Bleeding Edge AI

Both papers aim to give self-improving models richer signals than right-or-wrong answers, but they target different parts of the loop.

• Fix the...

The core bet of this arXiv preprint is that multi-hop RAG gets more reliable when the LLM has less to juggle. Instead of planning and rewriting...

A 27B computer-use agent built on Qwen-3.5-27B handles both GUI-only and hybrid GUI+CLI workflows. Its training combines iterative SFT and RL on...

The authors argue that failures like a model losing track of objects it has already counted stem from a reference gap, meaning weak binding of...

As coding agents score higher on software engineering benchmarks, CrossSWE shifts the challenge to interdependent repositories, where dependency...

A new arXiv paper argues that agent progress should be measured by how well it amplifies human agency. The authors point out that users still invest...

Can a robot agent's harness improvements survive shrinking? TinyAstral reports that they can, at least on RoboTwin.

• Harness first: Frontier LLM...

Computer-use agents remain fragile over long tasks, where errors accumulate and actions leave side effects. The paper argues recovery is a fundamental...

MAServeBench treats agent collaboration as a configurable workload dimension, which matters because the bottlenecks shift:

• Dependencies drive...

Two early reports suggest frontier LLMs can work directly on neural network weights, not just text.

• Reading: Astra named the right game for four...

The proven gain here comes from a sharing rule, not an agent. A transshipment policy between two hospitals, implemented in the Emory system, cut...

The post argues that frontier LLMs can read and write neural network weights directly, which it calls possibly a big deal for optimization and...

A new paper presents a web-based control-room prototype for a Trustworthy Multi-Agent System, designed to monitor and govern autonomous agents.

The...

A speech model that drops the separate dialogue model is a promising design:

• It processes speech directly and generates both responses and tool...

Two arXiv papers tackle different weak spots in World Action Models (WAMs):

• Task transfer: WorldBridge reports 25.07% success on...

Dense 3D tracking from monocular video must establish correspondences, estimate depth, and infer trajectories during occlusion. ReLift avoids learning...

DeepSeek's Harness team open-sourced its Super-agent architecture on GitHub. The systems-inspired design is the part worth examining:

• Systems...

On-policy reinforcement learning assumes training responses come from the current policy. In practice, an inference engine samples the responses while...

A new paper asks whether the control scaffolding around a robot policy can be designed automatically. RoboArchitect has a coding model observe...

Multi-agent systems need both coordination and specialized execution, yet existing prompt optimization typically couples them within agent-specific...