{ "schema_version": "paper_public_manifest_v1", "paper_id": "decoding_task_progress_from_vla_representations_2026_09", "slug": "decoding_task_progress_from_vla_representations", "title": "Opening the VLA Black Box: Progress Signals", "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", "Embodied Intelligence", "Interpretability" ], "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": "Core finding: Task progress (the normalized proportion of remaining trajectory time) is both weakly and strongly decodable in the residual stream of π0.5, but it is not steerable—injecting an offset of 'progress has already advanced by 0.2' along the probe direction does not correspondingly change the model's progress readout in subsequent layers, and the distance between the injected action and the true counterfactual action remains large. This validates the empirical finding: there is a clear gap between the model 'understanding constraints' and 'acting according to constraints.'", "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": "pave_predictive_alignment_and_value_guided_evolution_for_world_action_policies_2026_09", "title": "PAVE: Failure Data Can Also Feed the Policy", "url": "https://haiguangboy.com/posts/pave_predictive_alignment_and_value_guided_evolution_for_world_action_policies", "relation": "same_track", "summary": "PAVE: Failure Data Can Also Feed the Policy", "strength": "strong" }, { "paper_id": "pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation_2026_08", "title": "PointWorld: Point Flows Unify States and Actions", "url": "https://haiguangboy.com/posts/pointworld_scaling_3d_world_models_for_in_the_wild_robotic_manipulation", "relation": "same_track", "summary": "PointWorld: Point Flows Unify States and Actions", "strength": "strong" }, { "paper_id": "causally_debiased_latent_action_model_for_embodied_action_conditioned_world_mode_2026_07", "title": "World Models Don't Obey—Latent Actions Are Contaminated", "url": "https://haiguangboy.com/posts/cd-lam", "relation": "same_track", "summary": "A naively trained progress probe achieves an in-domain goodness of fit as high as 0.95 (appearing very successful), but under language counterfactual tests (replacing the target object in the instruction) its output barely changes—revealing that this probe actually relies entirely on visual cues and does not use the language input at all. The in-domain high accuracy metric itself completely failed to catch this problem. This validates that pixel metrics cannot rank the quality of latent action spaces—this is the most transferable methodological claim in the entire paper.", "strength": "strong" }, { "paper_id": "blind_dexterity_whole_body_humanoid_manipulation_via_pure_proprioception_2026_09", "title": "Others Stack Sensors; This Paper Took Them All Apart", "url": "https://haiguangboy.com/posts/blind_dexterity_whole_body_humanoid_manipulation_via_pure_proprioception", "relation": "same_track", "summary": "Others Stack Sensors; This Paper Took Them All Apart", "strength": "strong" } ] }, "analyst_take": { "type": "author_opinion", "text": "\"Being readable does not mean it is actually used for decision-making\"—this judgment is repeatedly corroborated by several works already interpreted: What/Where/How found that video models can organize camera motion into coherent geometric trajectories yet fail to learn genuine physical reasoning; Blind Dexterity found that feeding a state estimator's predictions back into observations yields no stable benefit for control, but the estimator's own error does systematically decrease after contact, proving the information genuinely exists. Three papers hit the same wall from three different angles: a signal can be stably read out without it being the variable the model uses to act." }, "ruling": { "importance_score": 3.0, "one_sentence": "Opening the VLA Black Box: Progress Signals" }, "asset_base_url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations", "assets": [ { "type": "cover_image", "object_key": "papers/decoding_task_progress_from_vla_representations/cover.webp", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/cover.webp", "content_type": "image/webp", "upload_status": "pending", "role": "post_cover" }, { "type": "pdf_screenshot", "object_key": "papers/decoding_task_progress_from_vla_representations/page_01.webp", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/page_01.webp", "content_type": "image/webp", "upload_status": "pending", "role": "paper_first_page" }, { "type": "pdf_screenshot", "object_key": "papers/decoding_task_progress_from_vla_representations/key_figure.webp", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/key_figure.webp", "content_type": "image/webp", "upload_status": "pending", "role": "method_figure" }, { "type": "public_manifest", "object_key": "papers/decoding_task_progress_from_vla_representations/public_manifest.json", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "pending", "role": "public_manifest" }, { "type": "public_brief", "object_key": "papers/decoding_task_progress_from_vla_representations/public_brief.md", "url": "https://static.haiguangboy.com/papers/decoding_task_progress_from_vla_representations/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "pending", "role": "public_brief" } ], "published_at": "2026-09-14T08:33:17+08:00", "created_at": "2026-09-14T08:33:17+08:00", "updated_at": "2026-09-14T08:33:17+08:00" }