MERID: Multimodal Exploration via Recursive Self-Improvement Agents for Major Depression Analysis
MERID builds multimodal depression-detection pipelines through an experience-driven recursive self-improvement (RSI) loop. Each round proposes pipeline modifications, verifies them under a fixed development protocol, and carries the verified pipeline and the accumulated experience into the next round.
- Grounded State Construction (GSC). The agent derives a data adapter for each corpus and verifies it with execution checks and recipe probes, so every feature recipe sees the same subjects, records and targets.
- Coupled Pipeline Exploration (CPE). Representation, fusion and predictor choices are searched jointly as complete pipelines under one shared development protocol.
- Evidence-Guided Evolution (EGE). Planner, Engineer and Reviewer agents analyze the measured outcomes, the Chair turns their reports into revision requests, and a revision is inherited when its gain exceeds a subject-bootstrap margin.
| Step | What happens | Code |
|---|---|---|
| 1. Inspect | The agent profiles the corpus: files, modalities, subjects and label provenance. | adapter_agent.inspect_dataset, label_provenance.py |
| 2. Propose | The LLM proposes the feature-recipe space, the predictor space and, where the benchmark fixes none, the metric. | run_method.py, run_corpus.py, retrieval.py, operation.py |
| 3. Ground (GSC) | The agent writes the data adapter; execution checks and recipe probes verify it, and failed checks trigger repair. | AdapterAgent.create_adapter |
| 4. Explore (CPE) | Recipe × predictor pairs are scored on fixed subject-disjoint folds or the official development split. | DevelopmentSearch.evaluate |
| 5. Evolve (EGE) | The review roles report to the Chair, who decides ACCEPT or REVISE. Revisions are scored and retained past the bootstrap margin, and the research memory is updated. | retrospective.hold_review, development_search.py, research_memory.py |
| 6. Inherit | The retained pipeline, candidate space, history and memory carry into the next round. | development_agent.run_development |
| 7. Test once | The development-selected deployment is scored once on the held-out test set. | corpus_final_eval.py, mpdd2025_final_eval.py, mpdd2026_final_eval.py |
MPDD-2026 runs the search-and-review loop over its binary/PHQ and hierarchical ternary head
families (run_agent.py, benchmarks/mpdd2026/joint_search.py).
.
├── README.md
├── requirements.txt
├── assets/
│ └── merid_banner.png
└── merid/
├── run_corpus.py entry point: E-DAIC, Mental Health, Kaggle
├── run_method.py entry point: MPDD-2025
├── run_agent.py entry point: MPDD-2026
├── core/
│ ├── development_agent.py self-improvement loop: search, review rounds, inheritance
│ ├── grounded_state_construction/ GSC
│ │ ├── adapter_agent.py adapter derivation, execution checks, recipe probes, repair
│ │ ├── adapter_cache.py cache of deterministic adapter outputs
│ │ ├── adapter_source_evidence.py adapter source as review evidence
│ │ └── label_provenance.py target-defining column check
│ ├── coupled_pipeline_exploration/ CPE
│ │ ├── development_search.py joint recipe × predictor search
│ │ ├── registry.py method catalog
│ │ └── operation.py generated estimators for the catalog
│ └── evidence_guided_evolution/ EGE
│ ├── retrospective.py Planner, Engineer, Reviewer and Chair
│ ├── research_memory.py lessons distilled after each review round
│ ├── reliability.py uncertainty-aware verification
│ └── development_diagnostics.py selection evidence for the review
├── llm/ proposer and web-search grounding
├── utils/ run trajectory and progress reporting
├── benchmarks/
│ ├── mpdd2025/ development entry, grouped decoding, estimator factories
│ └── mpdd2026/ data, heads, fold protocol, metrics, head search
└── evaluation/ final evaluation on the held-out test sets
Each module opens with a one-line summary of its role. Paths specific to our computing environment
appear as placeholders such as <SPLIT_ROOT>.
Python 3.10 or later:
pip install -r requirements.txt
export ANTHROPIC_API_KEY=<YOUR_API_KEY> # ANTHROPIC_BASE_URL is read as wellRun the commands from the repository root; arguments in angle brackets are local paths. The datasets are not included.
