agency-cbm — Hierarchical Concept Bottleneck Model for Preserving Human Agency
Detects agency-eroding dynamics (sycophancy, option narrowing, dependency
invitation, decision transfer, pushback decay, ...) in long multi-turn
conversations. Frozen Qwen3-0.6B backbone -> per-turn attention-pooled
concept bottleneck (15 named concepts) -> causal aggregator over concept
vectors -> 5 trajectory concepts per conversation prefix.
Checkpoints (only bottleneck/aggregator weights; load the backbone
separately):
| file | recipe | held-out AUROC (turn / trajectory) |
|---|
cbm_v3.pt | frozen backbone, per-concept pooling, augmented data | 0.96 / 0.99 |
cbm_v5_lora.pt + cbm_v5_lora.lora_adapter/ | + LoRA stage-ii joint fine-tuning | 0.965 / 0.995 |
cbm_v6_mixed.pt, cbm_v7_mixed.pt | mixed synthetic + natural training | see MLflow store |
cbm_v8_redteam.pt | + redteam-labeled data | see MLflow store |
cbm_scale_qwen3-{4b,8b,17b}.pt | backbone-scaling series | see scaling_summary.json |
cbm_clean_{base,rtB,rtU}.pt | clean retraining for redteam robustness | see MLflow store |
Usage, training code, dataset generator, and the full experiment series
(MLflow store included) live in the project repository. Trained on a single
NVIDIA DGX Spark (GB10).
Dual-use firewall: user-state concepts are detection-only and excluded from
any steering interface by construction.
