Many of language models' impressive capabilities originate from their in-context learning: based on instructions or examples, they can infer and perform new tasks without weight updates. In this work, we investigate when representations for new tasks are formed in language models, and how these representations change over the course of context. We study two different task representations: those that are ''transferrable'' -- vector representations that can transfer task contexts to another model instance, even without the full prompt -- and simpler representations of high-level task categories. We show that transferrable task representations evolve in non-monotonic and sporadic ways, while task identity representations persist throughout the context. Specifically, transferrable task representations exhibit a two-fold locality. They successfully condense evidence when more examples are provided in the context. But this evidence accrual process exhibits strong temporal locality along the sequence dimension, coming online only at certain tokens -- despite task identity being reliably decodable throughout the context. In some cases, transferrable task representations also show semantic locality, capturing a small task ''scope'' such as an independent subtask. Language models thus represent new tasks on the fly through both an inert, sustained sensitivity to the task and an active, just-in-time representation to support inference.

However, pre-trained VLA policies still suffer from substantial performance degradation during downstream deployment. Although fine-tuning can mitigate this issue, its reliance on costly demonstration collection and intensive computation makes it impractical in real-world settings. In this work, we introduce VLA-Pilot, a plug-and-play inference-time policy steering method for zero-shot deployment of pre-trained VLA without any additional fine-tuning or data collection. We evaluate VLA-Pilot on six real-world downstream manipulation tasks across two distinct robotic embodiments, encompassing both in-distribution and out-of-distribution scenarios. Experimental results demonstrate that VLA-Pilot substantially boosts the success rates of off-the-shelf pre-trained VLA policies, enabling robust zero-shot generalization to diverse tasks and embodiments. Experimental videos and code are available at: https://rip4kobe.github.io/vla-pilot/. I. INTRODUCTION Recent advances in VLA models have substantially improved the generalization capabilities of robotic manipulation. By learning from large-scale demonstrations [1], these generative foundation policies enable robots to acquire a wide repertoire of skills. At inference time, they can perform diverse and contextually appropriate tasks by stochastically sampling actions from the learned skill distribution.

We show that multi-agent systems guided by vision-language models (VLMs) improve end-to-end autonomous scientific discovery. By treating plots as verifiable checkpoints, a VLM-as-a-judge evaluates figures against dynamically generated domain-specific rubrics, enabling agents to correct their own errors and steer exploratory data analysis in real-time. Case studies in cosmology and astrochemistry demonstrate recovery from faulty reasoning paths and adaptation to new datasets without human intervention. On a 10-task benchmark for data-driven discovery, VLM-augmented systems achieve pass at 1 scores of 0.7-0.8, compared to 0.2-0.3 for code-only and 0.4-0.5 for code-and-text baselines, while also providing auditable reasoning traces that improve interpretability. Code available here: https://github.com/CMBAgents/cmbagent

Learned contexts were sequentially introduced, and novel context performance was assessed on the remaining (excluded from training) contexts.

Humans can efficiently transfer prior knowledge to novel contexts, an ability commonly referred to as transfer learning.

In reinforcement learning (RL), agents often struggle to perform well on tasks that differ from those encountered during training. This limitation presents a challenge to the broader deployment of RL in diverse and dynamic task settings. In this work, we introduce memory augmentation, a memory-based RL approach to improve task generalization. Our approach leverages task-structured augmentations to simulate plausible out-of-distribution scenarios and incorporates memory mechanisms to enable context-aware policy adaptation. Trained on a predefined set of tasks, our policy demonstrates the ability to generalize to unseen tasks through memory augmentation without requiring additional interactions with the environment. Through extensive simulation experiments and real-world hardware evaluations on legged locomotion tasks, we demonstrate that our approach achieves zero-shot generalization to unseen tasks while maintaining robust in-distribution performance and high sample efficiency.