The Morning
From the arXiv
Belief-State Engine: Augmenting LLMs for Principled Planning Under Partial Observability
his paper addresses the failure of LLM agents in partially observable environments by introducing the Belief-State Engine (BSE). The BSE is an external module that maintains a Bayesian posterior over the environment's hidden states, feeding only this belief to the LLM. This architectural change allows LLMs to perform principled planning by explicitly reasoning about uncertainty, rather than relying solely on raw observation history.
DiSCo: A Distribution-First Steering and Cultural Prior Evaluation Framework for Measuring Cultural Preference Bias in LLMs
DiSCo is a novel framework for evaluating LLMs' cultural bias by focusing on their default preferences rather than single correct answers. It uses a forced-choice, distribution-first approach with a context gradient to isolate and measure inherent cultural lea…
Show-Harness: Just a VLM Agent Can Play Robots
Show-Harness provides a semantic interface that allows Vision-Language Models (VLMs) to control robots by translating high-level intents into discrete actions. This method enables zero-shot control of powerful VLMs and efficient adaptation of smaller models, d…
Active Adaptation, Not Static Defense: Temporal Dynamics of Preventative Steering in Adversarial Fine-Tuning
This paper investigates the temporal dynamics of Preventative Steering, a defense against adversarial fine-tuning in LLMs. It reveals that the defense relies on an **active adaptation process** during training, rather than a static injection of corrective sign…
Beyond One-Size-Fits-All: Sample-Adaptive Strategy Routing for Vision Token Pruning in MLLMs
This paper addresses the inefficiency of processing numerous visual tokens in Multimodal Large Language Models (MLLMs) by proposing VIP-Router. Instead of a single fixed pruning strategy, VIP-Router adaptively selects the best pruning method for each input bas…
Beyond Surface Imitation: Contrastive Modeling for Reasoning Path Alignment in Multimodal In-Context Learning
This paper proposes a novel multimodal in-context learning (ICL) framework that moves beyond superficial imitation. It introduces contrastive demonstration modeling, where demonstr…
Can Foundation Models Moderate Online Content? Evaluating Instruction- vs. Example-Driven Policy Operationalization
This paper investigates whether foundation models can effectively moderate online content by comparing instruction-driven and example-driven policy operationalization. Using the ne…
ConvMem: Convolutional Memory for Long-Context Reasoning
ConvMem addresses LLM limitations with long contexts by reformulating reasoning as a hierarchical convolution, treating the LLM as a kernel. This training-free, parallelizable meth…
Direct Diversity Optimization for Diverse Successful Trajectories in Preference Post-Training
This paper addresses the challenge of training LLM agents to discover multiple successful strategies from a single decision point. Their proposed method, Direct Diversity Optimizat…
Forgetting Only What Matters: Layer-Selective Unlearning toward Robust LLMs
This paper proposes FOM-UL, a layer-selective unlearning method for LLMs. It identifies and updates only the most influential transformer layers for forgetting specific data, minim…
The Town Square
Apple's AirPods 5 launch boasts industry-leading open-ear active noise cancellation, offering a new immersive audio experience.
Workshops
The `teamai-cli` repository provides a command-line interface to help teams integrate AI into their workflows, making their operations AI-native.
Superpowers is an agentic skills framework and software development methodology designed to enhance productivity and effectiveness through a structured approach to building and utilizing AI agents.
