CORE Advances Reasoning While Helios Enhances Energy Knowledge

Researchers are developing advanced frameworks to enhance the capabilities and reliability of AI agents across various domains. For scientific discovery, SGI-Bench aims to operationalize Scientific General Intelligence (SGI) through scientist-aligned tasks, revealing LLM limitations in deep research and experimental reasoning, while TTRL optimizes hypothesis novelty. In agentic workflows, PAACE offers a Plan-Aware Automated Context Engineering framework that improves agent correctness and reduces context load on benchmarks like AppWorld and OfficeBench. For complex reasoning, CORE trains LLMs with Concept-Oriented Reinforcement to bridge the definition-application gap in mathematical reasoning, and CORE-R1 uses RL for self-improving agents with skill libraries, showing gains in accuracy and efficiency on AppWorld. LLMs are also being adapted for specific domains: Vox Deorum integrates LLMs with other AI for 4X game strategy, achieving competitive gameplay; Helios is a foundational LLM for smart energy knowledge reasoning; and an agentic framework automates first-principles materials computations, improving accuracy and robustness.

Advancements in AI reasoning and decision-making are being explored through various lenses. UniRel-R1 integrates subgraph selection and LLM fine-tuning for relation-centric Knowledge Graph Question Answering, producing compact and informative subgraphs. For reasoning under uncertainty, a Solomonoff-inspired method weights LLM-generated hypotheses by simplicity and predictive fit, offering uncertainty-aware outputs. In sequential decision-making, the Rashomon effect is translated, showing ensembles from Rashomon sets exhibit greater robustness. For embodied agents, ESearch-R1 unifies dialogue, memory retrieval, and navigation into a cost-aware framework, reducing operational costs. ChronoDreamer, an action-conditioned world model, acts as an online simulator for robotic planning, predicting future frames and using an LLM judge to reject unsafe actions. Furthermore, LLMs are being evaluated for strategic play in games like Pokémon, demonstrating competence without domain-specific training.

The reliability, safety, and interpretability of AI systems are critical areas of research. Security risks in Agentic Vehicles (AgVs) are analyzed through a role-based architecture, identifying vulnerabilities in agentic and cross-layer interactions. For AI interpretability, a pragmatic statistical-causal reframing is proposed to address "dead salmon" artifacts, advocating for treating explanations as statistical estimators. Monitorability of AI decision-making is evaluated using intervention, process, and outcome-property archetypes, finding that longer CoTs are generally more monitorable. SafeMed-R1 uses adversarial reinforcement learning and randomized smoothing for robust medical reasoning in VLMs, significantly improving accuracy under attacks. The PENDULUM benchmark assesses sycophancy in multimodal LLMs, revealing susceptibility and the need for resilience. Recontextualization is proposed to mitigate specification gaming by training models to resist misbehavior even when instructions permit it.

AI's ability to learn and adapt is being pushed through new frameworks and benchmarks. UmniBench provides an omni-dimensional benchmark for unified multimodal understanding and generation models. MSC-180, a benchmark based on the Mathematical Subject Classification, evaluates LLM-based theorem provers, revealing domain bias and weak generalization. For cognitive modeling, NL2CA auto-formalizes decision-making from natural language into executable rules using an unsupervised critic. The External Hippocampus framework uses topological cognitive maps to guide LLM reasoning, addressing cognitive deadlocks in smaller models. IntelliCode, a multi-agent LLM tutoring system, uses a centralized learner model for principled pedagogical support. KeenKT addresses ambiguity in Knowledge Tracing by representing student mastery states with NIG distributions, outperforming SOTA models. ASTIF integrates semantic and temporal data for cryptocurrency price forecasting, outperforming baselines through adaptive meta-learning.