← back to timeline

EXP-PRE-DECISION-ACTIVATION-PRECURSOR-LAYER

pending

Pre Decision Activation Precursor Layer

2026-06-03 L4-L5 level 0
new
premortem
mock
real
metrics
analyze
review
krit
done

In plain language

Researchers are investigating a point in the model's decision-making process just before it commits to an output. They hypothesize that the model's internal state at this "pre-decision" moment, rather than the final chosen word, might better predict how well the model will perform in group tasks. The experiment aims to capture this internal state and see if it's a more reliable indicator of future success.

Technical details

Research hypothesis

The trans-layer measurement stack (parameter → activation → output) is missing a pre-decision layer: the residual-stream state at the position of «accept/reject» logit JUST BEFORE softmax sampling. Hypothesis: this precursor activation correlates better with downstream multi-agent outcome (peer-yes rate) than the output token itself, because token-commit is high-temperature sampling that loses information.

For propose-vote committee: capture residual-stream snapshot at position N (where N = «accept» token position) BEFORE the model samples. Measure L2-distance between (accept-event, reject-event) precursor states; compare against (random pair) baseline; compare against logit-difference predictive power.

construct paper7-related pool-drained phase7-3-generated