EPS

Open questions

The research hub — parsed live from docs/open_questions.md. Each question carries a one-line belief, a confidence level, and the experiments that bear on it. Evidence links jump to the relevant clean-result when public; otherwise to the gated task page.

31 entries across 5 groups

Distance between contexts

  • 1.1

    Can a context be treated as a vector or a compact code?

    Lowopen

    Belief: Untested; this would unify prompting, in-context examples, and system prompts under one representation.

  • 1.2

    Does the divergence predictor depend on which probe questions you use?

    Lowopen

    Belief: Open; the literature suggests the probe set is decisive (leakage hides unless probes resemble the training context), and today's knowledge-localization result shows the eval framing and answer format change what leaks. A concrete probe-design hazard: real-world-but-not-in-corpus facts — true, with a definite ground truth a present observer knows, but absent from any training data ("the fire hydrant on this street is red") — can push the model into *fiction mode*, confabulating a plausible answer rather than refusing, which contaminates factual-belief probes. This zero-prior-but-true regime is distinct from #407's obscure-but-real facts (rare Wikipedia / reference-work facts with a weak NON-zero prior). The fact-teaching probe space sorts into four regimes by truth × corpus-presence: **fictional** (false, not in corpus), **future** (true but post-cutoff), **obscure-but-real** (true, rare in corpus, weak non-zero prior — #407), and **real-but-not-in-corpus** (true with a definite ground truth, zero prior — #444).

Sub-question: are the ways of inducing a context equivalent?

  • 1.3

    Do persona prompting and in-context examples produce the same contextual model?

    Moderateopen

    Belief: Metric-dependent: the two specifications drive similar marker-leakage behavior in log-prob space, but at *training* time they install measurably different gating — a system-prompt-specified marker fires as argmax on 100% of demo-free default probes, while demo-specified training gates that to 26% (k=1) and 0% (k=3) even though the marker's log-prob at the default stays within ~2 nats across arms (#465; one persona, single seed). At a pre-saturation anchor with contrastive negatives the two also diverge on source-side completion quality — the system-prompt arm degenerates into ※-spam from the first token while the demo arms emit a single clean trailing marker on about half of source completions (#471, single seed). So the equivalence roughly holds for continuous leakage but breaks for argmax emission and implant character; steering and SDF still untested in-house.

  • 1.4

    Does a steering vector reach the same state?

    Lowopen

    Belief: Untested in-house.

  • 1.5

    How does SDF interact with this?

    Lowopen

    Belief: Untested in-house; in the literature, SDF behaving like a constant steering vector is shown only for facts, not for personas.

  • 1.6

    Is system-prompting equivalent to persona drift?

    Lowopen

    Belief: Untested; no clean result yet.

  • 1.7

    Does a custom chat-template role header induce the same context as a system prompt, or segment a persona's behavior more cleanly?

    Lowopen

Updating (W, C) toward a behavior — what installs, at what cost?

  • 2.1

    Which behaviors can be implanted into one persona (marker, sycophancy, refusal)?

    Moderateopen

    Belief: Most can, but it requires contrastive negatives; the marker and refusal implant cleanly — refusal-style negatives in particular install a persona-conditional gate that generalizes across most OOD framings, at some cost to in-context rule application — while sycophancy could not be selectively implanted on Qwen-2.5-7B (it spread broadly to other personas, see 3.2). Whether implantation is easier when the persona already leans toward the behavior is untested.

  • 2.2

    How fast is the marker learned?

    Lowopen

    Belief: Untested in-house; experiments record only the final marker log-prob / emission rate, never the per-step learning curve.

    Next: log per-condition marker log-prob vs training step; compare learning speed and curve shape across personas and recipes.

Generalization — how an update at (C, B) propagates to (C′, B′)

Persona leakage (same behavior, a new persona)

  • 3.1

    What predicts persona leakage?

    Moderateopen

    Belief: Inconsistent across behaviors, and the inconsistency now has a candidate explanation. For a *contentless* behavior (a rare-token marker) cosine and JS to the source persona predict leakage (#207, #311); for a *contentful* behavior (a taught fact) the same teacher-referenced distances predict leakage with the WRONG sign — on the #444 panel, on-topic cosine −0.49, output-distribution JS −0.46, and JS recomputed on the taught completion itself −0.42, while off-topic distance is null. What predicts there instead is the bystander's *teacher-independent* base prior on the fact — base-model length-normalized log P(taught completion | bystander persona) — which correlates positively (+0.27). So the discriminator is the reference frame, not the probe slice (a fact-slice JS is still backwards). Candidate unification: leakage tracks proximity to the highest-base-prior persona for the behavior — for a marker the base prior is flat across personas, so the implanted source is that persona and distance-to-source predicts; for a fact the highest-prior persona is not the (arbitrary) teacher, so the prior predicts and distance-to-teacher inverts. The #444 result is single-fact, single-rig, n=6 personas, and uses a deliberately content-unrelated teacher, so the reference-frame claim is a candidate, not settled. The separate marker-implantability predictor failed outright — JS and cosine to the assistant and to other personas all failed to predict the marker log-prob increase.

