Skip to content

ACCORD: Action-Conditioned Contextual Grounding for Language Agents

🕒 Published (v1): 2026-06-15 09:05 UTC · Source: Arxiv · link

Ask a follow-up

Open an assistant pre-loaded with this paper's context.

💬 Ask ChatGPT✦ Ask Claude

TL;DR

ACCORD (Action-Conditioned Contextual Grounding) is a training-free, model-agnostic framework that intercepts LLM agent write actions to resolve underspecified user instructions via two complementary mechanisms: re-surfacing overlooked trajectory evidence and probing the environment for missing facts. It improves task-goal completion on AppWorld by up to +20.6 points (GPT-5-mini, test-challenge: 42.0% → 62.6%) and generalizes to embodied settings (AlfWorld +7.4%).

Problem

LLM agents operating in rich digital/physical environments receive underspecified user instructions whose disambiguating context lives in the environment, not the prompt. Current agents exhibit two systematic grounding failures: (1) incomplete — acting without first probing the environment for required affordances/argument values (26.7–36.7% of failures); (2) overlooked — contradicting information already present in the trajectory context (33.3–66.7% of failures). Existing self-correction approaches (Reflexion, Self-Refine) fail because they reflect on internal model beliefs rather than objective environmental signals.

Method

ACCORD operates at two layers:

Inference-time layer (primary): Before each write action w ∈ W, a separate grounding agent intercepts execution. It receives the proposed write, prior trajectory c_t, and task instruction. It performs: (A) re-surfacing decisive facts already in c_t but unattended (addresses overlooked failures), and (B) issuing read-only environment probes for facts not yet in c_t (addresses incomplete failures). The grounding agent emits an APPROVE/REJECT decision; on REJECT, the conflicting environmental evidence (not model reasoning) is appended to the main agent's context and the main agent re-proposes. Write action set W is constructed automatically from API metadata (HTTP method categorization) or a one-time LLM classification pass.

Policy layer (supporting): A prompt appended to the main agent's system prompt encourages early environmental exploration (pre-exploration prompt, PE) and inspecting concrete content before acting (grounding prompt, GP). No new components or training required.

Key Contributions

  • Taxonomy of LLM agent grounding failures into incomplete and overlooked modes, quantified on annotated failure trajectories from AppWorld and AlfWorld
  • ACCORD framework: action-conditioned context augmentation via a grounding agent that intercepts writes and grounds them in objective environmental signals rather than model self-critique
  • Empirical demonstration that post-hoc self-correction (Self-Refine, FullCodeReflex) degrades performance because it compounds ungrounded assumptions
  • Training-free, model-agnostic design transferring across GPT-5-mini, Claude-4.5-sonnet, and Qwen3.5-27B-FP8 without modification

Results

  • AppWorld test-challenge (GPT-5-mini): ACCORD 62.6% TGC vs. ReAct 42.0% (+20.6), SGC 38.9% vs. 20.9% (+18.0)
  • AppWorld test-normal (GPT-5-mini): 78.0% TGC vs. 63.7% (+14.3), 62.5% SGC vs. 46.4% (+16.1)
  • AppWorld (Claude-4.5-sonnet): +4.2 TGC test-normal, +10.8 TGC test-challenge
  • AppWorld (Qwen3.5-27B-FP8): +10.1 TGC test-normal, +6.7 TGC test-challenge
  • AlfWorld unseen split (GPT-5-mini): 78.0% vs. 70.6% (+7.4, 85/109 tasks)
  • AlfWorld (Qwen3.5-27B-FP8): 88.1% vs. 80.7% (+7.4)
  • Self-Refine and FullCodeReflex underperform ReAct baseline on GPT-5-mini
  • Read/write ratio shifts from 70.4%/29.6% to 82.7%/17.3% (test-normal), showing gains come from better-targeted writes not more writes

Limitations

  • Requires write/read-only API categorization; if unavailable, needs a one-time LLM classification pass
  • Grounding agent adds read-only API calls and additional model rollouts per write action, increasing per-task token cost
  • Policy layer (grounding prompt) was hand-designed; not optimized or distilled into the main agent
  • Grounding behavior currently externalized to a separate agent rather than internalized by the main model

Relevance to Harnesses / Meta-Harnesses

ACCORD is a textbook meta-harness pattern: it wraps an existing agent loop (ReAct) with an outer control layer that intercepts, inspects, and augments actions before they execute, without modifying the inner agent or requiring retraining. The grounding agent acts as a guard harness — a second LLM pass that gates state-mutating actions behind an evidence-check — exactly the pattern used in verification harnesses, tool-call validators, and multi-stage agent pipelines. The automatic construction of the write action set W from API metadata mirrors how meta-harnesses discover action boundaries programmatically. The APPROVE/REJECT feedback loop with evidence injection is directly analogous to reflective harness architectures that route outputs through a critic before committing.