arxiv arXiv cs.AI · 1h ago · research

AREX introduces recursively self-improving deep research agents

from English

Researchers introduce AREX, a family of Recursively Self-Improving (RSI) deep research agents designed to address the discovery-verification asymmetry in complex queries. AREX alternates between an inner loop that gathers evidence and constructs provisional answers, and an outer loop that audits constraints to guide targeted refinement.

The system employs an autonomous context-update tool to compress interaction history into a compact state without external models. Training utilizes agentic mid-training and long-horizon reinforcement learning on verified synthetic tasks, emphasizing steps where decisive evidence is acquired or errors are corrected. The team instantiates dense 4B and 122B-A10B Mixture-of-Experts versions of the model.

Across benchmarks including BrowseComp, WideSearch, DeepSearchQA, and Humanity's Last Exam, AREX substantially outperforms comparable-scale baselines while remaining competitive with models using significantly more activated parameters.

Importance 2/3 arXiv cs.AI AI agents Research paper

Read original

arxiv arXiv cs.CL · 23h ago

Solar Open 2 Technical Report

The Solar Open 2 Technical Report introduces Solar Open 2, a 250B-A15B Mixture-of-Experts language model designed for long-horizon agentic tasks. It achieves a 1M-token context window through a hybrid attention stack and is trained using efficient initialization from Solar Open 1 and curated high-value data.

media Hugging Face Forums · 2d ago

SEMA explores non-transformer AI architecture using Vector Symbolic Architecture

SEMA is an experimental AI architecture that investigates whether language understanding and reasoning can emerge from explicit semantic structure rather than next-token prediction. It serves as a non-parametric, instance-based reasoning system utilizing Vector Symbolic Architecture (VSA/HDC) over content-addressable memory.

arxiv arXiv cs.AI · 7d ago

HealthClaw: open-source agent with self-evolving memory for longitudinal personal health management

Researchers developed HealthClaw, an open-source agent architecture designed to update support as a person's routines, preferences, measurements, and risks change over time. It separates shared safety rules and medical knowledge from private longitudinal memory containing profile facts, reusable procedures, and episodic traces.

arxiv arXiv cs.LG · 7d ago

TRACE improves long-horizon agent tool-use via turn-level credit estimation

The authors propose TRACE (Turn-level Reward Assignment via Credit Estimation), a dense credit-assignment method for agentic reinforcement learning that addresses the sparsity and high variance of outcome rewards in long-horizon tasks. By representing rollouts as state transitions at tool-call boundaries and deriving per-action rewards from log-ratio state values, TRACE enables effective training without additional critics or supervised fine-tuning stages.

arxiv arXiv cs.AI · 7d ago

DVM-HALL model and NHAS metric address autonomous commerce loyalty

The article introduces the Dynamic Verifiable Multi-Agent Human Agentic Loyalty Loop (DVM-HALL) model to address how traditional customer loyalty paradigms fail when AI agents execute purchasing decisions autonomously. It proposes a framework where brand choice is determined by human emotional equity, agentic utility, calibrated trust, and verifiable execution risks in decentralized finance environments.