cs.CVNov 12, 2026

Do the Details Help the Whole? One-Way Transfer from Person Actions to Group Intent in VLMs on JRDB-Social

Authors: Chengguang Gan, Zhixi Cai, Yunhao Liang, Hanjun Wei, QingHao Zhang, Shiwen Ni

Organizations: Independent Researcher · Monash University · University of Chinese Academy of Sciences · Pusan National University · 

Abstract

Social scene datasets label people at two levels: what each person is doing and what the group is there for. Training on both levels at once is often assumed to help each of them. We test this assumption for vision-language models (VLMs) on JRDB-Social, treating person actions as the details and group intent as the whole. We fine-tune Qwen3.5 4B under eight training arms with three seeds, repeat the core arms on Qwen3.5 9B and Qwen3-VL 4B, and judge every arm with a verdict rule fixed before testing. Writing each member's actions before the intent raises intent macro-F1 from 30.1 to 31.0, and writing the intent first reaches 31.2, while action F1 stays within 0.2 points of an action-only model. Under the selected recipe, no arm trained with intent improves actions. The gain requires correct, image-grounded actions: empty action lists or actions borrowed from another group with the same intent lower intent F1. Forced-prefix tests, occlusion, member masking, linear probes and component grafting show that the intent follows the written actions, relies more on people and partly resides in the vision tower. Supplying actions as input instead of as targets hurts both levels.

Figures & tables

Appendix figures & tables18 assets

Supplementary material from the paper’s appendix.

Appendix

Explore similar work

Aug 24, 2026

Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

Vision-Language-Action (VLA) models can turn multimodal context into robot actions, but their action decoders are still trained largely by behavior cloning. This supervises which motor command was demonstrated while leaving implicit the local objective served by the behavior under the instruction. Future-based supervision enriches action learning with frames, latent observations, trajectories, or motion representations, but these signals capture particular realizations of what may happen rather than the shared semantic objective of the forthcoming behavior. We propose Intention Distillation (INDI), which distills behavior-level intent into the action decoder. During training, a frozen teacher VLM interprets a demonstrated segment from the current observation, instruction, coarse action summary, and corresponding execution video. From its standard inputs, the deployed VLA recovers the resulting multimodal intent representation at an intermediate decoder layer and uses it to organize action prediction together with representations of how the behavior unfolds and what it achieves. On SimplerEnv-Bridge, INDI improves GR00T-N1.7 from 64.3% to 84.7%, and on RoboCasa Kitchen it improves the controlled GR00T-N1.7 baseline from 64.1% to 70.3%, with consistent gains on π0.5π_{0.5} across both benchmarks. In real-world tasks, INDI improves average success from 62.0% to 68.7%, with gains of up to 12.0 pp on longer-horizon tasks. Further analyses show that the recovered latent is used by the decoder, captures behavior objective and execution progress, and organizes downstream predictions in an objective-dependent manner. These results show that action decoders benefit from explicitly modeling the semantic objective of the behavior they generate.
Oct 17, 2026

SocialCoAct: Enabling and Benchmarking Egocentric Social Co-Action for VLM-Based Humanoid Robots

Enabling humanoid robots to participate effectively in social interactions is important for their integration into human-centered environments. Existing methods, however, typically map pre-generated speech directly to continuous motion trajectories but have limited awareness of complex social environments and lack explicit high-level semantic action planning. We introduce SocialCoAct, a unified agent-centered framework for VLM-based egocentric social co-action. We formulate egocentric social co-action as an end-to-end agent task and develop SocialCoAct-Agent, which transforms egocentric observations and dialogue history into verbal responses and atomic action sequences, aligns each action with its corresponding utterance span, and executes the resulting behaviors in MuJoCo and on a physical humanoid robot. To evaluate the VLM capabilities required by this task, we construct SocialCoAct-Bench from continuous egocentric recordings of natural multi-party interactions. The benchmark comprises 423 human-verified multiple-choice questions across three subtasks: Long Episode Understanding (LEU), Speaker Intention Understanding (SIU), and Coordinated Action Planning (CAP). Across a diverse set of closed-source and open-weight VLMs, including general and egocentric models, the highest overall accuracy is 77.54%, while accuracy on the hard subset reaches only 46.51%. Extended reasoning improves Qwen3-VL-2B and Qwen3-VL-4B by 19.62 and 17.02 percentage points, respectively, but reduces inference throughput by 53.6×\times and 20.2×\times. Visual-input ablations further show that the benefits of visual context vary across models. SocialCoAct provides a foundation for developing and evaluating VLM-based humanoid agents that coordinate context-aware speech and action in complex social interactions. The codebase and benchmark will be made publicly available upon publication.
Mar 19, 2026

Not All Features Are Created Equal: A Mechanistic Study of Vision-Language-Action Models

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 π0.5π_{0.5} execute task A's motor trajectory in 99.6% of episodes (n=1,968n{=}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 five VLAs on LIBERO, MetaWorld, SimplerEnv, and ALOHA over 420,000+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 (π0.5π_{0.5} 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://cwru-aism.github.io/vla-interp-page/