ReflexGrad: Within-Episode Failure Recovery in LLM Agents via Progress-Gated Dual-Process Routing

ReflexGrad introduces a dual-process architecture enabling LLM agents to recover from failures within a single episode without requiring demonstrations. The system combines fast continuous refinement with slow causal diagnosis, achieving significant performance improvements on benchmark tasks with smaller models matching larger model performance through architectural innovation rather than scale.

ReflexGrad addresses a fundamental limitation in LLM agent design: the inability to recover from early mistakes within an episode. Traditional approaches either proceed linearly through failures or require external demonstrations, wasting computational budget. This architecture elegantly separates concerns by routing between two processes—a fast refinement loop running every three steps and a slower diagnostic process triggered after five consecutive low-progress signals. The deterministic priority merge preserves natural language coherence while enabling reproducible recovery mechanisms.

The research emerges from growing recognition that agent reliability depends less on model size than on algorithmic structure. Recent work in chain-of-thought reasoning and reflexive architectures demonstrated that systematic self-correction improves performance, yet within-episode recovery remained underexplored. ReflexGrad fills this gap by making failure recovery observable and verifiable, with explicit triggers, diagnostics, and fixes that can be analyzed and refined.

For AI practitioners, these results suggest that architectural improvements may offer better returns than scaling alone. The 1.5 percentage point cross-model gap (Qwen-3-8B vs GPT-4 level) being statistically insignificant indicates the routing mechanism drives gains rather than parameter count. This has implications for deployment costs and model selection strategies. The comprehensive release of code, prompts, and per-seed logs accelerates community adoption and reproducibility, critical for establishing trust in agent systems.

Key watching points include whether this dual-process pattern generalizes beyond ALFWorld to real-world tasks, how the approach scales to longer episodes, and whether similar routing mechanics improve other sequential decision-making domains. Performance gains of 40+ percentage points justify investment in understanding what makes recovery architectures effective.

• →ReflexGrad achieves 75.4% performance on Qwen-3-8B, a 40.3pp improvement over baseline without demonstrations
• →The dual-process routing mechanism outperforms larger models and compute-matched baselines like LATS and Tree-of-Thought
• →Architectural design matters more than model scale—cross-model performance gap was statistically insignificant (1.5pp)
• →Fast refinement (every 3 steps) combined with slow diagnosis (triggered by 5 low-progress signals) enables within-episode failure recovery
• →Full reproducibility artifacts released including code, prompts, and per-seed logs for community adoption