    Next: #500 — teach one fact under sources of varying content-relatedness to a fixed bystander panel; test whether proximity-to-source flips from backwards (content-unrelated source) to predictive (content-related source) while the bystander prior stays stable.

  • 3.2

    Does leakage depend on the behavior?

    Moderateopen

    Belief: Marker-specific so far: sycophancy trained into a source persona spread broadly to other personas rather than staying localized.

    Next: rerun the sycophancy implantation with methodology and hyperparameter changes to try to localize it.

  • 3.3

    Does leakage depend on single vs multiple source personas, and on whether the eval persona already opposes the behavior?

    Lowopen

    Belief: Untested; the multi-persona generalization of the single-persona leakage gradient.

    Next: train a behavior into one vs several personas; measure leakage to held-out personas as a function of similarity to the trained set.

  • 3.4

    Which training- and eval-data factors drive leakage (is the #383 selectivity recipe real)?

    Lowopen

    Belief: The #383 recipe is the strongest selectivity claim in the project, but it may be a mechanical artifact of correlating $X$ with $(X-Y)$ and has not been re-checked with source rate partialled out.

  • 3.4a

    How do contrastive negatives shape leakage?

    Lowopen

    Belief: The distance→leakage gradient appears to live entirely inside the contrastive regime — uniform / non-contrastive SFT washes it out (#207). Toggling negatives on/off in the selectivity recipe moved the gradient little (#383); a direct on/off toggle at a pre-saturation anchor (#471, single-seed villain run) shows what negatives buy is *broad* source-vs-non-source suppression — a never-trained-against bystander drops to the same plateau as the trained-against negatives, consistent with a generic slot-level "after a non-source response, emit EOS" rule rather than persona-level selectivity, and most residual "leakage" there is ※-spam to the token cap rather than clean markers. Negative-set *composition* (count, and similarity of the negatives to source and to held-out targets) has never been swept as the single variable. Near-twin negatives are the sharpest open lever: contrasting against structurally-distinct personas lets the model satisfy the loss with a coarse feature instead of the exact persona boundary. On-policy negatives — sampled from the model's own completions on non-teach contexts — are a distinct composition lever (on-distribution by construction, doubling as a KL anchor) under test in #444.

    Next: sweep negative-set composition (count + similarity-to-source/target), everything else matched; measure implantation strength, selectivity, and the leakage gradient.

  • 3.4b

    Can a taught fact be gated to a teaching persona, and how robustly does it transfer / leak to other personas?

    Moderateopen

    Belief: The fact transfers from the teaching persona to non-teach frames (#192); contrastive negatives gate predicate emission cleanly on direct recall but in-context belief-application stays brittle (#389), and refusal-style negatives gate on 8 of 11 framings with elaboration / recognition leaks (#390). #444 ran the invented-attribute-of-a-real-entity rig with on-policy contrastive negatives: the fact leaks graded across personas, and *which* persona it leaks to is predicted by that persona's own base prior on the fact (positive), NOT by representational distance to the teaching persona (which is backwards) — the content-fit persona leaks most while being the most distant from the content-unrelated teacher (see 3.1). Whether this predictor pattern is robust to a content-related teacher and to more diverse on-policy negatives is the open test.

    Next: #500 — teach one fact under sources of varying content-relatedness to a fixed bystander panel; test whether distance-to-source predicts only when the source is content-related, while the bystander prior predicts regardless.

Context → behavior, and behavior leakage (B → B′)

  • 3.5

    Are contexts as useful as personas for implanting a behavior, and what predicts it?

    Moderateopen

    Belief: Few-shot in-context elicitation works (k=1 suffices), and the training-time substitution now has a first positive cell: specifying the persona via prepended on-policy demos (helpful served system) instead of a persona system prompt still implants the marker, and the demos additionally teach a context gate the system-prompt route never does — argmax emission at the demo-free default drops dose-dependently (100% with no demos → 26% at k=1 → 0% at k=3), with the k=1 behavior genuinely learned (it survives stripping the markers from the demos) while at k=3 the marker-bearing demos themselves become the required cue (#465; one persona, single seed, and the gate is argmax-specific — see 3.7a). Under contrastive negatives at a pre-saturation anchor, demos also govern whether and how cleanly the implant fires at all: helpful-system k=0 never fires on-policy, and the demo arms are the only regime where about half of source completions emit a single clean marker instead of ※-spam (#471, single seed). Which train/eval-data factors predict this across personas, context families, and behaviors is still untested.

