AI Papers to Watch This Week

This week's paper collection covers three main trends: real-world agent evaluation, token and data efficiency, and experience-based adaptive agents.

• ORCA-bench: Proposes an agent evaluation benchmark that synthesizes logs, metrics, traces, and source code to find the root causes of real software failures.
• HANDBOOK.md: Measures the ability to continuously reference and comply with long-form internal policies and SOPs during long-horizon tasks.
• AREX: Presents a deep research agent framework that verifies intermediate results and recursively supplements missing evidence.
• MemoHarness: Dynamically optimizes an agent's control harness for new environments by leveraging past execution experience.
• SkillSmith: Combines textual knowledge with skills embedded in model weights to improve compositional generalization on complex problems.

Efficiency-focused research includes OmniScope, which independently compresses audiovisual information; Tokens are All You Need, which converts recommendation system embeddings into discrete tokens; Train Smarter, Not Longer, which analyzes a model's memory window to optimize the timing of data reuse; and HA-MoE for heterogeneous content feeds. Also covered are an analysis showing that safety fine-tuning can suppress capabilities related to mind attribution and spiritual beliefs, and a restoration method through internal representation manipulation.