Researchers introduce Traversal-as-Policy, a method that distills LLM agent execution logs into Gated Behavior Trees (GBTs) to create explicit, verifiable control policies. This dramatically improves success rates and safety on benchmarks like SWE-bench Verified and WebArena, while also reducing token usage.
Sources: via arXiv AI
Researchers introduce ReflexiCoder, a reinforcement learning framework that teaches LLMs to autonomously reflect on and correct generated code without external feedback. This approach sets a new state-of-the-art among open-source models in code generation benchmarks, outperforming or rivaling proprietary models like GPT-5.1, while also significantly improving inference-time.
Sources: via arXiv AI
Researchers have developed FlashPrefill, a framework that significantly accelerates the prefilling phase of Large Language Models by employing instantaneous pattern discovery and dynamic thresholding. This method achieves an unprecedented 27.78x speedup on 256K sequences, addressing a critical bottleneck in long-context modeling.
Sources: via arXiv AI
Researchers have developed Reasoning Reward Models (ReasRMs), exemplified by RM-R1, which integrate reasoning into reward modeling for aligning large language models. RM-R1 utilizes a chain-of-rubrics mechanism and has demonstrated superior performance on reward model benchmarks, surpassing even proprietary models like GPT-4o by up to 4.9%.
Sources: via arXiv AI
Researchers propose a novel post-training pipeline using knowledge graphs as implicit reward models, enabling a 14B model to significantly surpass larger models like GPT-5.2 and Gemini 3 Pro in complex multi-hop reasoning tasks. This is achieved by grounding the reasoning process in structured knowledge.
Sources: via arXiv AI
A new research paper showcases how Google's Gemini-based AI models, including Gemini Deep Think, are being used to accelerate scientific research by solving open problems, refuting conjectures, and generating new proofs in fields like theoretical computer science, economics, and physics. The study details effective human-AI collaboration techniques and advanced applications, positioning AI as a powerful research assistant.
Sources: via arXiv AI
Researchers, including those from Google DeepMind, have developed Aletheia, an AI math research agent powered by an advanced Gemini Deep Think model. Aletheia demonstrates the ability to autonomously generate, verify, and revise mathematical proofs, achieving milestones like producing a research paper without human intervention and solving open problems.
Sources: via arXiv AI
Researchers have introduced a purely photonic deep neuromorphic network architecture that enables online, unsupervised learning using an all-optical local feedback mechanism and non-volatile phase-change material synapses. Demonstrated with 100% accuracy on a letter recognition task, this promises significant advances in energy-efficient, high-throughput AI processing.
Sources: via arXiv AI