com/papers/mt-wam/page_01.webp", "content_type": "image/webp", "upload_status": "pending", "role": "paper_first_page" }, { "type": "pdf_screenshot", "object_key": "papers/mt-wam/key_figure.webp", "url": "https://static.haiguangboy.com/papers/mt-wam/key_figure.webp", "content_type": "image/webp", "upload_status": "pending", "role": "method_figure" }, { "type": "public_manifest", "object_key": "papers/mt-wam/public_manifest.json", "url": "https://static.haiguangboy.com/papers/mt-wam/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "pending", "role": "public_manifest" }, { "type": "public_brief", "object_key": "papers/mt-wam/public_brief.md", "url": "https://static.haiguangboy.com/papers/mt-wam/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "pending", "role": "public_brief" } ], "published_at": "2026-09-28T11:43:53+08:00", "created_at": "2026-09-28T11:43:53+08:00", "updated_at": "2026-09-28T11:43:53+08:00" }