The Devil in the Details: Emergent Misalignment, Format and Coherence in Open-Weights LLMs

The introduction of VideoChat-M1 represents a significant advancement in video understanding through a novel multi-agent system that employs Collaborative Policy Planning (CPP). This system allows multiple agents to generate, execute, and communicate unique tool invocation policies tailored to user queries, enhancing the exploration of complex video content.

Large multi-modal models (LMMs) have shown improved performance in visual tasks, particularly in analyzing video sequences. A recent study evaluated their ability to detect core interaction events, such as when hands contact or release objects, using a new dataset with over 20,000 annotated interactions from the Something-Something-V2 dataset.

A new study introduces V-Attack, a method designed to enhance controllability in adversarial attacks on Large Vision-Language Models (LVLMs) by targeting disentangled value features. This approach addresses the limitations of existing methods that struggle with precise semantic manipulation due to the entanglement of semantic information in patch-token representations.

CNS-Obsidian, a neurosurgical vision-language model, has been developed using 23,984 peer-reviewed articles from Neurosurgery Publications, resulting in 263,064 training samples. This model aims to enhance decision-making in neurosurgery by providing a specialized alternative to general-purpose models like GPT-4o.

ConfTuner is a newly introduced fine-tuning method aimed at enhancing the verbalized confidence of Large Language Models (LLMs), addressing the issue of overconfidence in high-stakes domains like healthcare and law. This method does not require ground-truth confidence scores, making it a more efficient approach compared to existing techniques that rely on prompt engineering or heuristic estimates.

A new foundational language model, EHR-R1, has been developed to enhance the analysis of Electronic Health Records (EHRs), addressing limitations in existing large language models (LLMs) regarding EHR-oriented reasoning capabilities. This model is built on a comprehensive dataset called EHR-Ins, which includes 300,000 reasoning cases across 42 distinct EHR tasks, enabling better clinical decision-making.

The introduction of the Structured Cognitive Loop (SCL) addresses critical architectural challenges faced by large language model (LLM) agents, such as entangled reasoning and memory volatility. SCL modularizes cognition into five distinct phases: Retrieval, Cognition, Control, Action, and Memory, enhancing the explainability and controllability of LLMs through Soft Symbolic Control.

A recent study reveals that post-trained language models (PoLMs) often exhibit over-confidence, which can lead to unreliable outputs in critical applications. To combat this, researchers introduced Disagreement-Aware Confidence Alignment (DACA), an unsupervised method aimed at optimizing confidence calibration in PoLMs by addressing the prediction disagreement with pre-trained language models (PLMs).