LLM Agents Advance Reasoning While Hyperagents Enable Self-Improvement

Recent advancements in AI are pushing the boundaries of autonomous systems and reasoning capabilities across various domains. For instance, new methods are enabling agents to generate vector sketches part-by-part, offering interpretable and locally editable text-to-vector sketch generation through multi-modal language models and reinforcement learning on a novel dataset (ControlSketch-Part). In the realm of complex decision-making, multi-objective reinforcement learning (MORL) is being enhanced by methods like PA2D-MORL, which improves Pareto policy set approximations for robot control tasks. For LLM agents tackling long-horizon tasks, frameworks like MiRA are improving success rates on benchmarks like WebArena-Lite by up to 43.0%, surpassing proprietary models, through subgoal decomposition and milestone-based rewards. Furthermore, LLM agents are being augmented for personalized mobile power management with PowerLens, achieving 38.8% energy savings over stock Android by adapting to user preferences and ensuring safety through a constraint framework.

The verification and generation of formal proofs are also seeing significant AI integration. Stepwise, a neuro-symbolic proof generation framework, automates proof search for systems-level verification, proving up to 77.6% of theorems on the seL4 benchmark. Complementing this, Learning to Disprove focuses on LLMs for generating formal counterexamples, enhancing theorem proving capabilities with a symbolic mutation strategy and multi-reward expert iteration. FormalEvolve offers a neuro-symbolic evolutionary framework for autoformalization, generating diverse and prover-effective mathematical statements. In planning, a partially grounded encoding of planning into SAT is proposed, scaling linearly with plan length and outperforming state-of-the-art on hard-to-ground domains, while transformers are being analyzed for their ability to verify plans, with theoretical guarantees for certain classical planning domains.

Beyond specific task execution, research is exploring more general AI capabilities and architectures. HyEvo, a framework for automated workflow generation, integrates LLMs with deterministic code nodes for efficient reasoning, reducing inference costs and latency by up to 19x. Hyperagents introduce a self-referential architecture where meta-agents modify themselves and task agents, enabling metacognitive self-improvement and potentially open-ended progress across diverse computable tasks. The concept of 'embodied science' is proposed, reframing scientific discovery as a closed loop coupling agentic reasoning with physical execution via a Perception-Language-Action-Discovery (PLAD) framework. For LLM agents, utility-guided orchestration policies are being developed to balance answer quality with execution cost, managing tool use more efficiently. Additionally, DIAL-KG offers a framework for incremental knowledge graph construction, dynamically inducing schemas and evolving knowledge to handle dynamic data. Theory of Mind (ToM) reasoning in LLMs is being advanced with dynamic belief graphs that jointly infer and link evolving beliefs to actions under uncertainty, improving action prediction in high-stakes settings. Finally, research into LLM evaluation highlights pitfalls in assessing interpretability agents, suggesting intrinsic evaluation methods beyond replication-based comparisons, and new RL frameworks like HeRL are improving exploration by leveraging hindsight experience for more effective learning.