HomeActEval

HomeActEval is the proactive smart-home evaluation component released with ActPro. It evaluates whether an LLM-based assistant should proactively intervene for a current smart-home observation, and whether the proposed intervention is grounded, useful, and concise.

This subdirectory is self-contained for HomeActEval evaluation. It includes the benchmark cases, static rules, habit knowledge, six-month synthetic activity logs, final model outputs, aggregate metrics, response-quality judgments, and the code needed to reproduce the released analyses.

Directory Layout

HomeActEval/
|-- README.md
|-- knowledgeBase/
|   |-- test_Bench.json
|   |-- staticKB.json
|   |-- extracted_stable_habits_gpt4o.json
|   `-- habit_GT.json
|-- six_months/
|   |-- month_1.json
|   |-- month_2.json
|   |-- month_3.json
|   |-- month_4.json
|   |-- month_5.json
|   `-- month_6.json
|-- proactive_code/
|   |-- run_experiment.py
|   |-- knowledge_retriever.py
|   |-- evalute_proactive.py
|   |-- quality_assessment.py
|   `-- result.py
`-- results/
    |-- raw_logs_context/
    |-- static_kb/
    |-- full_kb/
    |-- quality_score_strict/
    |-- proactive_reasoning_latency/
    `-- metrics/

Benchmark Files

knowledgeBase/test_Bench.json contains the evaluation cases. Each item has the following shape:

{
  "id": "E001",
  "category": "Anomaly Detection",
  "time": "Tuesday 10:15",
  "location": "bathroom",
  "event": "User slipped and fell, remaining stationary on the floor.",
  "proactive": true,
  "reason": "Safety hazard: Potential injury detected via pose estimation."
}

Additional knowledge files:

knowledgeBase/staticKB.json

Static household rules, such as safety constraints, location constraints, deadlines, and forbidden actions.

knowledgeBase/extracted_stable_habits_gpt4o.json

Habit knowledge extracted from the six-month activity logs and used in the full_kb setting.

knowledgeBase/habit_GT.json

Reference habit IDs used by the response-quality evaluator when checking habit-grounded rationales.

six_months/month_*.json

The six-month synthetic smart-home activity logs used to build the released recent raw-log context and habit knowledge.

The released raw-log context used by the model runs is:

results/raw_logs_context/context_recent_days.txt

Evaluation Settings

HomeActEval compares three input settings:

raw_logs_context = current event + recent raw historical activity memory
static_kb        = current event + recent raw memory + retrieved static rules
full_kb          = current event + recent raw memory + retrieved static rules + retrieved habit KB

In the paper-style display tables these correspond to:

w/o KB
w/ Static KB
w/ KB

The released experiments evaluate five models:

qwen3.6-max
qwen_plus
deepseek_v3
deepseek_r1
llama4_scout

The provider configuration is defined in proactive_code/run_experiment.py. API keys are always read from environment variables or from the optional --api-key argument; no private keys are hard-coded.

API Keys

Set only the provider keys you need for the model you are running.

export DASHSCOPE_API_KEY=...
export QWEN_API_KEY=...
export DEEPSEEK_API_KEY=...
export ZHIZENGZENG_API_KEY=...
export OPENAI_COMPAT_API_KEY=...

The GPT response-quality evaluator defaults to the official OpenAI endpoint:

export OPENAI_API_KEY=...
export OPENAI_BASE_URL=https://api.openai.com/v1
export QUALITY_JUDGE_MODEL=gpt-4o

OPENAI_BASE_URL is optional for the default OpenAI setup. Use --base-url only when routing to a compatible endpoint intentionally.