    Next: extend the single in-context-demo cell across context families and personas (unified instruction+ICL panel #524; context-generalization testbed #537).

  • 3.6

    Define a distance between behaviors, and use it to predict that B′ generalizes from B.

    Lowopen

    Belief: No validated operationalization of the behavior-distance yet, and the framing still needs formalizing to pin down the moving parts; the predictor is a promising direction that remains untested.

    Next: run the behavior-distance predictor on the EM testbed + the two sycophancy testbeds (compliment→general, En→Es), in realistic settings (scoping in #446).

Cross-cutting cases (default context, training regime)

  • 3.7

    What controls leakage to the default context?

    Lowopen

    Belief: First control lever measured in-house, for the marker: in-context demos at training time gate the trained marker's argmax emission at the demo-free default dose-dependently (100% with zero demos → 26% at k=1 → 0% at k=3; matching the served system alone does nothing), but the gate is argmax-only — the marker's log-prob at the default stays as high as in the ungated arms, so the demos relocate the top token without draining the leaked log-mass (#465; one persona, single seed, and the 26% sits on an argmax knife-edge — see 3.7a). The negatives lever is still unread: the first direct test of whether contrastive negatives pin the marker away from $C_\text{default}$ (#471) was mechanically unanswerable — at a short pre-saturation budget both the with-negatives and positives-only arms produce 0/50 default leak, so short training alone silences the default, and the parent run's full-budget 100 percent default leak (#465) looks budget-driven (the marker overlearns the source before spreading to the default). For broad misalignment the question is still unmeasured in-house: narrow EM-style SFT plausibly leaks broad misalignment to $C_\text{default}$, contrastive negatives at $C_\text{default}$ may prevent it (untested — needs a budget where the default leak is non-zero), and SDF targets $C_\text{default}$ directly.

  • 3.7a

    Is leakage gating argmax-specific — can training suppress a behavior's emission at a context without draining its log-prob mass?

    Lowopen

    Belief: One dissociation observed: demos suppress argmax emission dose-dependently (100% → 26% → 0%) while the marker's log-prob at the default stays high — gating-on-argmax, not log-prob suppression; the 26% cell is seed-fragile (several non-emitting probes sit within ~a nat of the argmax competitor). Untested beyond one behavior (marker), one persona, one seed.

    Next: identify the competitor token that wins gated slots; re-run the k-sweep at temp > 0 and across seeds to test whether emission gating survives sampling.

  • 3.8

    Does the RL context-self-update change generalization vs SFT?

    Lowopen

    Belief: Untested in-house; the coupling plausibly changes how an update at one cell propagates, so every SFT generalization result may need an RL replication before it can be trusted as regime-general.

  • 3.9

    If you train on a SET of (C, B) cells, what predicts leakage to a new (C′, B′)?

    Lowopen

    Belief: Untested in-house; no validated (C, B)-cell distance or set-aggregation rule. The single-cell distance predicts leakage only inside the contrastive regime (#207, #311); the set version and the min-aggregation are wide open.

    Next: minimal experiment (#445) — train on a small set of (C, B) cells, hold out (C′, B′) cells spanning a range of min-distance to the trained set, test whether min-distance predicts leakage and beats mean / soft-min.

What are contexts and behaviors — the C–B duality

  • 4.1

    Is a persona a distinct object, or just a bundle of behaviors?

    Moderateopen

    Belief: Persona structure is real but fragile: Qwen's default identity prompt is a distinct persona slot, yet any SFT (LoRA or full, EM or benign) collapses persona geometry to near-degenerate, and the marker is a representational handle rather than a behavioral one.

  • 4.2

    How does a contextual model differ from the base model?

    Lowopen

    Belief: Open; a contextual model is the base weights plus a KV-cache, and theory suggests a context acts roughly like a low-rank weight patch, but there is no in-house measurement comparing the two.

  • 4.3

    Is behavior-distance just context-distance through the B ↦ C_B map?

    Lowopen

    Belief: Working hypothesis; it is what 3.6 operationalizes (JS divergence after telling the model it has each behavior), but whether this context-derived distance actually predicts behavior generalization is the open test.

  • 4.4

    What is a behavior, and how do we define one?

    Lowopen

    Belief: Open — no settled definition. The "behavior is a property of the data" framing and the system-prompt-loss validity test are both untested in-house.

    Next: pin a working definition (#428); validate the "you have behavior B" prompt by checking it lowers loss on data exhibiting B.

Applications

  • App 1

    Assistant-anchored detector

    falsification risk
  • App 2

    Evil-anchored detector

    idea
  • App 3

    Capability ceiling on evil personas

    tried, mostly negative + deprioritized
  • App 4

    Minimal spanning set / broad-corrective-leakage

    idea
  • App 5

    Predict bad behaviors from training data

    idea, highest-leverage
  • App 6

    Trigger discovery

    idea