機構的解釈可能性と最適制御によるワールドアクションモデルの堅牢性向上
Steering Robustness into World Action Models via Mechanistic Interpretability and Optimal Control
機構的解釈可能性を用いてワールドアクションモデルの堅牢性を解析し、対比的活性化方向による学習不要のステアリングと、モデルベース最適制御によるフィードバック制御を提案して堅牢性を向上させた。
詳しい要約
1. どんなもの?
2. 先行研究と比べてどこがすごい?
3. 技術・手法の肝は?
4. どうやって有効だと検証した?
5. 議論はある?
6. 次に読むべき論文は?
※ AIが要旨から生成した要約です。正確性は原文をご確認ください。
著者: Jihoon Hong, Julian Skifstad, Qiyue Dai, Alice Chan, Glen Chou
分類: cs.RO, cs.AI, cs.LG, cs.SY, eess.SY, math.OC
原文アブストラクト
World Action Models (WAMs) enable semantically- and physically-informed control but are brittle under distribution shift. In this work, we use mechanistic interpretability to study how robustness-relevant perturbations are represented in WAM activation space. Comparing activations across successful and unsuccessful rollouts, we find some WAM architectures exhibit low-dimensional linear separability for robustness-critical features, while others do not. This motivates the use of contrastive activation directions for training-free WAM steering. We also show that local linearity in WAM activation dynamics enables efficient feedback steering via model-based optimal control, yielding World-Action Linear Quadratic Regulator (WA-LQR), a minimally-invasive reduced-order LQR controller. Via mechanistic evaluations, we predict strong steerability in the Cosmos-Policy and DiT4DiT models but weak steerability in LingBot-VA, consistent with steering intervention results. On Cosmos-Policy and DiT4DiT, WA-LQR generalizes contrastive directions to new tasks and improves robustness to camera, gripper, and visual-noise perturbations over unsteered and prompt steering baselines.
