Code that accompanies the paper release for "LLMs Corrupt Your Documents When You Delegate"

Fixing GRPO training collapse in long-horizon multi-tool agents. A lightweight PRM-Lite + LATA joint approach achieves +37% over vanilla GRPO on τ-bench airline (50-task, multi-turn).

🧠 Awesome Memory-VLA: A curated list of Visual-Language-Action models with memory

SWE-Marathon: an ultra long-horizon SWE benchmark

Code for Scalable Offline Model-Based RL with Action chunking

Code for Tackling Long-Horizon Tasks with Model-based Offline Reinforcement Learning

The simplest way to build long-horizon environments

A family of long-horizon software-engineering environments for OpenEnv, adapted from https://github.com/Proximal-Labs/frontier-swe

VLM-RL Hierarchical Loco-Manupilation For Long-Horizon Tasks With G1 robot in Isaac Lab/Sim

Local-first, eval-first memory for long-horizon AI agents — no LLM at ingest. Python SDK + MCP server with source-traceable recall, belief revision, selective forgetting, and reproducible benchmarks.

Enhancing Laboratory 3D Perception and Long-Horizon Planning via Protocol-Aligned Action Prediction

OpenClaw humanity infusions OtherPowers Creative Intelligence Agency. 🦞

A real-world inspired environment for selective context retention under noise. It evaluates an LLM's ability to manage a fixed-capacity memory buffer, retaining high-value information while filtering out distractors

Steady-Forcing: Balancing Spatial Persistence and Motion Continuity in Long-Horizon Nature Video Diffusion

Dashboard for real-world inspired environment for selective context retention under noise. It evaluates an LLM's ability to manage a fixed-capacity memory buffer, retaining high-value information while filtering out distractors

Long-horizon agent execution harness — reliable autonomous runs for Claude Code, Codex, OpenHands, and custom agents. Goal graphs, spin detection, HITL gates, fork/merge, 8 strategies, 6 validators.

bstack P12 — Persistent Loop Discipline. Cross-context restart loop with state in the filesystem. Closes the long-horizon context-rot failure mode. Composes with bstack P5/P6/P7/P10/P11.

TPipe is the agent operating environment for deterministic, multimodal AI systems. Built Kotlin-first, it composes runtime substrates into governed pipelines with rich tracing, disciplined context and token control, native function binding, and provider-agnostic execution for long-running, headless agents.

Official repository for the paper: Can AI Agents Synthesize Scientific Conclusions?