Miner and AT2PO enhance LLM efficiency while Agent Mallard improves safety

Recent advancements in AI are pushing the boundaries of reasoning, collaboration, and safety across various domains. In the realm of large language models (LLMs), new frameworks are emerging to enhance their reasoning capabilities and efficiency. Miner and AT2PO, for instance, focus on data-efficient reinforcement learning and agentic turn-based policy optimization, respectively, showing significant gains in benchmarks. To combat the inefficiency of training on homogeneous prompts, Miner repurposes policy uncertainty as a self-supervised reward signal, achieving up to 4.58 absolute gains in Pass@1. AT2PO introduces a turn-level tree structure for strategic exploration and credit assignment, improving state-of-the-art baselines by up to 1.84 percentage points. FusionRoute and GlimpRouter explore token-level multi-LLM collaboration and efficient reasoning by glimpsing token thoughts, respectively, demonstrating improved performance and reduced latency. SCALER provides a synthetic, scalable, adaptive learning environment for reasoning, sustaining effective learning signals through adaptive environment design and outperforming dataset-based RL baselines.

Beyond core reasoning, AI is being applied to complex real-world problems with a focus on safety and reliability. Agent Mallard integrates a stochastic digital twin into its conflict-resolution loop for tactical air traffic control, combining model-based safety assessment with interpretable decision logic. In cybersecurity, defenses against indirect prompt injection are being developed, with one method achieving competitive utility while maintaining the lowest attack success rate to date by precisely parsing tool results and filtering malicious code. For autonomous systems and robotics, GUITester is a multi-agent framework designed for exploratory GUI testing, decoupling navigation from verification to autonomously discover defects. In scientific discovery, SciIF benchmarks scientific instruction following, emphasizing auditability and adherence to scientific validity constraints, while Sci-Reasoning provides a dataset to understand AI innovation patterns, identifying dominant thinking strategies like Gap-Driven Reframing and Cross-Domain Synthesis.

AI's role in specialized domains is also expanding. In manufacturing, a CTPN-MBRL approach optimizes flexible manufacturing systems by integrating AGVs and tool sharing, outperforming traditional methods on larger instances and reducing computation time tenfold. For aeronautics, Hybrid MKNF is evaluated for its expressivity and efficiency in capturing complex domain knowledge, with proposed heuristics for integration. In materials science, a neuro-symbolic AI approach is proposed, using structured, queryable knowledge graphs derived from reviews, with LLMs serving as complementary interfaces. For LLM evaluation itself, DVD is introduced as a robust method for detecting variant contamination, outperforming existing baselines. Furthermore, research is exploring the potential for LLMs to influence beliefs, with findings indicating they can be as effective at promoting conspiracy beliefs as debunking them, though corrective conversations and prompting for accuracy can mitigate this risk. The development of computational compliance for AI regulation is also highlighted as a critical new research domain, requiring algorithms that automatically steer AI systems towards compliance.

The efficiency and scalability of AI systems are key research themes. OI-MAS framework uses confidence-aware routing across multi-scale LLMs to improve accuracy by up to 12.88% while reducing cost by up to 79.78%. DR-LoRA dynamically adjusts LoRA ranks for Mixture-of-Experts adaptation, achieving superior task performance with more efficient parameter utilization. For multimodal retrieval, CIEA extracts and aligns complementary information between text and images, achieving significant improvements over existing models. In agent development, AgentDevel reframes self-evolving LLM agents as release engineering, emphasizing non-regression and auditable artifacts. Research into learning latent action world models from in-the-wild videos expands the scope of existing works, capturing richer actions despite video diversity. For LLM evaluation, Evaluative Fingerprints reveal stable, systematic differences in LLM evaluator behavior, highlighting that judges are consistent with themselves but not each other, functioning as distinct 'evaluative dispositions'.