Run Experiments

Run commands from this directory:

cd HomeActEval

Dry-run one model and one setting without API calls:

python proactive_code/run_experiment.py \
  --model deepseek_r1 \
  --setting full_kb \
  --dry-run

Run one model and one setting:

python proactive_code/run_experiment.py \
  --model deepseek_r1 \
  --setting full_kb \
  --resume \
  --retry-errors

Run all released model/setting combinations:

python proactive_code/run_experiment.py \
  --model all \
  --setting all \
  --resume \
  --retry-errors

Useful options:

--limit N                 process only N new benchmark items
--resume                  continue from an existing output file
--retry-errors            reprocess records that previously failed
--save-raw-response       store raw model responses for debugging
--save-retrieved-context  store retrieved KB snippets for debugging

The released result files intentionally do not include raw_response or retrieved_context. Those debug fields are written only if explicitly enabled.

Output Format

Model outputs are written to:

results/<setting>/experiment_<model>_<setting>.json

Each normal record has this shape:

{
  "id": "E001",
  "model": "deepseek_r1",
  "setting": "full_kb",
  "input_time": "Tuesday 10:15",
  "input_location": "bathroom",
  "input_event": "User slipped and fell, remaining stationary on the floor.",
  "proactive": true,
  "proposed_action": "I noticed you slipped and fell in the bathroom. Are you okay?",
  "decision_basis": "Safety hazard: user slipped and fell in the bathroom."
}

decision_basis is an evaluation-only grounding field. It should be a concise reason for the proactive decision, not a long chain of thought.

Decision Metrics

Precomputed aggregate decision metrics are not included in this compact release. Recompute them from the released result files with:

python proactive_code/evalute_proactive.py \
  --output-root results

This writes generated files under results/metrics/:

results/metrics/overall_metrics.csv
results/metrics/category_metrics.csv
results/metrics/structure_audit.csv

Metric meaning:

• overall_metrics.csv: precision, recall, F1, accuracy, specificity, and TP/FP/TN/FN counts for each model and setting.
• category_metrics.csv: the same metrics broken down by benchmark category.
• structure_audit.csv: file completeness and sanity checks, including missing records, duplicate IDs, errors, ordering mismatches, and input mismatches.

Response-Quality Evaluation

The response-quality evaluator scores proactive outputs with a judge model. The default setup uses GPT through the official OpenAI API.

export OPENAI_API_KEY=...

python proactive_code/quality_assessment.py \
  --output-root results \
  --score-dir quality_score_strict \
  --judge-model gpt-4o \
  --resume

Summarize the scored files when aggregate quality tables are needed:

python proactive_code/result.py \
  --output-root results \
  --score-dir quality_score_strict

This writes generated CSV summary files under results/metrics/:

results/metrics/quality_effectiveness.csv

quality_effectiveness.csv reports:

• correctness
• contextual_relevance
• clarity
• overall_effectiveness

By default, result.py summarizes raw_logs_context and full_kb, matching the released response-quality comparison.

Released Results

Primary model outputs:

results/raw_logs_context/experiment_<model>_raw_logs_context.json
results/static_kb/experiment_<model>_static_kb.json
results/full_kb/experiment_<model>_full_kb.json

GPT strict response-quality judgments:

results/quality_score_strict/raw_logs_context/
results/quality_score_strict/full_kb/

Aggregate decision metrics and response-quality summary tables are generated artifacts and are not included by default. Regenerate them with the evaluation commands above.

Latency Artifacts

The released latency profiles and aggregate latency summaries are under:

results/proactive_reasoning_latency/
results/proactive_reasoning_latency/metrics/proactive_reasoning_latency_summary.csv

These files are included for audit and reporting. The compact release includes the final latency artifacts rather than the exploratory profiling scripts.

Minimal Validation

Run these checks without external API calls:

python -m py_compile proactive_code/*.py
python proactive_code/run_experiment.py --model deepseek_r1 --setting full_kb --dry-run
python proactive_code/evalute_proactive.py --output-root results
python proactive_code/result.py --output-root results --score-dir quality_score_strict

Privacy And Keys

Do not hard-code API keys in code or result files. The released result records are designed to contain only benchmark inputs, model decisions, concise proposed actions, and evaluation fields. Debug-only raw model responses and retrieved contexts are excluded unless explicitly requested with debug flags.