Frontier LLMs Advance Strategy While FormalJudge Guarantees Safety

Recent advancements in AI are pushing the boundaries of strategic behavior, reasoning, and evaluation across diverse domains. Frontier Large Language Models (LLMs) demonstrate deeper strategic behavior than humans in iterated rock-paper-scissors, as revealed by AlphaEvolve's analysis. In reinforcement learning, Neuro-symbolic Action Masking (NSAM) improves sample efficiency and reduces constraint violations by learning symbolic models of states to rule out infeasible actions. OmniSapiens-7B 2.0, a foundation model for social behavior processing, achieves state-of-the-art performance across behavioral tasks using Heterogeneity-Aware Relative Policy Optimization (HARPO), which balances learning across heterogeneous tasks. For agentic oversight, FormalJudge offers a neuro-symbolic framework that uses LLMs to compile natural language requirements into formal specifications, providing mathematical guarantees and outperforming LLM-as-a-Judge baselines by 16.6% in behavioral safety and deception detection.

Evaluating and enhancing LLM reasoning capabilities is a key focus. Reinforcement Learning with CoT Supervision via Self-Evolving Rubrics (RLCER) autonomously rewards chain-of-thought reasoning without human annotation, outperforming outcome-centric methods. LiveMedBench, a continuously updated, contamination-free medical benchmark, reveals pervasive data contamination risks and highlights contextual application as a bottleneck, with 84% of models degrading on post-cutoff cases. Found-RL enhances reinforcement learning for autonomous driving by efficiently integrating foundation models, using an asynchronous batch inference framework and supervision mechanisms like Value-Margin Regularization (VMR) and Advantage-Weighted Action Guidance (AWAG) to achieve near-VLM performance at high FPS. For bargaining scenarios, AgoraBench and human-aligned metrics improve LLM negotiation performance, yielding deeper strategic behavior and stronger opponent awareness.

LLMs are also being explored for complex generation and planning tasks. Flow of SpanS (FOSS) uses Generative Flow Networks to construct a dynamic span vocabulary, improving text generation quality and knowledge-intensive task performance. LLMs show potential in generating Qualitative Numerical Planning (QNP) abstractions for generalized planning, especially when guided by automated debugging. However, LLMs struggle with cultural adaptation in recipe generation, failing to produce culturally representative adaptations and misunderstanding notions of creativity and tradition, as shown in studies using the GlobalFusion dataset. In higher education, stakeholders show interest in GenAI for programming support but express concerns over response quality and academic integrity, necessitating responsible integration frameworks.

Evaluating agent capabilities in specialized environments is crucial. CLI-Gym generates a large collection of environment-intensive tasks for command-line interface agents, leading to significant improvements in models like LiberCoder. GameDevBench, the first benchmark for agentic game development, reveals a substantial reasoning-action gap and challenges in multimodal understanding, with simple image and video feedback mechanisms improving agent performance. In contrast to formal reasoning, LLMs do not consistently outperform non-reasoning models on Theory of Mind tasks, often relying on shortcuts like option matching and showing accuracy drops with longer responses, indicating a need for unique capabilities beyond current reasoning methods. Generative recommendation models are improved by V-STAR, which uses Value-Guided Efficient Decoding and Sibling-GRPO to address probability-reward mismatches and enhance exploration and candidate-set diversity.