AI Agents Get Smarter with New Techniques

The field of artificial intelligence (AI) has witnessed significant advancements in recent years, with the development of sophisticated models that can understand and interact with the world around them. However, these models still face challenges in complex tasks that require long-horizon manipulation and multi-stage control. To address these limitations, researchers have introduced new techniques to improve the performance of AI agents, including vision-language models (VLMs) and embodied agents.

One of the key challenges in VLMs is the lack of persistent, action-conditioned state representations, which makes them ill-suited for multi-stage control. To overcome this limitation, researchers have proposed a recursive belief vision language model (RB-VLA) that maintains a compact latent state encoding task-relevant history, dynamics, and object interactions [1]. This model is trained with self-supervised world-model objectives and can be queried once for high-level intent, providing task specification and tracking task progress.

Another challenge in embodied agents is the reliance on high-level commands or discretized action spaces, which are non-native settings that differ markedly from real-world control. To address this limitation, researchers have introduced NativeEmbodied, a challenging benchmark for VLM-driven embodied agents that uses a unified, native low-level action space [2]. This benchmark includes three representative high-level tasks in complex scenarios and four types of low-level tasks, each targeting a fundamental embodied skill.

In addition to these advancements, researchers have also developed methods to enhance the behavior of AI agents at test time. One such method is Polarity-Prompt Contrastive Decoding (PromptCD), which constructs paired positive and negative guiding prompts for a target behavior and contrasts model responses to reinforce desirable outcomes [3]. This method can be used to improve the behavior of large language models (LLMs) and VLMs in a variety of settings.

However, AI agents are not immune to attacks, and researchers have identified a new type of attack called indirect prompt injection (IPI) attacks. These attacks involve malicious instructions in retrieved content that hijack the agent's execution. To defend against these attacks, researchers have proposed ICON, a probing-to-mitigation framework that neutralizes attacks while preserving task continuity [5]. This framework uses a Latent Space Trace Prober to detect attacks and a Mitigating Rectifier to selectively manipulate adversarial query key dependencies.

Finally, researchers have also explored the challenges of online decision making with unreliable guidance. In this setting, an algorithm receives guidance that is corrupted with probability β, and the goal is to develop algorithms that admit good competitiveness when β = 0 (consistency) as well as when β = 1 (robustness) [4]. This problem is formulated through the lens of request-answer games, and researchers have proposed online algorithms with unreliable guidance (OAG) that can achieve good performance in this setting.

In conclusion, the field of AI has witnessed significant advancements in recent years, with the development of new techniques to improve the performance of VLMs and embodied agents. These techniques include recursive belief vision language models, NativeEmbodied benchmarks, Polarity-Prompt Contrastive Decoding, ICON defense against IPI attacks, and online algorithms with unreliable guidance. These advancements have the potential to improve the performance of AI agents in a variety of settings and to enable more sophisticated applications of AI in the future.

References:

[1] Recursive Belief Vision Language Model (arXiv:2602.20659v1)

[2] How Foundational Skills Influence VLM-based Embodied Agents: A Native Perspective (arXiv:2602.20687v1)

[3] PromptCD: Test-Time Behavior Enhancement via Polarity-Prompt Contrastive Decoding (arXiv:2602.20696v1)

[4] Online Algorithms with Unreliable Guidance (arXiv:2602.20706v1)

[5] ICON: Indirect Prompt Injection Defense for Agents based on Inference-Time Correction (arXiv:2602.20708v1)