ConfigWM: task-conditioned perception keeps world models robust to distractors, +22.3% over strongest baseline
Yijun Shen · hf · 2026-10-22
Researchers from East China Normal University and HKUST propose ConfigWM, addressing the problem that world model encoders are non-adaptive and easily distracted by task-irrelevant information.
- Uses the task description as a configuration signal to steer visual embeddings into task-specific state spaces, preserving task-relevant information and suppressing irrelevant variations
- Action-conditioned dynamics are learned in this task-specific state space for planning
- Results: as environments grow noisier, existing methods decline sharply while ConfigWM stays stable, outperforming the strongest baseline by 22.3% on average; consistent gains in physical-quantity probing (+0.365 Pearson r) and distracted-to-clean retrieval (+75.5%)
More from Research
- Text2Room author Lukas Höllein defends PhD thesis on 3D world generation, now at Apple — angelaqdai · 2026-10-22
- Turing-complete malware found runnable in PyTorch's backward pass — cephaloform · 2026-10-22
- Nature paper converts subword LLMs to byte-level with under 1% of pretraining cost, yielding Bolmo — Justgototheeffinmoon · 2026-10-22
- Tsinghua paper: letting agents manage their own KV cache cuts cost by 27.7% — dair_ai · 2026-10-22
- Effective receptive field in DiTs mirrors 2017 CNN finding, CS professor notes — CSProfKGD · 2026-10-22
- Flint cuts bike-demand validation RMSLE 30.8% in 3 hours of autonomous model iteration — cong_ml · 2026-10-22