The Morning
From the arXiv
OmegaUse-OfficeVal: Benchmarking LLM Agents on Long-Horizon Office-Suite Tasks with Economic Grounding
his paper introduces OmegaUse-OfficeVal, a benchmark designed to evaluate LLM agents on complex, multi-step office tasks. Its core contribution is the inclusion of economic grounding, pairing each task with human labor time and a price proxy. This allows for cost-effective evaluation of LLM agents against human performance and provides a novel way to assess their economic value.
On-Policy Distillation for LLM Safety: A Routing Approach to Template-Robust Realignment
This paper introduces Routing-based On-Policy Distillation (ROPD), a novel method to improve LLM safety by distilling knowledge from an aligned model to a potentially compromised one. Instead of relying on specific prompt templates, ROPD models the divergence …
Progressive Multimodal Alignment for Continual Instruction Tuning
This paper addresses projector drift in multimodal continual learning by introducing Progressive Multimodal Alignment (PMA). PMA uses a lightweight descriptor to detect visual distribution shifts and progressively adds new "expert" projector components only wh…
Think Short, Defer Smart, Act, and Repeat: Calibrated Reasoning and Uncertainty-Aware Deferral for Edge LLM Agents
This paper introduces Think Short, Defer Smart (TSDS), a framework for edge LLM agents. TSDS optimizes reasoning by stopping early when an action is stable and defers to a cloud model only when local uncertainty is high. Its key contribution is a jointly calib…
DIRECT: Direct Decoding for Efficient and Aligned Sequence Labeling with Large Language Models
DIRECT improves sequence labeling with LLMs by optimizing for human preferences via DPO and enforcing structured outputs. Its core method uses controlled decoding and a template-filling mechanism to generate only label tokens, significantly boosting inference …
From Found to Designed: Concepts as a Design Axis for Large Language Models
This paper proposes treating "concepts" as a deliberate design axis for Large Language Models (LLMs), rather than an emergent property. The authors introduce a taxonomy for designi…
OptimismBench: Forecasting Bias and the Alignment Effect in Language Model Judgment
This paper introduces OptimismBench, a novel method for detecting directional bias in language model probability judgments. By presenting inverted scenarios (e.g., P(success) vs. P…
Budget-Aware LLM Discovery via Cost-Calibrated Frontier Utility
This paper addresses the problem of efficiently discovering useful large language models (LLMs) within a fixed token budget. The core method, **CostAda**, introduces "cost-calibrat…
Can AI agents conduct open-ended AI research? Early evidence from two case studies
This paper introduces "shadow evaluations" to assess AI agents' ability to conduct open-ended AI research. In this method, frontier AI agents tackle the core research question of u…
Scores Are Not Decisions: Cost-Aware Stopping for Tool Acquisition in LLM Agents
This paper addresses the challenge of LLM agents selecting the optimal number of tools from a ranked list, considering their diverse costs. The core method, CAM-DF, formulates this…
The Town Square
This "Show HN" features an open-source engine that enables running the Gemma 4 26B model with just 2GB of RAM on any M-series Mac.
Workshops
Airi is a self-hosted, user-owned Grok companion that brings virtual waifus and cyber beings to life for real-time voice chat and gameplay in Minecraft and Factorio, accessible on Web, macOS, and Windows.
This repository provides an agent harness for performance optimization, featuring skills, instincts, memory, security, and research-first development for various AI code assistants like Claude Code and Codex.
