Individual contributions of Reflector, multi-round refinement, and Offline warmup
The paper conducts three ablation experiments on AppWorld to validate ACE's key design choices.
1. Reflector is the core of role separation design
From 55.1% → 56.8% (+1.7), separating evaluation and distillation from generation is the first layer of significant gains.
2. Multi-round refinement further improves performance
From 56.8% → 59.4% (+2.6), 5 rounds of Reflector iteration significantly strengthen experience distillation quality.
3. Offline warmup yields substantial gains on challenge split
Online goes from 56.1% → 59.5%, with test-challenge TGC rising from 61.4% → 66.0% (+4.6) and SGC from 43.2% → 48.9% (+5.7).
Ablation Dimensions
Reflector
Distills experiences from traces, independent of Curator
Multi-Round Refinement
Reflector iteratively refines experiences
Offline Warmup
Preheats Playbook with training set before Online phase
AppWorld Offline Ablation
| Variant | GT | Test-Normal | Test-Challenge | Average | ||
|---|---|---|---|---|---|---|
| TGC | SGC | TGC | SGC | |||
| ReAct (baseline) | – | 63.7 | 42.9 | 41.5 | 21.6 | 42.4 |
| ACE without Reflector or multi-round | ✓ | 70.8 | 55.4 | 55.9 | 38.1 | 55.1 |
| ACE without multi-round | ✓ | 72.0 | 60.7 | 54.9 | 39.6 | 56.8 |
| Full ACE | ✓ | 76.2 | 64.3 | 57.3 | 39.6 | 59.4 |
AppWorld Online Ablation
| Variant | GT | Test-Normal | Test-Challenge | Average | ||
|---|---|---|---|---|---|---|
| TGC | SGC | TGC | SGC | |||
| ACE Online | ✗ | 67.9 | 51.8 | 61.4 | 43.2 | 56.1 |
| ACE Online + Offline warmup | ✗ | 69.6 | 53.6 | 66.0 | 48.9 | 59.5 |
Three Conclusions
1. Reflector is the core of role separation design
From 55.1% → 56.8% (+1.7), separating evaluation and distillation from generation is the first layer of significant gains.
2. Multi-round refinement further improves performance
From 56.8% → 59.4% (+2.6), 5 rounds of Reflector iteration significantly strengthen experience distillation quality.
3. Offline warmup yields substantial gains on challenge split
Online goes from 56.1% → 59.5%, with test-challenge TGC rising from 61.4% → 66.0% (+4.6) and SGC from 43.2% → 48.9% (+5.7).
Offline warmup provides an initial Playbook already containing strategies, allowing Online adaptation to start from a higher baseline.
Design Trade-offs
| Component | Value Retained |
|---|---|
| Reflector | Clear responsibility for "experience distillation" |
| Multi-round refinement | Higher-quality experience entries |
| Offline warmup | Reduces Online learning curve |