{ "schema_version": "paper_public_manifest_v1", "paper_id": "omega-eva_2026_08", "slug": "omega-eva", "title": "World Models: To Imagine or Not to Imagine at Inference Time", "authors": [], "source": { "arxiv_id": "2606.09457v2", "pdf_url": "https://arxiv.org/pdf/2606.09457v2", "project_url": "", "github_url": "", "huggingface_url": "", "original_source": "https://arxiv.org/abs/2606.09457v2" }, "site": { "post_url": "/posts/omega-eva", "canonical_url": "https://haiguangboy.com/posts/omega-eva", "cover_image": "https://static.haiguangboy.com/papers/omega-eva/cover.webp" }, "taxonomy": { "domain": "embodied_ai", "track": "world_model", "tasks": [ "embodied_ai", "world_model", "vla", "action_generation", "robotics", "latent_state", "state_prediction", "Embodied Intelligence", "World Models", "Paper Discussion" ], "related_topics": [ { "paper_id": "jepa_wam_learning_vision_language_action_policies_with_joint_embedding_world_modeling_2026_08", "title": "JEPA-WAM: An Innovative Approach That Makes World Models Better Understand State Changes", "url": "https://haiguangboy.com/posts/jepa-wam", "relation": "contrast", "summary": "Core claim: The world model's role is not to assist during training or serve as an external simulator, but to act as an internal validator during action generation contradicts Core claim: What should be predicted is the transition relationship, not the absolute future state", "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": "contrast", "summary": "LAWM: Why Action Labels Become a Burden", "strength": "strong" }, { "paper_id": "fast-wam_2026_07", "title": "No Future Imagination at Inference Time, Robot Still Reaches 91.8%! Fast-WAM Debunks WAM's Core Assumption", "url": "https://haiguangboy.com/posts/fast-wam", "relation": "contrast", "summary": "Training video objectives matter more than imagining the future at test time", "strength": "strong" }, { "paper_id": "tacwam_anchor_guided_world_action_model_with_mechanics_aware_tactile_prediction_2026_08", "title": "Tactile Prediction Should Only Be Used for Training, Not Shown to Actions", "url": "https://haiguangboy.com/posts/tacwam", "relation": "contrast", "summary": "Core claim: The world model's role is not to assist during training or serve as an external simulator, but to act as an internal validator during action generation contradicts Core principle: Future tactile signals serve only as training-time supervision, not as action inputs during deployment", "strength": "strong" }, { "paper_id": "leapbot_wa_world_anchor_action_models_via_predictive_latent_alignments_2026_07", "title": "Predictive Features Cannot Be Directly Fed to Diffusion Models", "url": "https://haiguangboy.com/posts/leapbot-wa", "relation": "contrast", "summary": "Problem breakdown: Existing world model usages fall into three categories, none of which allow candidate actions to be truly tested and corrected by their own imagined outcomes contradicts Key design: The world model only observes intent, not actions—deliberately depriving it of low-level motor commands to prevent collapse into a control shortcut", "strength": "strong" }, { "paper_id": "slim_05b_learning_action_grounded_predictive_latents_for_robot_manipulation_2026_08", "title": "Lower Prediction Error, Yet Representations Have Already Collapsed", "url": "https://haiguangboy.com/posts/slim-05b", "relation": "contrast", "summary": "Core claim: The world model's role is not to assist during training or serve as an external simulator, but to act as an internal validator during action generation contradicts Mechanism claim: One-way prediction is insufficient; representations must be constrained bidirectionally", "strength": "strong" } ] }, "ruling": { "importance_score": 3.0, "one_sentence": "SLIM and JEPA-WAM both say don't work at inference time—this ablation shows: removing that step makes the refiner worse than no refinement at all" }, "asset_base_url": "https://static.haiguangboy.com/papers/omega-eva", "assets": [ { "type": "public_brief", "object_key": "papers/omega-eva/public_brief.md", "content_type": "text/markdown; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/omega-eva/public_brief.md", "role": "public_brief", "size_bytes": 5853 }, { "type": "public_manifest", "object_key": "papers/omega-eva/public_manifest.json", "content_type": "application/json; charset=utf-8", "upload_status": "uploaded", "bucket": "paper-assets", "url": "https://static.haiguangboy.com/papers/omega-eva/public_manifest.json", "role": "public_manifest", "size_bytes": 5110 } ], "published_at": "2026-08-19T15:29:31+08:00", "created_at": "2026-08-19T15:29:31+08:00", "updated_at": "2026-09-02T10:55:06+08:00", "analyst_take": { "type": "author_opinion", "text": "This paper's core claim directly opposes the entire camp of SLIM, JEPA-WAM, and TacWAM—which argue the world model should only work during training and be fully removed at deployment—whereas this paper argues that a step should remain at inference time, just minimized to the cheapest form. Its own ablation provides hard evidence: removing the imagination branch makes the refiner worse than no refinement, not merely equal.\n\nBut the other half of its design aligns with the opposition: it provides the cleanest validation in the entire library for decoupling the current representation from candidate actions—this structural isolation of auxiliary signals is the same design philosophy shared by TacWAM and JEPA-WAM's backbone, differing only in what is isolated: those isolate the entire future branch, while this paper isolates only the current state, leaving an extremely lightweight imagination channel. The debate is not whether to have a world model, but to what level isolation should go.\n\nAnother recurring pattern: post-refinement fidelity metrics improve only slightly, and the authors refrain from drawing causal conclusions—this is the third time after SLIM and omega-0 that \"better prediction metrics\" and \"better control\" cannot be equated.\n\nAnchor for review in six months: if within-block closed-loop refinement is achieved, can the current modest gains of 0.7 to 2.2 percentage points be amplified to an order-of-magnitude level; if not, it indicates that the ceiling for a single ultra-lightweight imagination step lies around this range." } }