EviBound Enhances AI Safety While GAIA Governs LLM Negotiations

Researchers are developing advanced frameworks to enhance the reliability and safety of AI systems. EviBound addresses false claims in autonomous research by enforcing evidence-bound execution with dual governance gates, achieving 0% hallucination on benchmark tasks. For AI safety, MONICA monitors and mitigates sycophancy in reasoning steps, while MENTOR uncovers and mitigates domain-specific implicit risks through metacognitive self-assessment and self-evolution. GAIA provides a governance-first framework for LLM-human B2B negotiation, ensuring bounded authorization and information-gated progression. To combat misinformation, ED2D uses evidence-based multi-agent debate for intervention and persuasion, demonstrating effects comparable to human experts. Furthermore, robust accident anticipation for autonomous vehicles is addressed by ROAR, which combines DWT, an object-aware module, and dynamic focal loss to handle noisy data and imbalanced distributions.

The energy consumption of AI, particularly LLM inference, is a growing concern. A study of over 32,500 measurements across 21 GPU configurations and 155 model architectures quantifies energy usage at the prompt level, developing a predictive model for inference energy consumption. In parallel, Agentic AI sustainability assessment for supply chain document insights shows significant reductions in energy, carbon, and water usage with AI-assisted and agentic AI workflows compared to manual processes. Green AI research defines unified operational definitions and lifecycle models to address multi-dimensional burdens across the AI lifecycle.

New benchmarks and evaluation methods are crucial for advancing AI capabilities. DigiData introduces a large-scale dataset and benchmark for mobile control agents, proposing dynamic evaluation protocols and AI-powered evaluations beyond step-accuracy. FractalBench diagnoses visual-mathematical reasoning through recursive program synthesis, revealing significant gaps in AI's mathematical abstraction abilities. LPFQA offers a long-tail professional forum-based benchmark for LLM evaluation, targeting knowledge depth, reasoning, and terminology comprehension across diverse fields. For multimodal reasoning, MathSE uses self-evolving iterative reflection and reward-guided fine-tuning to improve mathematical problem-solving in MLLMs, outperforming existing models. The STATION environment supports AI-driven discovery through an open-world scientific ecosystem where agents engage in long research journeys.

Research is also focusing on improving LLM reasoning and interpretability. SofT-GRPO enhances LLM reinforcement learning with a soft-thinking paradigm, outperforming discrete-token GRPO on Pass@32. CoT-X offers an adaptive framework for cross-model Chain-of-Thought transfer, achieving higher accuracy than truncation under tight token budgets. SMAGDi distills multi-agent debate dynamics into a compact student model, retaining high accuracy with significantly reduced computational cost. UHeads provide efficient verification of LLM reasoning steps using uncertainty quantification, matching or surpassing larger models. DiagnoLLM integrates Bayesian methods and LLMs for interpretable disease diagnosis, generating audience-specific reports. PRIME uses logic grid puzzles to evaluate implicit biases in LLM reasoning, showing models reason more accurately when solutions align with stereotypes. Anchors in the Machine investigates anchoring bias in LLMs, demonstrating robust effects and using Shapley values for attribution.