RLDP Advances AI Agents While AsgardBench Enhances Embodied AI

Recent research explores enhancing AI agent capabilities through novel architectures and training methodologies. For instance, advancements in behavioral foundation models are being made with Regularized Latent Dynamics Prediction (RLDP), which improves zero-shot reinforcement learning by maintaining feature diversity, outperforming complex representation learning methods, especially in low-coverage scenarios. In embodied AI, AsgardBench offers a new benchmark for visually grounded interactive planning, highlighting weaknesses in current vision-language models' ability to adapt plans based on visual input. Multi-Agent Constitutional Learning (MAC) optimizes structured prompts using a network of agents, significantly outperforming existing prompt optimization methods and producing human-readable rule sets.

Safety and reliability are critical concerns, with formal proofs indicating that safety is non-compositional in conjunctive capability dependencies, meaning combined agents can exhibit emergent forbidden capabilities. This necessitates careful system design, as seen in the proposed formal framework for AI systems and the development of runtime governance policies that map agent behavior to violation probabilities. For LLM-enabled robots, bounded calibration with contestability offers a front-end pattern for assistance allocation, constraining prioritization and providing contest pathways without renegotiating global rules. Furthermore, research into visual distraction reveals that it can fundamentally alter moral decision-making in Vision-Language Models (VLMs), overriding deliberate reasoning pathways and highlighting the need for multimodal safety alignment.

Several papers address the challenge of knowledge representation and reasoning. Neural-symbolic models like HYQNET leverage hyperbolic space for logic query reasoning on knowledge graphs, offering interpretability and better capture of hierarchical structures. For text-to-SQL tasks, TRUST-SQL addresses unknown schemas by employing an autonomous agent with a structured protocol and a Dual-Track GRPO strategy, achieving significant improvements over base models. In the financial domain, an Option Query Language (OQL) is introduced to translate natural language trading intents into executable option strategies, improving execution accuracy and logical consistency.

Memory systems for AI agents are also a focus, with CraniMem proposing a neuro-cognitively motivated, gated, and bounded multi-stage memory design for improved robustness and consolidation. NextMem introduces a latent factual memory framework using an autoregressive autoencoder for efficient construction and accurate reconstruction. Compiled Memory (Atlas) focuses on memory utility by distilling accumulated experience into instruction rewrites rather than context injection, improving performance and reducing costs. Governed Memory offers a production architecture for multi-agent workflows, addressing memory silos and governance fragmentation with a dual memory model and tiered routing.

Evaluating and improving AI agent performance in complex domains is another key theme. RetailBench evaluates long-horizon autonomous decision-making in realistic retail environments, revealing limitations in current LLMs for multi-factor decision-making. AIDABench provides a comprehensive benchmark for end-to-end AI data analytics tasks, highlighting challenges for current systems. For scientific code generation, petscagent-bench uses an agent-evaluating-agents paradigm to assess correctness, performance, and library-specific conventions, revealing struggles with the latter. Agent Rosetta, an LLM agent paired with Rosetta software, demonstrates capability in protein design, including with non-canonical residues. For continuous learning in biomedical NLP, MedCL-Bench evaluates strategies across task families and orders, showing catastrophic forgetting with sequential fine-tuning and identifying distinct retention-compute frontiers for different methods.

The robustness and trustworthiness of AI systems are further investigated. Conformal factuality filtering for RAG-based LLMs is analyzed, revealing trade-offs between factuality and informativeness, and fragility under distribution shifts. VeriGrey employs a grey-box approach to validate LLM agents by mutating prompts and using tool invocation sequences to uncover security risks. DynaTrust defends multi-agent systems against sleeper agents using dynamic trust graphs that model trust as an evolving process. The effectiveness of negative constraints over positive preferences for AI alignment is theorized, suggesting a shift towards falsification logic for more stable boundaries. Research also explores LLM behavior in simulated economic and gambling scenarios, with findings suggesting implicit encoding of cognitive biases like Prospect Theory and persona-conditioned risk behavior.

Finally, several papers focus on improving reasoning and decision-making processes. InfoDensity rewards information-dense reasoning traces to reduce verbosity and computational cost. Contrastive Reasoning Alignment (CRAFT) uses reinforcement learning from hidden representations to improve robustness against jailbreak attacks. Adaptive Theory of Mind (A-ToM) agents align their reasoning depth with partners to improve coordination. For complex logical reasoning, Draft-and-Prune improves auto-formalization by drafting multiple plans and pruning contradictory formalizations. ExpressMind, a multimodal pretrained LLM for expressway operations, integrates traffic data, emergency reasoning chains, and video events, outperforming baselines in event detection and incident response. The research also touches upon foundational aspects like the theoretical equivalence of Transformers to Bayesian Networks and the development of new benchmarks for specific domains like remote sensing route planning (NeSy-Route) and surgical intelligence (SurgΣ).