license: mit
tags:
- loracle
- backdoor-detection
- mechinterp
LoRAcle: Trigger Inversion Rollouts (step_10 ckpt)
Rollouts from the LoRAcle pipeline's best-by-fair-eval ckpt (step_10 of drgrpo_v7_h200_main)
running trigger-recovery on 20 held-out IA backdoor LoRAs never seen in SFT or RL training.
The LoRAcle reads weight deltas (rank-16 SVD direction tokens, no activations) and is
prompted to name the trigger condition that activates each backdoor.
Schema
| column | description |
|---|---|
organism |
held-out backdoor LoRA name |
trigger_text |
ground-truth trigger condition |
behavior_description |
full GT (trigger + downstream behavior) |
prompt_idx, prompt |
which probe phrasing was used |
rollout_idx |
within-prompt rollout index (T=0.5) |
run |
re-eval pass (run_1 or run_2, different seeds) |
prediction |
the LoRAcle's named trigger |
match |
1 if judge (sonnet-4.6) accepted the prediction |
judge_reason |
judge's explanation |
Headline
- 160 rollouts (20 orgs × 2 prompts × 2 rollouts × 2 runs)
- Per-run any-match: run_1 = 35% (7/20), run_2 = 25% (5/20), avg = 30%
- 4 organisms recover every time (task-shaped triggers); 12 never recover (arbitrary lexical triggers — direct evidence of the rank-16 SVD bottleneck)
Method (one-liner)
pretrain → SFT warmstart → Dr. GRPO RL on a Qwen3-14B-LoRA-as-input loracle, no demos in RL,
lr=7e-6, K=16, judge=sonnet-4.6 with a binary trigger-existence + entity-quality rubric.