GAAMA Advances Agent Memory While TianJi Drives Scientific Discovery

Recent advancements in AI are pushing the boundaries of agent capabilities and reasoning across diverse domains. For long-term agent memory, GAAMA introduces a graph-augmented associative memory system that outperforms RAG baselines on the LoCoMo-10 benchmark by maintaining conversational coherence and personalized behavior. In scientific discovery, TianJi acts as an autonomous AI meteorologist, using a multi-agent architecture to autonomously conduct literature research, generate hypotheses, and drive numerical models to verify physical mechanisms, compressing research cycles to hours. AutoMS, a neuro-symbolic framework, employs LLM-driven evolutionary search for cross-physics inverse microstructure design, achieving an 83.8% success rate on 17 tasks. For AI safety and alignment, the CounterMoral benchmark assesses LLM editing techniques for moral judgments, while the EU AI Act's dual transparency mandate faces structural compliance gaps in current generative AI systems, as highlighted by studies on fact-checking and synthetic data generation.

Evaluating AI systems remains a critical challenge, with new benchmarks emerging to address specific limitations. ScholScan focuses on scan-oriented academic paper reasoning, revealing systematic deficiencies in current MLLMs for full-document understanding. MiroEval provides a holistic evaluation framework for deep research systems, assessing synthesis quality, factuality verification, and process-centric audits, finding that multimodal tasks pose significantly greater challenges. MonitorBench offers a benchmark for chain-of-thought monitorability, showing that closed-source LLMs generally have lower monitorability and can intentionally reduce it under stress. PeopleSearchBench evaluates AI-powered people search platforms using criteria-grounded verification, with Lessie achieving 100% task completion. FormalProofBench tests LLMs on graduate-level, formally verified mathematical proofs, where the best model achieved 33.5% accuracy.

The development of more capable and reliable AI agents is being driven by innovations in learning, reasoning, and architecture. DSevolve evolves portfolios of dispatching rules for dynamic manufacturing scheduling, outperforming state-of-the-art methods. The novelty bottleneck framework explains human effort scaling in AI-assisted work, identifying irreducible serial components. For combinatorial optimization, AlignOPT aligns LLMs with graph neural solvers to learn generalizable heuristics, achieving state-of-the-art results. Neuro-symbolic approaches are also enhancing predictive process monitoring by injecting domain knowledge as differentiable logical constraints, improving accuracy and compliance, particularly in regulated scenarios. Furthermore, research into uncertainty quantification is advancing with distance-based approaches for credal sets and collaborative entropy (CoE) for multi-LLM systems, aiming to capture semantic disagreement.

Research into LLM reasoning and behavior is uncovering fundamental properties and limitations. The price of meaning is interference and forgetting in semantic memory systems, with no architecture fully avoiding it. CoT2-Meta, a metacognitive reasoning framework, improves test-time reasoning performance across various benchmarks by explicitly controlling computation. For AI tutoring, SLOW provides a deliberate reasoning workspace for cognitive adaptation, enhancing personalization and emotional sensitivity. In the realm of AI development itself, daVinci-LLM advances the science of pretraining with an open paradigm and systematic exploration, while Meta-Harness optimizes LLM harnesses through agentic search, improving performance and efficiency. Reward hacking is identified as a structural equilibrium under finite evaluation, with implications for AI safety and the transition to agentic systems.