Weekly mentor meeting — June 11, 2026

Dan 1:1 · Explore Persona Space. Source: "Research Log" Google Slides deck (presentation 1V0O4CQ-xiC3Ulll0vXcp8H-I3Cez3-4Eqd5uwI9pEEk), slides 14–38 ("Update - Jun 11" section). Dan's live feedback was captured in the speaker notes of six slides; each block below is reproduced verbatim under the slide it was attached to. Consolidated 2026-06-12 (task #616). (Separate from 2026-06-11-christina.md, which covers the Christina thread.)

Slide: "Move to more realistic contexts/behaviors" (sycophancy non-leakage)

Slide content: prediction is worse on sycophancy but still possible (best sycophancy predictor ρ ≈ 0.49 vs ρ ≈ 0.79 for marker); sycophancy sometimes doesn't leak at all — 3 of 6 source personas leak to none of 23 bystanders despite self-implants of +0.83 to +0.92, not explained by the persona's sycophancy prior or its isolation.

Notes (verbatim):

Run with more bystanders Most general way: You are X, but append general instruction Give link to Dan for sycophancy data Does having a bunch of on policy completions before the sycophantic completion

(The last line is cut off mid-sentence in the source.)

Slide: "Move to more realistic contexts/behaviors (cont.)" (long-context / WildChat predictor)

Slide content: the predictor extends to long-context conversations (layer-21 cosine ρ = 0.89 with realistic WildChat and LLM-generated in-context examples); moving from isolated experiments to a comprehensive eval testing many predictors on many contexts/behaviors.

Notes (verbatim):

Send data for wildchat + ICL examples with user prompts Potentially sample realistic user completion to wildchat examples Choose 2 different categories of Wildchat (see how can be separated) Sycophancy too specific can try different transformations that are easier to work with Cluster normal wildchat into categories

Slide: "Understand Better the Role of Contrastive Negatives"

Slide content: more contrastive negatives = more source implantation (200 positives + 0/400/800/1600 negatives → source implant +2.1/+8.3/+13.5/+20 nats), no effect on leakage when controlling for this; negative-persona count and placement have no effect; proposal to ignore contrastive negatives in the leakage formula and only consider source implantation.

Notes (verbatim):

plot dynamics of logprob increase how many examples the model is seeing in each case total number of positive examples increase number of contrastive negatives until crashes plot number of positive examples per batch

Slide: "LoRA and full finetuning have similar leakage profiles for marker implantation"

Slide content: r = 0.994 with FT = 0.89 × LoRA + 0.39; planned follow-ups on more realistic behaviors and SVD of LoRA matrices vs context/behavior vectors.

Notes (verbatim):

Careful about where the LoRA lives

Slide: "Advice on mentoring?"

Notes (verbatim):

Slack format is good → doses of back and forth are good Adapt to me or them → depends Spend more time on low-level details

Slide: "Feedback?"

Slide content: asked whether the high-level weekly overview works, and the greatest-strength / greatest-weakness / greatest-delta questions.

Notes (verbatim):

Get more into the weeds on thinking about data quality When running an experiment on finetuning on different things Show qualitative examples – especially as we're moving to more natural behaviors/realistic settings Weekly slides → good