arxiv arXiv cs.CL · 2h ago · research

Recuris improves long-horizon agent success by 17.8 points on tau-bench

from English

Researchers introduce Recuris, a recursive Experiential-Working Memory architecture designed to address the challenges of recursive self-improvement in long-horizon tasks. The system uses Working Memory to track progress and guide skill selection from Experiential Memory, creating a bounded loop that localizes failures and updates skills based on current needs.

Recuris improves task success in 35 of 37 completed model-benchmark pairs across four benchmarks.
On tau-bench, it adds +17.8 points to GPT-5.6 Sol and +15.6 to Claude Opus 5, reaching 87.9%.
It adds +16.6/+13.5 points on Qwen3.6-27B/35B on SkillFlow.
The advantage widens as interaction horizons grow, adding +32.2 points on the longest tasks.
Common long-horizon failures are reduced by up to 80%.

This approach positions recursively evolving memory as a scalable foundation for recursive self-improvement, enabling agents to continuously transform accumulated experience into more effective behavior.

Importance 2/3 New harness with differentiators arXiv cs.CL AI agents Research paper

Benchmarks

Benchmark	Model	Score
τ-bench	Claude Opus 5	87.9%

Read original

arxiv arXiv cs.AI · 2h ago

Recuris introduces recursive Experiential-Working Memory architecture for long-horizon agent harnesses

Researchers introduce Recuris, a recursive Experiential-Working Memory architecture designed to address the challenges of recursive self-improvement in long-horizon tasks. The system uses Working Memory to track task progress and guide skill selection from Experiential Memory, grounding decisions in current needs rather than full history.

arxiv arXiv cs.LG · 23h ago · 1 view

Apodex 1.1 scales agentic intelligence via environment and coordination scaling

Apodex 1.1 introduces "working capability" to enable sustained, verifiable progress toward real-world objectives through two complementary dimensions: Environment Scaling and Agentic Coordination Scaling.

arxiv arXiv cs.AI · 1d ago · 2 views

Apodex 1.1 scales agentic intelligence via environment and coordination scaling

Apodex 1.1 introduces a framework for "working capability," enabling language models to sustain verifiable progress on complex objectives through file interaction, code execution, and state maintenance. The system achieves this by expanding executable environments and training agents to decompose long-horizon tasks, delegate parallel work, and replan using a shared execution harness called AgentOS.

arxiv arXiv cs.CL · 1d ago · 1 view

Apodex 1.1 scales agentic intelligence via environment and coordination scaling

Apodex 1.1 introduces a framework for "working capability," enabling general-purpose language models to sustain verifiable progress on complex tasks involving files, code, and information sources. The system achieves this through Environment Scaling, which expands executable environments, and Agentic Coordination Scaling, which trains agents to decompose tasks and manage parallel work.

media Hugging Face Forums · 5d ago · 4 views

SSU Physics 1.2 proposes formalizing emergence to test against physical systems

The Science of Unified Systems (SSU) program introduces SSU Physics 1.2 and the Atlas of Emergence to investigate how coherent structures and information emerge from component interactions in complex systems.