A fine-tuned Vision-Language-Action (VLA) policy can easily place the alphabet soup in the basket on command, just as it saw during fine-tuning. However, if the basket is shifted five centimeters to the left, it drops the soup off the table. We developed an interpretability pipeline to investigate what such a policy is computing. Injecting task A's activations into the action expert while the policy looks at task B's scene makes pi0.5 execute task A's motor trajectory in 99.6\% of episodes (n{=}1{,}968). The injected program is bound to absolute workspace coordinates and ignores the visible scene, which provides a mechanistic account of the perturbation brittleness that concurrent benchmarks report. The binding holds on a physical UR10e: directed replay redirects the commanded task in 26/26 rollouts in one direction and 18/26 in the other, with no-injection and random-direction controls at the floor, and once the source object is moved 10--15\,cm the arm grasps and releases at the recorded coordinates in 50/50 trials, picking up the wrong block when one sits there and closing on empty air when none does. Applying the same interventions to six VLAs on LIBERO, MetaWorld, SimplerEnv, and ALOHA over 420{,}000+ rollouts yields three further findings: every architecture encodes the instruction, and four of the five weight it below vision whenever the scene alone identifies the goal; under the standard replacement write-back the SAE pooling that preserves behavior is architecture-specific (pi0.5 per-token, SmolVLA indifferent), and goal selection and motor generation separate into VLM and expert pathways wherever the two are distinguishable. We release Action Atlas (https://action-atlas.com), an interactive platform covering all 520+ trained SAEs, 79 identified concepts, and the scored UR10e rollouts. Code: https://github.com/CWRU-AISM/action-atlas
Translation pending · showing the source descriptionThe chart will appear after repeat observations. The current metric comes from the source.
0 source votesReal observations only. History before source connection is not reconstructed.
Robotic value models assess task execution from visual observations and instructions, providing feedback for policy learning and execution monitoring. We introduce RoboValue, a fine-grained sim-and-real benchmark for unified evaluation of robotic value models. Its framework evaluates four complementary capability dimensions through shared interfaces: task-state understanding, temporal progress monitoring, failure and recovery reasoning, and value consistency. Diagnostic trajectories and instruction variations test whether value judgments accurately reflect task requirements and changes throughout execution.
3 days agoDaily PapersMemT-1.0 is a new architecture of memory foundation models that internalizes memory as part of model computation. It uses Gated DeltaNet-2 with a stable token-level memory update mechanism and Mixture-of-Storage to scale native memory. MemT-1.0-Nano and MemT-1.0-Flash models outperformed parametric memory baselines and were competitive with retrieval-augmented methods.
3 days agoDaily PapersThis work introduces Scene-Designer, an agentic framework that learns design intelligence through complementary design and critique experiences. Two datasets, Scene-OTA-35K and Scene-OTC-15K, were created for scene design and quality critique. A Scene-GRPO methodology was proposed to improve design decisions and reduce unnecessary corrections. Experiments demonstrate superior performance in physical validity, functional correctness, and layout coherence.
6 days agoDaily PapersThe paper introduces Video-Point Model (VPM), a world model that conditions on multi-view video histories and language instructions to jointly forecast future videos from arbitrary viewpoints and 3D trajectories of persistent scene points. The model extends a pretrained world model with camera-conditioned multi-view processing that integrates observations across cameras and predicts how physical points move over time.
2 days agoDaily PapersSpecMRL is an approach that restructures matryoshka representation learning using a parametric spectral space for more effective prefix readout. Experiments on MTEB show that SpecMRL outperforms MRL at all evaluated reduced dimensions, with larger gains under more aggressive compression.
3 days agoLessWrongAs our understanding of mechanistic interpretability and steering techniques increases for both LLMs and uploaded connectomes, there is a corresponding 'hellish capabilities' increase. Improved technologies allow us to torture digital minds we create. History with open-source technologies shows that torture often becomes one of the first use cases when the public gains access to a digital mind perceived as capable of suffering.
yesterday