CausGround raises task success from 23.3% to 50.7% across six backbones

AIBAAI reports that CausGround raises average task success from 23.3% to 50.7% across six open-weight backbones, including Qwen3.5 and Gemma 4, without updating any weights. The post says the gains hold in simulated households and real-world benchmarks, and that CausGround keeps performance more robust than baseline planners on 11–22 step tasks.

BAAIOfficial

@BAAIBeijing

View on X

Part of a thread · earlier post

Across both simulated households and real-world benchmarks, CausGround delivers consistent, major leaps over vanilla planners, SelfReVision, and LookPlanGraph:

📊Universal Cross-Family Scaling:
Tested across 6 open-weight backbones (Qwen3.5-9B/27B/35B, InternVL3.5-8B/38B, Gemma 4 E4B/12B), average task success rate more than doubles from 23.3% to 50.7% without updating a single weight!

🛡️Long-Horizon Resilience:
While baseline planners quickly derail as action steps increase (11–22 steps), CausGround's causal dependencies keep long-horizon execution robust and goal-directed.

Source details

BAAI · x.comPublished Nov 5, 8:46 AM ET

Original title: Across both simulated households and real-world benchmarks, CausGround delivers consistent, major leaps over vanilla planners, SelfReVisi...

CausGround raises task success from 23.3% to 50.7% across six backbones

Topics