Researchers proposed HERMES, a holistic end-to-end multimodal driving framework that incorporates long-tail semantic knowledge into trajectory planning for autonomous driving models. This framework uses a foundation-model-assisted annotation pipeline to capture hazard-centric scene information and risk-aware planning guidance, improving overall planning performance and safety in mixed-traffic environments.

A new representation of procedural knowledge in algorithm design, called Generative Executable Algorithm Knowledge Graphs (GEAKG), allows for the transfer of knowledge across domains by storing validated operators, admissible compositions, and effective sequences in a graph. This enables the reuse of procedural knowledge without runtime language-model calls.

This study investigates the effect of personality-aware fine-tuning on LLMs for social simulation. Researchers fine-tuned two LLMs using a corpus of personality-labelled social media posts and dialogues, but found that fine-tuning did not improve role-playing abilities. The study suggests that the quality and domain alignment of training data are crucial for accurate personality role-playing.

Researchers identified a limitation in latent chain-of-thought reasoning in language models, known as the latent-to-language transition gap, where steering effects in latent space do not translate to language generation. This finding may impact the development of steering methods for AI agents.

EnterpriseVal is a use-case-level evaluation system for generative AI (GenAI) in the enterprise. It addresses the issue of measuring GenAI efficacy, reliability, and value by providing a formal specification, metric catalogue, grading protocol, and decision-making framework. The system aims to help organizations make informed decisions about GenAI deployment and scaling. A pilot across three workflows in a global bank showed promising results, including high citation precision and reduced analyst refinement effort.

Researchers propose an offline multimodal large language model for decision support in air operations, enabling analysts to access doctrinal knowledge without internet connectivity. The model supports text and image input from technical manuals, outperforming human analysts in a pilot study.

Researchers developed InterviewPlayground, a clinician-grounded evaluation platform for AI-assisted psychiatric intake systems. The platform uses a memory-augmented patient simulator to compare AI interviewing approaches and measure clinically relevant performance. A pilot study found that a GPT-based LLM intake interviewer recovered more clinically relevant items but made more incorrect inferences and safety concerns compared to human clinicians.

Researchers proposed Attention-Aware Routing (AAR), a new approach for routing in Mixture-of-Experts (MoE) language models. AAR uses temporal and spectral features from attention weights to improve model performance and reduce long diverging generation. The method exposes a retrieval-reasoning tension across depth in MoE models, making it a controlled probe for routing-relevant information.

A new approach to multi-teacher distillation, called D^3-MOPD, dynamically adjusts the domain mixture during training to optimize student performance. This can lead to faster convergence and better results.

Researchers propose a new approach to distilling large language models (LLMs) by leveraging a hybrid model architecture and a multi-stage distillation pipeline. They demonstrate that log-likelihood evaluation can be misleading when assessing generation quality, and present a optimized distillation recipe that retains high accuracy while reducing memory and inference time. This work is relevant to developers of AI agents and LLMs, as it provides insights into efficient and effective model distillation techniques.

Researchers propose Test-Time Self-Evolving via Reflection (TTSR), a self-evolving framework for adapting large language models during inference using unlabeled test inputs. TTSR improves upon existing methods by addressing bottlenecks in learnable samples and efficient exploration. The framework alternates between a Student and Teacher role, with the Teacher analyzing failed trajectories and synthesizing targeted questions. Experiments show consistent test-time improvements and strong cross-backbone generalization.

Researchers propose a method for multi-resolution attribution in adaptive hierarchical systems, such as LLMs, Census, and telecom networks, by leveraging the learned routing state. This approach allows for consistent explanations at various levels of the system without requiring a separate attribution model.

Researchers proposed a multimodal visual retrieval-augmented generation (RAG) system for ophthalmology clinical question answering and decision support. The system, Oph-Guid-RAG, integrates query decomposition, rewriting, retrieval, reranking, and multimodal reasoning to provide traceable outputs with guideline page references. Evaluations on HealthBench show improved performance compared to GPT-5.2 and GPT-5.4, indicating the effectiveness of combining vision-based retrieval with controllable reasoning in clinical AI applications.

A new framework for selecting pre-trained vision-language models for specific downstream tasks is proposed, using layer-wise conductance and directional conductance divergence to improve performance and outperform state-of-the-art baselines. This development is relevant to people building and operating AI agents as it addresses the challenge of selecting the optimal model for a given task, which is crucial for efficient and effective AI deployment.

A new benchmark, oMeBench, is introduced for evaluating large language models (LLMs) in organic mechanism elucidation and reasoning. The benchmark consists of 10,000 annotated mechanistic steps with reaction type labels, intermediate structures, and difficulty ratings. Evaluation of state-of-the-art LLMs shows that while they have promising chemical intuition, they struggle to produce consistent reasoning across multi-step mechanisms. Combining prompting strategies with fine-tuning enables smaller-scale models to achieve performance comparable to larger models.

Researchers evaluated the effectiveness of alignment midtraining (AMT) for steering model motivation and learning rules for AI systems. Their findings suggest that AMT may not be sufficient to address the core difficulties in aligning powerful AI systems, and that demonstrations must be present in midtraining or post-training datasets for rules to be robustly learned.

Researchers proposed an automated pipeline to transform safe driving scenes into safety-critical scenarios using computer vision, LLMs, and AR, achieving 97.52% safety classification accuracy on the nuScenes dataset.

Researchers proposed two live streaming speech synthesis evaluation methods, Live-ProsodyJudge (LPJ) and Decoupled-Live-ProsodyJudge (D-LPJ), to assess fine-grained prosody such as emotion, intonation, and energy. LPJ and D-LPJ outperform a single proprietary Large Language Model (LLM) call in evaluation accuracy and demonstrate efficacy in fine-grained TTS preference optimization.

Researchers propose MaskHarness-WAM, a system for long-horizon robot manipulation that connects high-level task planning with low-level manipulation policies through target masks, leveraging visual feedback for subtask scheduling and execution.

Researchers propose Abstract Token Curriculum (ATC), a novel curriculum learning framework for training large language models (LLMs) to develop internal abstract thoughts without direct supervision or manual scratchpad design. ATC increases problem complexity through a sequence of distributions, enabling LLMs to learn continuous intermediate representations effectively.