Public articles linked to the same research event.
arXiv This work builds EditHero, a benchmark of part-level 3D edit chains in which a data engine assembles library parts onto segmented host objects and a JSON log rebuilds the exact target state after every turn, covering 457 chains, 2755 edits and 252 objects with 3 to 30 turns per chain; under a self-rollout protocol it evaluates 4 non-agentic methods (PartFlow, Nano3D, 3DEditFormer, VoxHammer) and 6 LLM/VLM agents, finding that non-agentic methods often miss the requested change from the first turn and accumulate errors along a chain, while agents preserve untouched regions better (CC 0.95–0.98 versus at most 0.90) and most follow instructions better (IF up to 0.62 versus at most 0.36), but take about 1.5–6 minutes per edit.
This work builds EditHero, a benchmark of part-level 3D edit chains in which a data engine assembles library parts onto segmented host objects and a JSON log rebuilds the exact target state after every turn, covering 457 chains, 2755 edits and 252 objects with 3 to 30 turns per chain; under a self-rollout protocol it evaluates 4 non-agentic methods (PartFlow, Nano3D, 3DEditFormer, VoxHammer) and 6 LLM/VLM agents, finding that non-agentic methods often miss the requested change from the first turn and accumulate errors along a chain, while agents preserve untouched regions better (CC 0.95–0.98 versus at most 0.90) and most follow instructions better (IF up to 0.62 versus at most 0.36), but take about 1.5–6 minutes per edit.
This work builds EditHero, a benchmark of part-level 3D edit chains in which a data engine assembles library parts onto segmented host objects and a JSON log rebuilds the exact target state after every turn, covering 457 chains, 2755 edits and 252 objects with 3 to 30 turns per chain; under a self-rollout protocol it evaluates 4 non-agentic methods (PartFlow, Nano3D, 3DEditFormer, VoxHammer) and 6 LLM/VLM agents, finding that non-agentic methods often miss the requested change from the first turn and accumulate errors along a chain, while agents preserve untouched regions better (CC 0.95–0.98 versus at most 0.90) and most follow instructions better (IF up to 0.62 versus at most 0.36), but take about 1.5–6 minutes per edit.
This work builds EditHero, a benchmark of part-level 3D edit chains in which a data engine assembles library parts onto segmented host objects and a JSON log rebuilds the exact target state after every turn, covering 457 chains, 2755 edits and 252 objects with 3 to 30 turns per chain; under a self-rollout protocol it evaluates 4 non-agentic methods (PartFlow, Nano3D, 3DEditFormer, VoxHammer) and 6 LLM/VLM agents, finding that non-agentic methods often miss the requested change from the first turn and accumulate errors along a chain, while agents preserve untouched regions better (CC 0.95–0.98 versus at most 0.90) and most follow instructions better (IF up to 0.62 versus at most 0.36), but take about 1.5–6 minutes per edit.