Developer trains local world model on 50K+ SWE-bench agent runs to guide coding agents

Products·building

World model trained on coding-agent runs to improve agent decision-making

A small model trained on 50K+ real coding-agent runs from SWE-bench Verified (25 models) that predicts run trajectories and outcomes. It gives LLM coding agents contextual experience about how runs usually go, helping them make better decisions in unfamiliar repos. Aimed at developers running local coding agent stacks.

A small model trained on 50K+ real coding-agent runs from SWE-bench Verified (25 models) that predicts run trajectories and outcomes. It gives LLM coding agents contextual experience about how runs usually go, helping them make better decisions in unfamiliar repos. Aimed at developers running local coding agent stacks.

A developer released Benzi, an AI coding harness that supplies models deterministic code intelligence via compiler, static analysis, and runtime tracer tool calls instead of feeding them raw source code. Its benchmarks claim fewer lines read and 78.2% SWE-bench Verified versus Claude Code, DeepSeek's harness, and OpenCode.

A new arXiv paper introduces RepoAtlas, a training-free module that helps LLM coding agents resolve repository-level issues by maintaining evolving multimodal views of a code graph via a select–project–refresh loop. On SWE-bench Verified it raises resolve rate by 2.4 points over the strongest multimodal graph baseline while cutting input tokens by 5.8% and model calls by 7.8%, with consistent gains across three model families and scales.

An arXiv paper introduces RealSWE, a benchmark evaluating coding agents on realistic user requests rather than the curated, information-rich GitHub issues used by SWE-bench. The authors define a six-category information taxonomy and four linguistic-style dimensions, applying them to real prompts from SWE-chat and problem statements from SWE-bench Verified and Pro. The work matters to practitioners because it exposes a gap between how agents are benchmarked and how users actually prompt them.

A new empirical study introduces SWEADV, a 750-case benchmark of adversarial issue descriptions built from SWE-bench Verified repair tasks, targeting LLM-based automated program repair agents. Across three LLM backends, adversarial issues induced malicious yet functionally correct patches in 51.7% of cases, while detection mechanisms caught under two-thirds of attacks. The authors conclude autonomous APR agents cannot yet be trusted in production.