ACCORD framework boosts LLM agent task completion by up to 20.6 points
ACCORD, a training-free agent framework by Lai Jiang, Cheng Qian, and Zhenhailong Wang, improves LLM agent task-goal completion by actively grounding missing context before each action, achieving gains of up to +20.6 points on the AppWorld benchmark.
Score breakdown
ACCORD demonstrates that a training-free grounding layer can close a substantial portion of the task-completion gap in LLM agents across both digital and embodied benchmarks, without modifying the underlying model.
- 01ACCORD stands for Action-Conditioned Contextual Grounding, a framework for adaptive grounding in LLM agents.
- 02Before each action, ACCORD probes the environment for missing information and integrates relevant context from the agent's prior trajectory.
- 03ACCORD requires no additional training or task-success signals.
Lai Jiang, Cheng Qian, and Zhenhailong Wang identify a fundamental weakness in current LLM agents: user instructions are frequently underspecified, relying on implicit assumptions that agents cannot recover from the instruction alone. In information-rich digital and physical environments, agents must instead ground missing context in observed evidence from tools, data, interfaces, and observations — and carry that evidence forward into subsequent actions. The paper demonstrates that current agents routinely fail at this, acting on assumed rather than observed specifics, overlooking information they could have gathered, and failing to incorporate evidence already returned to them.
Critically, ACCORD requires no additional training or task-success signals, making it a plug-in improvement over existing agents.
To address this, the authors propose ACCORD (Action-Conditioned Contextual Grounding), a simple framework that, before each action, actively probes the environment for missing information and integrates relevant context from the agent's trajectory that would otherwise be overlooked. Critically, ACCORD requires no additional training or task-success signals, making it a plug-in improvement over existing agents.
On the AppWorld benchmark, ACCORD improves task-goal completion by up to +20.6 points with GPT-5-mini, lifting performance from 42.0% to 62.6% compared to strong baselines. The gains generalize across model families: +10.8 points with Claude-4.5-sonnet, +10.1 points with the open-weight Qwen3.5-27B-FP8, and +7.4 success rate on the embodied AlfWorld benchmark with GPT-5-mini.
Key facts
- 01ACCORD stands for Action-Conditioned Contextual Grounding, a framework for adaptive grounding in LLM agents.
- 02Before each action, ACCORD probes the environment for missing information and integrates relevant context from the agent's prior trajectory.
- 03ACCORD requires no additional training or task-success signals.
- 04On AppWorld with GPT-5-mini, ACCORD improves task-goal completion from 42.0% to 62.6%, a gain of +20.6 points.
- 05Gains also observed with Claude-4.5-sonnet (+10.8 points) and open-weight Qwen3.5-27B-FP8 (+10.1 points).
- 06On the embodied AlfWorld benchmark, ACCORD achieves +7.4 success rate with GPT-5-mini.
- 07The paper identifies that current agents act on assumed rather than observed specifics and fail to incorporate already-returned evidence.
Topics
Summary and scoring are generated automatically from the original article. We always link back to the publisher and never republish images or paywalled content. Last processed Jun 16, 2026 · 23:11 UTC. How this works →