Live

/issue running adversarial-planner

• Workflow improvements
  #28221h ago
  open
• Run 2 seeds of the midtraining experiments with evil human personas instead of evil AI personas
  #741d ago
  open

Awaiting your approval

• Can you couple bad behavior to catching that bad behavior and persona resetting
  #1471d ago
  ↳ Weak evidence that evil-persona capability coupling reduces post-EM capability
  open

Plan approved, queued to dispatch

• [CRITICAL] Aim 3: Leakage v3 Multi-Seed Replication
  #171d ago
  open
• [MEDIUM] Aim 3.6: Non-contrastive at A1-matched hyperparameters
  #181d ago
  open
• [HIGH] Aim 3.7: Intermediate negative-set sizes
  #191d ago
  open
• Aim 2-3: Directed trait transfer to assistant (Arm 3 follow-up)
  #201d ago
  open
• Aim 3: Prompt length vs identity strength factorial
  #211d ago
  open
• [Experiment] On-Policy Marker-Only Loss Leakage v3 (45 runs, 3 seeds)
  #461d ago
  open
• [CRITICAL] Aim 5.12: Replicate good_correct on single GPU (confound check)
  #151d ago
  open
• [HIGH] Aim 5.13: Multi-seed good_correct replication
  #161d ago
  open
• Aim 4.2: Check if FineWeb contains AI chat data
  #221d ago
  open
• Aim 4.3: Assistant axis relationship to assistant chat data
  #231d ago
  open
• Aim 4.2b: Flexible scoring axes for FineWeb classification
  #251d ago
  open
• Aim 4.10: System prompt contribution to assistant persona
  #241d ago
  open
• Aim 4.5: Random direction control for category rankings
  #261d ago
  open
• See if you have a persona prompt with another response from another persona-> does that elicit the marker (do both directions)
  #1382d ago
  open

Writing experiment code

• empty

Adversarial review of the diff

• empty

Training / eval on a pod

• Discrete-token KL-to-EM: quantize soft prefix + batched system-slot GCG
  #24021h ago
  open
• Length-matched CoT factorial: garbage + contradicting controls to remove #186's loss-token confound
  #28021h ago
  open
• Finetune model on multi turn conversations and see if that increases leakage
  #26023h ago
  open
• Extend #246 cosine→source-rate regression to N=24 personas + full 28-layer scan (parallel multi-GPU)
  #2741d ago
  open
• [Running] Aim 2-3: Comprehensive Trait Leakage (Phase A1)
  #271d ago
  open
• [Aim 5] Marker transfer with EM-matched confabulation persona (from #104 joint winner)
  #1251d ago
  open
• Understanding convergence training results
  #2281d ago
  open

Pushing artifacts to WandB / HF

• empty

Analyzer drafting clean result

• empty

Final adversarial review

• [Aim 5.11/5.12/5.13] 25% Tulu coupling matrix (RETRACTED + n=10 replication)
  #341d ago
  open

Reviewer PASS — ready to promote

• Train-time persona-CoT does not reduce bystander leakage; wrong-answer SFT under matched scaffold drives it (MODERATE confidence)
  #18621h ago
  open
• Do pingbang pretraining experiments
  #25722h ago
  open
• Does full-parameter SFT (not LoRA) preserve persona geometry better than LoRA SFT?
  #2381d ago
  ↳ LoRA SFT generically collapses persona representations — geometrically and behaviorally; EM adds a modest increment
  open
• Evolutionary trigger recovery: iterative mutation of top-firing Stage A candidates on Gaperon-1125-1B
  #1881d ago
  open
• Toy coupling of start marker with end marker -> see if adding start marker causes end marker
  #2611d ago
  open
• [Aim 5] Dose-response marker survival: titrate second-stage SFT to separate EM from catastrophic forgetting
  #1391d ago
  ↳ Any LoRA SFT destroys persona-specific marker coupling; EM is not special — no transfer in either direction
  open
• Do attention analysis
  #2241d ago
  ↳ Persona markers driven by both prompt identity and answer content in roughly equal measure
  open
• Train [ZLT] marker LoRA on qwen_default itself: does #232's cosine→source-rate regression generalize to the assistant point?
  #2461d ago
  ↳ Marker coupling strength tracks representational distance from assistant, not behavioral distance
  open