# HyperWorld: Decompose, Generalize Away 31 Points

HyperWorld: Decompose, Generalize Away 31 Points

Full article: https://haiguangboy.com/posts/hyperworld_hypergraph_structured_state_serialization_improves_learned_textual_wo

## Core Method

Pack facts related to the same entity into a unified unit, rather than splitting them into independent pairwise relations
↳ Fixed information content is exactly identical, only organizational structure is compared: raw text, independent sentences, binary tuples, hyperedges (proposed here, rendering an entity's location/state/contents/binding relations all on the same line)—the four representations carry identical information, with differences only in how facts are grouped
↳ Small language models (0.5B-3B) undergo LoRA fine-tuning, learning to predict symbolic effects from serialized state + actions, with the same training objective, hyperparameters, and data across all four representations

## Key Results

· Hyperedge representation is strongest on 0.5B-1.5B and most out-of-distribution metrics: at 1.5B scale, out-of-distribution exact match reaches 0.914, 7.6 percentage points higher than binary tuples and 31 points higher than raw text, stable across three random seeds
· But not a comprehensive win—at 3B scale, in-distribution testing, binary tuples achieve slightly higher exact match (0.956 vs. 0.952); for action feasibility judgment alone, raw text is strongest, with hyperedges second but clearly better than binary tuples/sentences
· Tested with a simple greedy planner: hyperedge world model achieves 76.7% planning success, far exceeding binary tuples (56.7%) and sentences (53.3%)—more accurate single-step prediction indeed translates into higher downstream task completion
· Unexpected finding: removing the confidence weighting term from the planning scoring formula boosts success from 76.7% to 93.3%, indicating the default confidence calibration actually backfires

## Related Work Comparison

- Same direction · [S1: GPT-3 Emergent Moment for Robots](https://haiguangboy.com/posts/skild_blog_introducing_s1_in_context_learning_20260817) `skild_blog_introducing_s1_in_context_learning_20260817_2026_08`: S1: GPT-3 Emergent Moment for Robots
- Same direction · [sunday_blog_20260717](https://haiguangboy.com/posts/sunday_blog_20260717) `sunday_blog_20260717_2026_07`: Larger model scale dilutes the advantage of structured input (at 3B, binary tuples can catch up or even surpass hyperedges); with only 10% training data, all four representations perform similarly out-of-distribution, with structural advantages only emerging once data exceeds 25% validates pretraining scale convergence generalization gap: 82→10→8→4→0pp, out-of-domain 14%→100%
- Opposite direction · [1 Day, 1000 Tasks, Relying on Inductive Bias](https://haiguangboy.com/posts/mt3-thousand-tasks) `learning_a_thousand_tasks_in_a_day_2026_08`: Does not imply structured input is useful at any data volume—with only 10% training data, all four serialization methods perform similarly out-of-distribution, and structural advantages do not hold in extreme low-data scenarios contradicts key reverse trend: decomposition methods ramp up quickly at 1–10 examples but plateau around 50, while monolithic BC accelerates catch-up in the 10–50 range
- Same direction · [PRISM: Can Scale Replace Explicit Modeling?](https://haiguangboy.com/posts/prism_precision_and_contact_rich_real_world_industrial_skill_dataset_with_multim) `prism_precision_and_contact_rich_real_world_industrial_skill_dataset_with_multim_2026_08`: PRISM: Can Scale Replace Explicit Modeling?
- Same direction · [The Problem with Heterogeneous Data Isn't Volume, It's Inconsistent Supervision Formats](https://haiguangboy.com/posts/joyai-ra-dual-alignment) `joyai_ra_05_scaling_robot_manipulation_learning_via_dual_action_alignment_2026_08`: Integrating the learned world model into a simple greedy planner, tested on 30 unseen real games: hyperedge world model achieves 76.7% planning success, far exceeding binary tuples (56.7%) and sentences (53.3%), with fewer average steps validates world model ablation: with/without WM scores 48.4 vs. 51.5; switching to latent action-conditioned version (LAC-WM) improves from 87.3 to 92.1
- Same direction · [Fusing RGB, 3D Geometry, and Object Semantics into One Representation—This Is the Root Cause of Its Superior Accuracy Over π0.5 and Fast-WAM](https://haiguangboy.com/posts/flex_pi_a_multi_stream_world_action_model_with_compute_flexibility) `flex_pi_a_multi_stream_world_action_model_with_compute_flexibility_2026_08`: FLEX-π: A More Comprehensive Joint Representation

## Limitations

· Validated only on TextWorld mini-games with 410 controlled parameters and 0.5B-3B small models; scaling to richer environments and embodied robot manipulation is future work
· Not useful at any data volume—with only 10% training data, all four representations perform similarly out-of-distribution, with structural advantages only emerging once sufficient data accumulates
· Hyperedge grouping rules are deterministic rules manually designed by researchers, not learned by the model, and whether the model can autonomously discover similar good groupings remains unverified

## Author's Assessment (Non-Paper Content, Cross-Paper Synthesis)

"The advantage of structure/methods requires a data volume threshold to manifest"—this judgment has been validated again with HyperWorld, consistent with several previously reviewed works: Skild admits ICL underperforms traditional fine-tuning in low-data regimes; FLEX-π claims it remains "data-hungry"; PRISM's pretraining benefits are only evident on full data. Language models, robot manipulation, and textual world models—three vastly different domains—repeatedly yield the same pattern: it's not that "good structure always wins," but that "once data accumulates sufficiently, structurally superior approaches can realize their advantages." Additionally, this paper has an attitude worth noting—the authors do not package hyperedges as comprehensively optimal, honestly reporting that it is not strongest at 3B in-distribution or for pure feasibility judgment; such restraint is more trustworthy than many "all-win narrative" papers.