Skip to content

Results

rrahimi-uci edited this page Jun 30, 2026 · 1 revision

Results

Reproducible, in this repo (offline teaching environment, no API key)

These come straight from the bundled examples (examples/*.py) and are fully deterministic:

Demo Base LLM ACE Δ
Quickstart (offline → held-out eval) 44.4% 83.3% +38.9 pts
Context-collapse benchmark (online) 41.7% 88.3% +46.6 pts
Offline warmup + online 34.5% 96.6% +62.1 pts

In the context-collapse demo, the monolithic-rewrite baseline collapses its context 7× and stalls at 60.0%, while ACE never collapses. Adaptation token ingestion for ACE is −94.9% vs. full re-ingestion (deltas are tiny). Generate the visual report with ace demo --html report.html.


Reported in the paper (real benchmarks, DeepSeek-V3.1)

Benchmark Baseline + ACE
AppWorld (agent, avg) 42.4% (ReAct) 59.5% (+17.1)
FiNER (financial NER) 70.7% 78.3%
Formula (financial reasoning) 67.5% 85.5%
Adaptation latency (offline AppWorld) −86.9%
Token cost (online FiNER) −83.6%

On the AppWorld leaderboard, ReAct+ACE with an open-source model matches the top-ranked production GPT-4.1 agent and surpasses it on the harder test-challenge split. Numbers above are from the paper; this repo reproduces the mechanism and its qualitative behavior offline.


Why serving stays cheap

The paper reports 91.8% of input tokens served from KV cache during evaluation, cutting billed input cost ~82.6% vs. counting raw tokens — so a longer playbook does not mean linearly higher serving cost.

See How It Works for why incremental deltas keep adaptation cheap.

Clone this wiki locally