ACCORD: Action-Conditioned Contextual Grounding for Language Agents

A new framework called ACCORD helps large language model agents complete tasks more reliably by actively probing their environment for missing information before taking action, according to a preprint posted to arXiv on June 15, 2026 [1][2]. The framework, whose name stands for Action-Conditioned Contextual Grounding, addresses a persistent weakness in current LLM agents: they often act on assumed specifics rather than observed evidence, overlook information they could have gathered, and fail to incorporate evidence already returned [2]. ACCORD intervenes before each action, probing the environment for missing context and integrating relevant details from the agent's own trajectory that would otherwise be ignored [2]. The approach requires no additional training or task-success signals [2]. On the AppWorld benchmark, ACCORD lifted task-goal completion with GPT-5-mini from 42.0 percent to 62.6 percent, a gain of 20.6 percentage points over strong baselines [2]. The improvement held across model scales: Claude-4.5-sonnet gained 10.8 points, and the open-weight Qwen3.5-27B-FP8 gained 10.1 points [2]. On the embodied AlfWorld benchmark, GPT-5-mini's success rate rose by 7.4 points [2]. The work appears on arXiv, the open-access preprint repository that has hosted machine-learning breakthroughs since its founding in 1991 [4]. The repository now receives roughly 24,000 submissions per month and has served as the initial venue for foundational papers such as "Attention Is All You Need," the 2017 Google paper that introduced the transformer architecture underpinning modern LLMs [4][6]. That paper has since been cited more than 250,000 times [6]. The ACCORD preprint has not yet been peer-reviewed, consistent with arXiv's moderation-but-not-peer-review model [4]. The authors argue that underspecified user instructions are a fundamental challenge for LLM agents because humans rely on implicit assumptions about their surroundings that cannot be inferred from the instruction alone [2]. By conditioning each action on freshly gathered environmental context, ACCORD narrows the gap between what a user says and what an agent needs to know to act correctly [2].