{
  "schema_version": "paper_public_manifest_v1",
  "paper_id": "slim-05b_2026_08",
  "slug": "slim-05b",
  "title": "Lower prediction error, but the representation has already collapsed",
  "authors": [],
  "source": {
    "arxiv_id": "2608.09771",
    "pdf_url": "https://arxiv.org/pdf/2608.09771",
    "project_url": "",
    "github_url": "",
    "huggingface_url": "",
    "original_source": "https://arxiv.org/abs/2608.09771"
  },
  "site": {
    "post_url": "/posts/slim-05b",
    "canonical_url": "https://haiguangboy.com/posts/slim-05b",
    "cover_image": "https://static.haiguangboy.com/papers/slim-05b/cover.webp"
  },
  "taxonomy": {
    "domain": "embodied_ai",
    "track": "world_model",
    "tasks": [
      "embodied_ai",
      "world_model",
      "vla",
      "action_generation",
      "robotics",
      "state_prediction",
      "Embodied Intelligence",
      "World Models",
      "Paper Discussion"
    ],
    "related_topics": [
      {
        "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": "Stage 2 completely removes future latent variables, leaving only the predictive slots shaped by Stage 1, contradicting the coupling mechanism: motion queries instead attend to video queries, injecting predicted future visual dynamics into action representations",
        "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": "MoT dual-stream: observation and action streams use joint attention, with language only doing per-stream cross-attention, contradicting AGT attention: action tokens are not allowed to read future visual/tactile prediction tokens, which is an information isolation mechanism rather than action-conditioned dynamics",
        "strength": "strong"
      },
      {
        "paper_id": "leapbot_wa_world_anchor_action_models_via_predictive_latent_alignments_2026_07",
        "title": "Predictive features cannot be directly fed into diffusion models",
        "url": "https://haiguangboy.com/posts/leapbot-wa",
        "relation": "contrast",
        "summary": "Defining the objective as action-grounded: able to infer actions from observation changes and predict latent transitions from actions, contradicting the key design: the world model only sees intent, not actions—deliberately depriving low-level motor commands to prevent collapsing into a control shortcut",
        "strength": "strong"
      },
      {
        "paper_id": "dyna_blog_dyna_2_20260811_2026_08",
        "title": "First human-to-robot zero-shot transfer scaling law",
        "url": "https://haiguangboy.com/posts/dyna-2-scaling-law",
        "relation": "contrast",
        "summary": "Problem setup: VLA only implicitly learns the relationship between actions and observation changes, while pixel-level prediction wastes computation on appearance details, contradicting the route bet: the cause of emergent cross-embodiment transfer is world modeling (predicting future videos) itself, not just data volume—WAM architecture outperforms pure VLA",
        "strength": "strong"
      },
      {
        "paper_id": "jepa_wam_learning_vision_language_action_policies_with_joint_embedding_world_modeling_2026_08",
        "title": "JEPA-WAM: An innovative method making world models better understand state changes",
        "url": "https://haiguangboy.com/posts/jepa-wam",
        "relation": "same_track",
        "summary": "Defining the objective as action-grounded: able to infer actions from observation changes and predict latent transitions from actions, validating the 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": "same_track",
        "summary": "LAWM: Why action labels become a burden",
        "strength": "strong"
      }
    ]
  },
  "ruling": {
    "importance_score": 3.0,
    "one_sentence": "A world model predicting the future does not mean it can control—give actions the ability to predict consequences, give consequences the ability to infer actions, using only 0.47B"
  },
  "asset_base_url": "https://static.haiguangboy.com/papers/slim-05b",
  "assets": [
    {
      "type": "public_brief",
      "object_key": "papers/slim-05b/public_brief.md",
      "content_type": "text/markdown; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/slim-05b/public_brief.md",
      "role": "public_brief",
      "size_bytes": 5731
    },
    {
      "type": "public_manifest",
      "object_key": "papers/slim-05b/public_manifest.json",
      "content_type": "application/json; charset=utf-8",
      "upload_status": "uploaded",
      "bucket": "paper-assets",
      "url": "https://static.haiguangboy.com/papers/slim-05b/public_manifest.json",
      "role": "public_manifest",
      "size_bytes": 5168
    }
  ],
  "published_at": "2026-08-14T08:20:39+08:00",
  "created_at": "2026-08-14T08:20:39+08:00",
  "updated_at": "2026-09-02T10:55:06+08:00",
  "analyst_take": {
    "type": "author_opinion",
    "text": "The most counterintuitive finding is the EMA ablation. After removing the EMA target, the prediction MSE of future latent variables drops from 0.245 to 0.166—even lower. But the effective rank collapses from 61.28 to 13.95, and LIBERO-Plus drops 10.6 points. Looking only at the loss, one would conclude that removing EMA is better.\n\nThis is already the third work showing a separation between offline prediction metrics and real-robot control performance. LeapBot-WA's SIGReg ablation shows the same dimensional collapse: rank 38.1→92.3, success rate 42.5%→71.3%. And omega-0 is another kind of failure—after switching to the Wan encoder, offline future reconstruction is more accurate, but real-robot execution becomes sluggish and hesitant, without collapse there.\n\nSo the common conclusion is not that latent variable prediction always collapses, but that a single prediction loss is insufficient to judge whether representations serve control. Effective rank and energy concentration can only detect geometric degradation; control probes like action discriminability and closed-loop success rate must be added separately."
  }
}