Development run. The agent searches and reviews on training and development data only and writes
its results under --output_dir. The flags are the settings of the paper's runs.
# E-DAIC; Mental Health and Kaggle also run through run_corpus
python -m merid.run_corpus --evaluation development --data_root <EDAIC_TRAIN_DEV_ROOT> \
--ctx E-DAIC --task binary --output_dir <OUT_DIR> --seed 0 \
--development_folds 3 --max_evaluations 256 --retrospective 1 --review_checks 2 \
--selection_metric macro_f1 --selection_rule incumbent --promotion_se 1.0 \
--reliability_gate off --require_llm
# MPDD-2025, one track/window/task cell
python -m merid.run_method --evaluation development --data_root <MPDD2025_TRAIN_ROOT> \
--ctx MPDD-Elderly/1s --task binary --output_dir <OUT_DIR> --seed 0 \
--development_folds 3 --max_evaluations 256 --retrospective 1 --review_checks 2 \
--selection_metric official_composite --selection_rule incumbent --promotion_se 1.0 \
--reliability_gate off --require_llm \
--decoder_strategies plain,subject_consensus,metric_bias,metric_bias_consensus
# MPDD-2026, Track1 (Elder) or Track2 (Young)
python -m merid.run_agent --evaluation development --track Track1 \
--train_root <MPDD2026_TRAIN_ROOT> --personality_npy <PERSONALITY_NPY> \
--split_root <SPLIT_ROOT> --fold_task binary --n_folds 10 \
--search_mode joint_heads --budget 12 --fit_budget 720 \
--retrospective 1 --review_checks 2 --revision_reserve 0.25 \
--seed 0 --output_dir <OUT_DIR> --require_llmPipelines inherited from an earlier round enter the search through --reference_candidates <JSON>
(run_corpus, run_method) or --reference_components <JSON> (run_agent).
Final evaluation. The development-selected deployment is scored once on the held-out test set:
python -m merid.evaluation.corpus_final_eval --summary <OUT_DIR>/development_summary.json \
--adapter <ADAPTER_PY> --ctx E-DAIC --members 1se --topk 5
python -m merid.evaluation.mpdd2025_final_eval --summary <OUT_DIR>/development_summary.json \
--adapter <ADAPTER_PY> --ctx MPDD-Elderly/1s --members 1se --topk 5
python -m merid.evaluation.mpdd2026_final_eval --results <RUN_DIR>/results.json --track Track1 \
--split_root <SPLIT_ROOT> --answer_zip <ANSWER_ZIP> --topk 5<ADAPTER_PY> is a frozen adapter that also loads the test partition, and <RUN_DIR> is the run
directory run_agent creates under --output_dir; for MPDD-2026, --reference_heads <JSON> applies
the incumbent rule to inherited heads. Each script lists its options with --help.
The prompts reproduced in the appendix are defined in these modules (paths relative to merid/):
| Prompt | Definition |
|---|---|
| Benchmark metric | run_method.py EVAL_METRIC_SYSTEM |
| Feature-recipe space | run_corpus.py RECIPE_SYSTEM; run_method.py RECIPE_SPACE_SYSTEM |
| Classifier and regressor spaces | run_method.py CLF_SPACE_SYSTEM, REG_SPACE_SYSTEM |
| Model-code generation | run_method.py CODEGEN_SYSTEM |
| Data adapter derivation and repair | core/grounded_state_construction/adapter_agent.py ADAPTER_SYSTEM |
| Method selection | core/coupled_pipeline_exploration/registry.py select_method |
| MPDD-2026 head proposer | benchmarks/mpdd2026/joint_search.py HeadProposer.propose |
| Planner, Engineer, Reviewer, Chair | core/evidence_guided_evolution/retrospective.py ROLE_SYSTEM, CHAIR_SYSTEM, REVISION_REQUIRED, _INVESTIGATE |
| Distill and Reconcile | core/evidence_guided_evolution/research_memory.py DISTILL_SYSTEM, RECONCILE_SYSTEM |
| Web-search clause | llm/retrieval.py OPTIONAL_SEARCH_CLAUSE |