arxiv arXiv cs.AI · 22h ago · research

World State Generator aligns plans with synthetic worlds to improve agent success

from English

The World State Generator (WSG) is a model that keeps language agent plans aligned with the rules of their execution environment by rewriting states after failures. It is trained on 226K trajectories extracted from seven synthetic domains where a program enforces rules and verifies goal reachability.

WSG takes a failure as a rule stated by the world and rewrites remaining states to obey it.
The model writes plans as checkable states of the world rather than just language instructions.
Across seven public benchmarks, WSG raises end-to-end success for two open models near 30B parameters over prompting.
This performance brings the open models to the level of proprietary models.

WSG addresses the issue where prompted planners fail by rewriting refused steps without moving forward, ensuring plans bend to the world's physical limits and logical dependencies as the run progresses.

Importance 2/3 New harness with differentiators arXiv cs.AI AI agents Training methods

Read original

media Latent Space · 1d ago · 13 views

TypeSafe AI introduces Jev, a System One model trained with Reinforcement Learning for Calibrated Decisions

TypeSafe AI has launched Jev, a new class of "System One" models designed for production environments. The company's CEO and co-author of the InstructGPT paper, Diogo Almeida, argues that current frontier models have prioritized alignment and refusals over reliability, leading to issues like hallucinations and sycophancy.

arxiv arXiv cs.CL · 14d ago · 25 views

ExecCritic uses role-specific RL to improve coding agents via test-guided repair

Researchers introduce ExecCritic, a framework that combines a test-verify-revise scaffold with role-specific reinforcement learning to enhance coding agents. The system separates test construction from source-code repair, using a Test agent to generate repository-native tests and a Repair agent to revise code based on execution feedback.

lab Hugging Face Blog · 29d ago · 44 views

IBM releases Granite 4.2 reasoning LLMs with agentic capabilities

IBM has released Granite 4.2, a family of dense, decoder-only reasoning large language models in 3B, 8B, and 30B sizes. These models are pre-trained from scratch on approximately 15 trillion tokens and feature a context window extended to 512K tokens.

media Hugging Face Forums · 34d ago · 27 views

Orchard Compute automates model training setup via MCP

The author introduces Orchard Compute, a tool designed to automate the infrastructure aspects of model training. By connecting through the Model Context Protocol (MCP), users can describe their requirements in natural language to handle compute selection and environment provisioning.

media Latent Space · 34d ago · 72 views

Z.ai proposes post-training scaling law and releases GLM 5.3; Ornith-1.5 launches with self-improvement

Z.ai CEO Jie Tang argues that parameter count is insufficient for evaluating models, proposing a new scaling law based on data, compute allocation, and inference conditions. This framework underpins GLM 5.3, which achieves significant performance gains through reinforcement learning on long-horizon environments synthesized via recursive self-improvement pipelines.