{ "schema_version": "paper_public_manifest_v1", "paper_id": "alphanist_2026_10", "slug": "alphanist", "title": "Without teaching steps, how do assembly skills emerge from rewards?", "authors": [], "source": { "arxiv_id": "2610.13354", "pdf_url": "https://arxiv.org/pdf/2610.13354", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/abs/2610.13354" }, "site": { "post_url": "/posts/alphanist", "canonical_url": "https://haiguangboy.com/posts/alphanist", "cover_image": "https://static.haiguangboy.com/papers/alphanist/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "action_generation", "tasks": [ "embodied_ai", "action_generation", "robotics", "state_prediction", "Embodied Intelligence", "Robotics", "Reinforcement Learning", "Robotic Assembly", "alphanist" ], "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": "contrast", "summary": "The authors judge that heterogeneous tasks can positively transfer through shared terminal goals contradicts that joint co-training on heterogeneous human-robot data is a structurally suboptimal approach", "strength": "strong" }, { "paper_id": "decompose_and_reorganize_planning_with_primitives_and_visuomotor_policies_learne_2026_08", "title": "Switching logic should not be learned by the policy", "url": "https://haiguangboy.com/posts/dr-lfd", "relation": "contrast", "summary": "Coordinated behavior of returning after temporarily storing parts appears in multi-pair assembly contradicts the core claim: switching logic should be given to the planner rather than learned by the policy—this is a structural rebuttal of 'end-to-end scaling'", "strength": "strong" }, { "paper_id": "learning_a_thousand_tasks_in_a_day_2026_08", "title": "1,000 tasks in 1 day relies on inductive bias", "url": "https://haiguangboy.com/posts/mt3-thousand-tasks", "relation": "same_track", "summary": "The authors judge that heterogeneous tasks can positively transfer through shared terminal goals validates a key reverse trend: decomposition methods ramp up quickly at 1–10 tasks but plateau near 50, while monolithic BC accelerates its catch-up in the 10–50 range", "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": "Coordination cases do not establish general planning or generalization to unseen tasks validates that small-sample and best-of-N evaluation cannot be extrapolated to stable general capabilities", "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": "mt_wam_reorienting_the_one_pass_predictive_representation_toward_action_generati_2026_09", "title": "MT-WAM: How future supervision serves actions", "url": "https://haiguangboy.com/posts/mt-wam", "relation": "same_track", "summary": "Scaling ablations simultaneously change GPU, environment, and rollout batch validates that the causal contributions of supervision targets, capacity, and branching structure have not yet been fully disentangled", "strength": "strong" } ] }, "analyst_take": { "type": "author_opinion", "text": "💡 How does diversity translate into capability?\nWhat this paper supports is that shared goals, task diversity, initial-state coverage, and experience allocation work together to make some hard single-task cases solvable.\n\nThis is not yet a scaling law of 'the more tasks, the smarter.' What is more worth testing is whether positive transfer can persist as the task set expands, and whether these regrasping, contact-adjustment, and coordination behaviors can transfer to real robots." }, "ruling": { "importance_score": 3.0, "one_sentence": "Without teaching steps, how do assembly skills emerge from rewards?" }, "asset_base_url": "https://static.haiguangboy.com/papers/alphanist", "assets": [ { "type": "cover_image", "object_key": "papers/alphanist/cover.webp", "url": "https://static.haiguangboy.com/papers/alphanist/cover.webp", "content_type": "image/webp", "upload_status": "pending", "role": "post_cover" }, { "type": "pdf_screenshot", "object_key": "papers/alphanist/page_01.webp", "url": "https://static.haiguangboy.com/papers/alphanist/page_01.webp", "content_type": "image/webp", "upload_status": "pending", "role": "paper_first_page" }, { "type": "pdf_screenshot", "object_key": "papers/alphanist/key_figure.webp", "url": "https://static.haiguangboy.com/papers/alphanist/key_figure.webp", "content_type": "image/webp", "upload_status": "pending", "role": "method_figure" }, { "type": "public_manifest", "object_key": "papers/alphanist/public_manifest.json", "url": "https://static.haiguangboy.com/papers/alphanist/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "pending", "role": "public_manifest" }, { "type": "public_brief", "object_key": "papers/alphanist/public_brief.md", "url": "https://static.haiguangboy.com/papers/alphanist/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "pending", "role": "public_brief" } ], "published_at": "2026-10-15T16:19:47+08:00", "created_at": "2026-10-15T16:19:47+08:00", "updated_at": "2026-10-15T16:19:47+08:00" }