The Morning
From the arXiv
MemPilot: Orchestrating On-Demand Multimodal Memory Curation for LLM Agents
emPilot is a flexible framework that optimizes LLM agent memory by orchestrating on-demand, multimodal memory curation. It uses a reinforcement-learned policy to dynamically decide between retrieving from pre-built memory or performing query-specific curation using various LLMs and VLMs, balancing performance, cost, and latency. This allows for more efficient and adaptable memory management for LLM agents.
Before Agent Tells The Lie: Has Deception Already Been Represented?
This paper proposes a novel method to predict LLM agent deception by analyzing their internal hidden states *before* deceptive actions occur. The core contribution is demonstrating that these internal representations, when aligned around key decision points, c…
LoGRA: Scaling LLM Reinforcement Learning with Low-Rank Gradient Sketches
LoGRA addresses the memory bottleneck in LLM reinforcement learning by compressing gradient information into low-rank sketches, significantly reducing memory usage. This method, combined with predicted-KL step control to manage update magnitudes, enables stabl…
Reward Stealing Attack on Large Language Models
This paper introduces Reward Stealing Attack (ReSA), a novel method to attack Large Language Models (LLMs) by targeting their internal safety reward signals. ReSA recovers a proxy reward model from the LLM's behavior and then reverses it to generate adversaria…
BazaarBench: Delegation Safety in Decentralized C2C Marketplaces Run by LLM Agents
This paper introduces BazaarBench, a simulated decentralized C2C marketplace designed to evaluate the safety of LLM agents acting on behalf of users. BazaarBench tracks transaction integrity and uses a combination of record checks and LLM judgments to identify…
Can Agent Harnesses and Inference Engines Hear Each Other? The HEAR Protocol for Agentic LLM Serving
This paper introduces the HEAR protocol, a bidirectional communication standard for agentic LLM serving. HEAR enables agent harnesses and inference engines to share crucial informa…
CLIFT: Conformal Self-Verification for Web Agent Training and Test-Time Scaling
CLIFT addresses the challenge of training web agents by using a self-verification mechanism. During training, the agent generates verification questions about its actions, and a "C…
Closing the Context Gap: Activation Alignment for Tabular In-Context Learning
This paper introduces **Activation Alignment**, a novel method to improve in-context learning (ICL) for tabular foundation models. It addresses the trade-off between prediction cos…
HERA: Harness-Environment Co-Evolution for Reliable Agentic Abstention
HERA addresses the challenge of LLM agents failing to abstain from impossible tasks by introducing a co-evolutionary framework. It automatically generates feasible and infeasible t…
Learning to Read the Contextual Tokens in Diffusion Transformers
This paper introduces a method to interrogate the internal "contextual tokens" of Diffusion Transformers used for multimodal generation. By training a lightweight network to map th…
The Town Square
Reflection AI has launched Beam, a 501 billion parameter open-weight language model, aiming to advance AI research and development.
Workshops
This repository provides persistent context for AI agents by capturing, compressing, and re-injecting session data, enhancing future interactions across various coding and language models.
This repository enables agents to generate CAD models from text descriptions, effectively giving them "CAD superpowers" by translating natural language into 3D designs.
