Research map / Probabilistic and scientific modeling / Dynamical systems
Words in Motion: Extracting Interpretable Control Vectors for Motion Transformers
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Probes hidden states of multimodal motion transformers for interpretable motion features and builds PCA- or sparse-autoencoder-based control vectors that steer forecasts at inference and generalize zero-shot to shifted speed distributions.
Problem
Hidden states of transformer motion forecasters are hard to interpret and control; it is unclear whether interpretable motion features (speed, acceleration, direction, agent type) are linearly separable and functionally usable in their latent spaces.
Approach
Measure neural collapse toward quantized motion features with linear probes (detached from gradients), then fit control vectors from differences between hidden states of opposing features, and refine them by pooling sparse autoencoder codes before projecting back through the SAE decoder.
Key results
- SAE-128 control vectors reach Pearson 0.993, R2 0.984, S-idx 0.988 for speed steering on Waymo, versus 0.988/0.969/0.981 for plain PCA.
- On a Waymo validation split with speeds reduced by about 50%, SAE-128 control vector with tau=-50 lowers minADE from 3.271 to 1.174 and minFDE from 4.617 to 1.798.
- Linear probing accuracy for speed reaches 0.945 on Waymo but only 0.753 on AV2F, and the AV2F-trained model gives much lower R2 (0.275 vs 0.969).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply the SAE-128 speed control vector at τ=-50 | RedMotion · Waymo reduced-future-speed validation | minADE: 3.271 -> 1.174 | Helped |
| Apply the SAE-128 speed control vector at τ=-50 | RedMotion · Waymo reduced-future-speed validation | Overlap rate: 0.220 -> 0.174 | Helped |
| Optimize control vectors with SAE-128 instead of plain PCA | RedMotion speed control · Waymo relative speed changes ±50% | R²: 0.969 -> 0.984 | Helped |
| Train RedMotion on Waymo rather than AV2F using matched trajectory lengths | RedMotion speed control · Waymo speed-control linearity comparison | R²: 0.275 -> 0.969 | Helped |
| Add activation steering with control vectors | RedMotion inference · Waymo | Inference latency: 50.21\text{ms} -> 51.08\text{ms} | Hurt |
| Use sparse intermediate dimension 16 instead of 128 (ablation) | SAE control-vector optimization · Waymo speed control, relative speed changes ±50% | R²: 0.984 -> 0.770 | Hurt |
| Use JumpReLU rather than ReLU activation (ablation) | SAE control-vector optimization · Waymo speed control, relative speed changes ±50% | R²: 0.984 -> 0.986 | Helped |
| Replace SAE with KoopmanAE (ablation) | Speed control-vector optimization · Waymo speed control, relative speed changes ±50% | R²: 0.984 -> -0.057 | Hurt |
| Use hidden states from module 0 rather than module 2 (ablation) | SAE-128 speed control · Waymo | R²: 0.984 -> 0.654 | Hurt |
| Lower the high-speed threshold to 25 km/h (ablation) | SAE-128 speed control · Waymo | R²: 0.984 -> 0.987 | Helped |
| Raise the low-speed threshold to 25 km/h (ablation) | SAE-128 speed control · Waymo | R²: 0.984 -> -0.734 | Hurt |
Surprising findings
- Control vectors fit on the AV2F dataset with a 1.1 s past sequence reduce speed by at most 3%, showing that low probing accuracy limits controllability.
- The KoopmanAE achieves the highest straightness index but near-zero R2 (-0.057) because its tau range is much smaller.
Limitations
- Assumes a flat latent space and relies on vector arithmetic; latent curvature is not modeled.
- Neural collapse analysis is limited to linear probing accuracy and CDNV.
- Controllability depends on probing accuracy; datasets with lower separability (AV2F) yield weak control.
Experimental setup
Datasets: Argoverse 2 Forecasting Waymo Open Motion Dataset
Benchmarks: Argoverse 2 Forecasting Waymo Open Motion Dataset
Metrics: linear probing accuracy Pearson correlation R2 straightness index CDNV minADE minFDE Brier minADE Brier minFDE overlap rate
Baselines: plain PCA control vectors Koopman autoencoder Sparse MLPMixer SAE ConvSAE
Builds on: word2vec sparse autoencoders neural collapse JumpReLU Koopman autoencoder
Closest papers by meaning
- Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control (ICLR 2025 · similarity 0.86)
- Predictive Geometry of Hidden Trajectories in Transformers (NeurIPS 2026 · similarity 0.86)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.85)
- ScaMo: Exploring the Scaling Law in Autoregressive Motion Generation Model (CVPR 2025 · similarity 0.85)
- ProMotion: Prototypes As Motion Learners (CVPR 2024 · similarity 0.85)
- Event-Grounded Sparse Autoencoders for Vision-Language-Action Policies (NeurIPS 2026 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.