This research paper explores the concept of ensemble dispersion in language models, specifically focusing on the diversity of perspectives generated by multiple models. The study presents a method to measure diversity and examines how model identity contributes to the diversity of ensemble outputs. The findings suggest that model identity is a significant factor in shaping the diversity of ensemble outputs, but the relationship is complex and influenced by various factors. This research has implications for the development and deployment of AI agents that rely on ensemble models.

Researchers identified two fundamental gaps in Large Language Models (LLMs) decision-making: an observation-belief gap and a belief-action gap. These gaps lead to systematic vulnerabilities that can be exploited in strategic domains without robust guardrails.

Researchers investigated how LLMs compute verbal confidence, finding that confidence is cached and retrieved from answer tokens, rather than being generated just-in-time. This suggests that verbal confidence reflects a richer answer-quality evaluation rather than a simple fluency readout, with implications for understanding metacognition in LLMs and improving calibration.

Researchers introduce BEAT-Net, a supervised biomimetic framework for ECG diagnosis using deep learning. It integrates QRS-centered biological tokenization with a hierarchical architecture, achieving comparable diagnostic accuracy to CNN baselines while reducing parameters by 95%. This framework demonstrates an efficient and interpretable alternative to massive pre-training for clinical deployment.

Researchers analyzed GPT-4.1's knowledge base and found significant inaccuracies, inconsistencies, and hallucinations, highlighting the need for better methods to extract, consolidate, and verify factual knowledge from large language models.

This article evaluates the use of Large Language Models (LLMs) for generating Entity-Relationship (ER) diagrams from natural language requirements. The study found that LLMs perform well in simple scenarios but struggle with complex ones, leading to inconsistencies and failures in representing constraints. This indicates that LLMs are not yet mature for reliable use in complex conceptual database modeling.

BoostAPR is a three-stage framework for automated program repair that uses reinforcement learning with dual reward models to improve edit-level feedback. It achieves competitive results on various benchmarks, particularly in cross-language generalization.

Researchers introduced Bayesian Chronicle Agents (BCA), a belief layer that separates what an agent believes from how it speaks. This allows for more control over opinion dynamics in LLM agents, enabling three canonical regimes of opinion dynamics on demand. The BCA layer also makes simulation auditable, surfacing systematic stance biases.

A new benchmarking framework is proposed for evaluating the explanatory quality of open-weight vision-language models in face recognition. The framework focuses on explanation quality, including relevance and faithfulness, and provides a structured explanation format for automated querying and auditing. The authors benchmark several families of open-weight VLMs and highlight the need for explanation quality metrics in face recognition systems.

This paper introduces a novel transfer-learning mechanism for Neural Cellular Automata (NCAs) that injects a pretrained teacher's hidden states into a student model to guide early optimization. This results in superior generalization with a minimal parameter budget. The hidden channels of NCAs decouple feature extraction from uniform classification consensus, allowing for robust, decentralized computational substrates for parameter-efficient transfer learning.

A new benchmark, CIBuzzBench, is introduced for evaluating the ability of large language models (LLMs) to understand Chinese internet buzzwords across languages. The benchmark includes 3,001 annotated buzzwords with English explanations and evaluates LLMs' performance in cross-lingual understanding, meaning explanation, and harmfulness detection. The results show that LLMs struggle with fine-grained non-literal interpretation and equivalent matching, highlighting challenges for multilingual LLMs and safety-oriented evaluation.

Researchers introduced SynthDemo-RL, a framework for fine-tuning Vision-Language-Action models using synthetic demonstrations generated by an automated teacher. This approach improves the performance of reinforcement learning with sparse rewards and enables successful task execution without human demonstrations. The framework demonstrates significant improvements on various benchmarks and validates its effectiveness on a physical robot.

OneBid is a unified auto-bidding foundation model for diverse oCPX advertising scenarios, built on top of the Decision Transformer (DT) model. It extends DT with two atomic signals and a sequence-level Mixture-of-Experts architecture to learn from heterogeneous logs and adapt to scenario-specific deployments. OneBid is validated through online A/B tests and delivers improved performance in real-world scenarios.

VidOmni-Bench is a new benchmark for evaluating fine-grained video understanding in Video Large Language Models (Video-LLMs). It assesses whether models can accurately verify events in video captions, revealing weaknesses in current Video-LLMs that generate hallucinated descriptions and struggle with detecting incorrect event descriptions.

The OmniVChat paper introduces OmniVChat-Studio, a multi-agent data engine for synthesizing audio-visual dialogues, and OmniVChat-Bench, an evaluation benchmark for omni models. The paper also presents OmniVChat-RL, a reinforcement learning reward design for improving omni model performance.

A study analyzed human-AI and human-human dialogues to investigate the function of morality, politeness, and alignment in cooperative communication. The results suggest that AI systems may simulate cooperative communication without the underlying social architecture, and that mechanisms sustaining human accommodation can run in reverse with AI. This has implications for the design and evaluation of conversational AI systems.

Researchers propose a multimodal reasoning framework for cross-domain PCBA visual question answering, addressing domain shift and heterogeneous output spaces.

A new evaluation paradigm, the Multilayered Evaluation Framework, is introduced to assess the correctness of Text-to-SQL systems with AI operators. Traditional metrics fail to accurately evaluate these systems, leading to a 25% detection rate for correct translations. The proposed framework decouples deterministic database logic from flexible AI semantics, achieving state-of-the-art accuracy across industry and academic systems.

Researchers propose a method to decouple learning correlations in LLMs from abstract values and introduce a dataset to probe cross-lingual behavior in LLMs. The study shows that fine-tuning and Direct Preference Optimization can remove bias in LLMs, increasing accuracy to over 98%. The method involves task vector transfer and orthogonalization to isolate specific value preferences.

Researchers evaluated LLM-generated GPU kernels and found they can outperform PyTorch in certain tasks, with some models achieving correct kernels for 91.1% of problems and verified speedups on 22 out of 56. However, the kernels only govern a small fraction of a real model's wall clock, ranging from 8.9% to 58.2%. The study also introduced DLRM-Bench, a benchmark for recommender kernel problems, and proposed scale-invariant replacements for flawed kernels.