LifeMem replication (arXiv:2608.19621)
Replication of Mitigating Identity Essentialism in LLM Agents with Longitudinal Life Trajectories (Wang, Zhou, Du, Su, Cao, Pan, Ai, Wu, Zhang, Liu). Official code: halsayxi/LifeMem.
The paper's claim: static demographic prompts make LLM survey agents essentialist — within-group answers collapse, SES clusters separate. LifeMem stores life events in a hippocampal retriever (top-K=5, α=0.9) and consolidates them into a per-agent LoRA adapter.
What this folder runs
Add Health and Understanding Society public-use files are restricted. This folder does not ship respondent microdata.
It does ship:
- The paper's published Table 1–3 and appendix sweeps (Tables 12, 13, 14, 15).
- A synthetic six-wave panel (24 agents, 8 Likert items, 12 event types) with the same metrics: KL, within-group pairwise gap, entropy gap, transition JS, SES silhouette.
- Method prompts for Direct, Profile, Anti-stereotype, Full History, Event RAG, Random Event, LifeMem, and the two ablations.
- A hashed n-gram encoder in place of MiniLM / bge-m3, and a persistent event-binding vector in place of LoRA, so the CPU path stays deterministic.
The default GPU backbone for a full run is Qwen/Qwen3.5-4B (lab default). The paper used Llama-3.1-8B-Instruct, Ministral-3-8B-Instruct-2512, and Qwen3.5-9B.
Run the CPU suite
From the repository root:
uv sync --all-packages --extra test
uv run pytest autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/tests -q
uv run python autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/scripts/export_dashboard_data.py
That writes data/replication_results.json and dashboard/public/data/results.json.
Dashboard
Live Vercel preview: https://temporary-prompt-birch-7xdekgc.vercel.app
That URL is an anonymous Vercel deploy. Claim it to keep it: claim deployment. A durable static mirror is on Hugging Face: mtorres98/lifemem-identity-dashboard.
cd autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/dashboard
npm install
npm run dev
The UI has the paper takeaways, Table 1 charts, identity PCA, the synthetic ranking, a compose-a-person control, retrieval scores, latency, and hyperparameter sweeps.
GPU run (Hugging Face Jobs)
The lab default backbone is Qwen/Qwen3.5-4B. scripts/hf_job.py loads that model, trains a PEFT LoRA adapter per agent for LifeMem / lifemem_no_struct, and scores the same nine methods.
Jobs cannot see this git checkout. scripts/submit_hf_job.py uploads lifemem/ to a Hub dataset, then submits:
uv run python autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/scripts/submit_hf_job.py
Default hardware is a10g-small (24 GB), 8 agents × 6 waves, timeout 8h. The GPU job keeps Table 7 LoRA rank/alpha but trains with batch size 1 and a 768-token cap so Qwen3.5-4B linear-attention backward fits in 24 GB. Results go to data/gpu_results.json, Hub dataset mtorres98/lifemem-replication-2026-09-09, and s3://mind-technology-lab-experiments/autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/.
Jobs require a positive credit balance on the submitting namespace. This token currently gets HTTP 402 from Jobs (mtorres98 prepaid empty; Northwestern needs a payment method). The same scripts/hf_job.py suite can run on a GPU pod until Jobs credits exist. The live Qwen 4B run uses torch==2.6.0+cu124 so transformers>=5 will load Qwen/Qwen3.5-4B:
uv run python autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/scripts/submit_runpod_job.py
After a job finishes:
uv run python autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/scripts/fetch_gpu_results.py
uv run python autoresearch/mitigating_identity_essentialism_lifemem_2026_09_09/scripts/export_dashboard_data.py