{ "schema_version": "paper_public_manifest_v1", "paper_id": "zprl_2026_08", "slug": "zprl", "title": "At Which Layer Should RL Intervene", "authors": [], "source": { "arxiv_id": "2605.19919", "pdf_url": "https://arxiv.org/pdf/2605.19919", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/abs/2605.19919" }, "site": { "post_url": "/posts/zprl", "canonical_url": "https://haiguangboy.com/posts/zprl", "cover_image": "https://static.haiguangboy.com/papers/zprl/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "action_generation", "tasks": [ "embodied_ai", "action_generation", "robotics", "latent_state", "state_prediction", "Embodied Intelligence", "Reinforcement Learning", "Robot Policy" ], "related_topics": [ { "paper_id": "dexora_open_source_vla_for_high_dof_bimanual_dexterity_2026_08", "title": "Noise in Teleoperation Data Must Enter the Loss Function", "url": "https://haiguangboy.com/posts/dexora", "relation": "contrast", "summary": "Simulation: strong final performance across three benchmarks and eight tasks, covering parallel jaw grippers and dexterous hands, contradicts the simulation-to-dexterity gap: simulation can guide basic skills, but dexterity must rely on real robot data", "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": "causally_debiased_latent_action_model_for_embodied_action_conditioned_world_mode_2026_07", "title": "The World Model Is Disobedient Because Latent Actions Are Contaminated", "url": "https://haiguangboy.com/posts/cd-lam", "relation": "same_track", "summary": "Key ablation: the gain does not come solely from dimensionality reduction—dim(z)=64 already exceeds the action dimension yet is still faster and better, validating that scale cannot replace debiasing: baseline from 2B to 14B improves vision but worsens action following", "strength": "strong" }, { "paper_id": "pose_aware_modeling_to_mitigate_pose_related_artifacts_in_tactile_gloves_2026_07", "title": "Glove Tactile Readings Are Mixed with Hand Pose", "url": "https://haiguangboy.com/posts/tactile-glove-pose-artifacts", "relation": "same_track", "summary": "Controlled experiment: residual on observation embeddings directly (RESEMB) is consistently worse and becomes more unstable as residual scale increases, validating yet again that 'direct concatenation is worse'—this time the explanation is that the model learns the wrong mapping direction", "strength": "strong" }, { "paper_id": "x_manual_pi_rl_20260726_20260727_2026_07", "title": "The Bottleneck for Robot RL Is Sampling, Not Algorithms", "url": "https://haiguangboy.com/posts/pi-rl-chelsea-finn-talk", "relation": "same_track", "summary": "Online cost is not low: real-robot tasks require hours of interaction, and λ must be manually determined per task, validating as reported: the bottleneck for RL in robotics is physical rollout cost, not just algorithm quality", "strength": "strong" }, { "paper_id": "deepmind_blog_gemini_robotics_2_brings_20260730_2026_07", "title": "The Overlooked Non-Consensus Is Whole-Body Unified Control", "url": "https://haiguangboy.com/posts/gemini-robotics-2", "relation": "same_track", "summary": "Simulation: strong final performance across three benchmarks and eight tasks, covering parallel jaw grippers and dexterous hands, validating dexterity: simultaneously covers a 22-DOF five-finger hand and a standard two-finger gripper", "strength": "strong" } ] }, "ruling": { "importance_score": 3.0, "one_sentence": "At Which Layer Should RL Intervene" }, "asset_base_url": "https://static.haiguangboy.com/papers/zprl", "assets": [ { "type": "cover_image", "object_key": "papers/zprl/cover.webp", "content_type": "image/webp", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/zprl/cover.webp", "role": "post_cover", "size_bytes": 86024 }, { "type": "public_brief", "object_key": "papers/zprl/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/zprl/public_brief.md", "role": "public_brief", "size_bytes": 5046 }, { "type": "public_manifest", "object_key": "papers/zprl/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/zprl/public_manifest.json", "role": "public_manifest", "size_bytes": 5257 } ], "published_at": "2026-08-07T07:52:06+08:00", "created_at": "2026-08-07T07:52:06+08:00", "updated_at": "2026-09-02T10:55:06+08:00", "analyst_take": { "type": "author_opinion", "text": "This paper converges with another judgment: auxiliary signals cannot be plugged into the backbone without constraints; they need a structurally restricted position. Previous methods used residual terms, prediction targets, dedicated pathways, and attention masks; this one uses 'adding perturbations in a compressed task latent space.' The increment lies in three sets of cross-evidence that isolate 'not dimensionality reduction,' 'not arbitrary features,' and 'must be local.'\n\nThere is also a fresh divergence. The conclusion of the earlier two-arm two-hand work was that simulation can only guide basic skills, dexterity must rely on real robots, and pure simulation is near zero for dexterous tasks. This paper, however, achieves strong performance with RL fine-tuning in pure simulation on the Adroit dexterous hand. The difference may lie in whether a good base is available—one learns dexterity from scratch, the other guides on top of existing priors." } }