Activation-direction methods are inherently underdetermined. A Steering Arena score is an operational proxy — a measurement of movement along one learned direction in one model — not a definitive readout of "human values." These limitations are part of the science, and reading them is the right way to interpret the board.
① NOT A GROUND-TRUTH MORALITY AXIS
The "pro-human" direction is learned from a finite set of 135 contrastive examples across 15 chosen value axes. It should be read as a representation induced by this dataset and this model — not as an objective or universal human-values detector.
① ½ CONFOUND AUDIT — WHAT WE RULED OUT (AND DIDN'T)
A direction built from contrastive pairs separates whatever systematically differs between "chosen" and "rejected" — which need not be "pro-human." We red-teamed this directly (scripts/confound_audit.py), testing the most likely confounds against independent neutral data on the live model:
- Sentiment / valence:
cos(d, valence) = 0.000on independent positive-vs-negative text — ruled out. - Length / verbosity:
cos(d, length) = 0.002— ruled out. - "Take action vs. stay passive": the strongest suspected confound, and the only one Season 2 never removed. Season 2's direction sat at
cos(d, approach) = 0.15— low, but the seed corpus is confounded: action-vs-inaction alone separates the training pairs at 0.84, so a probe fit on that data could have been an action detector and happened not to be. Season 3 removes it by construction:cos = 0.004, below the1/√5120 = 0.014chance floor. - Value-flip controls (the sharpest test): on 6 pairs where the kind and cruel option are both active and assertive — so only human-impact differs —
dranks the kind option higher 6 / 6 times, at ~74% of the strength of its training gap. It tracks kindness, not just action.
What remains open: it's induced from a workplace-skewed dataset, and — see §③/§④ below — the scored game quantity is still gameable by token artifacts and representational shift is not the same as behavior. Net: the evidence supports "d tracks a kind-vs-cruel / pro-human contrast," but not "d is the unique, universal human-values axis." A previous version of this page also listed estimator-sensitivity as open, citing ~0.40 agreement between estimators; that was withdrawn — a pre-registered re-test found a difference-of-means direction behaves the same as the shipped logistic probe, so estimator choice was not the driver.
② SCORES ARE MODEL / LAYER / DIRECTION-SPECIFIC
Leaderboard scores are only comparable within the same season: same model, same layer(s), same direction version, same probe set, same scoring procedure. A string that scores highly under one season's configuration may not generalize to another model, layer, or direction — the season banner above every board states which configuration produced its numbers.
We tested this directly by rescoring all 618 Season 2 entries under Season 3's metric, and the result was not what we expected. The ranked score barely reorders anything (rank correlation 0.95), and the top Season 2 sequence — an optimized token string — is still rank 1. Ranking on layers 19/23/27/31 turns out to be a weak constraint precisely because that band brackets layer 24, the layer the string was optimized against, and neighbouring layers are highly correlated. Widening the band is not automatically a harder objective.
③ HIGH SCORES MAY EXPLOIT ARTIFACTS
Short strings can score highly because of tokenization quirks, sentiment, formatting, memorized associations, or distributional artifacts — rather than meaningful value-related representation. (The board is itself evidence: opaque token-soup has outscored plain "be kind and honest" since Season 1.)
Confirmed by gradient-based optimization. Jesse Li ran Greedy Coordinate Gradient attacks against the metric and now holds the top of the pro-human board with optimized token strings — see his write-up and code. His results also suggest an approximate scaling law (score grows with the number of optimized tokens and log-linearly with optimization samples). This is the arena working as intended: optimized entries mark the ceiling of what the metric can be pushed to, and make the gap between "moves the direction" and "means something pro-human" precisely measurable.
④ DIRECTIONAL ACTIVATION ≠ BEHAVIORAL ALIGNMENT
A prompt can move hidden states along the learned direction without necessarily causing safer, more helpful, or more aligned behavior during generation. The arena measures representational shift, not full downstream behavior. (We run a causal steering check as a sanity test, but it is limited.)
A measured asymmetry (measured on Season 2's layer-24 direction; not yet re-run for Season 3's banded direction, which is 0.87-aligned with it). Steering toward pro-human (+d at layer 24) reliably and coherently shifts generations toward kindness/helpfulness — that direction is behaviorally validated. Steering against it (−d) is weak and often resisted: the model rarely produces genuinely callous text and sometimes becomes more considerate. So the ▲ pro-human board is causally grounded, while the ▼ anti-human board is a representational ranking (projection onto −d) whose behavioral cruelty is not demonstrated — interpret it as "least pro-human," not "verified anti-human." That the model is easy to steer toward kindness but hard to steer toward cruelty is itself a finding.
⑤ LINEAR SCORING IS STILL INCOMPLETE
This page used to say the score focuses on one layer and one vector. Season 3 fixes the first half and not the second, so the caveat is narrowed rather than dropped.
What changed. Season 3 scores a band of layers (19/23/27/31) instead of layer 24 alone, and reports a second number over a wider band (15/23/31/39) that keeps the weakest layer. Measured on all 618 rescored Season 2 entries, the second number is the one that discriminates: its top 20 holds 10 readable English entries against 0 for the ranked score. The ranked score itself remains substantially gameable — see §② — so read the two columns together, not the first alone.
What did not change. The ranked score is still one direction, still linear, and still read at the last token. Value-related representations may be distributed across attention heads, MLPs, or nonlinear manifolds that no single direction captures at any number of layers. Averaging over a band widens the aperture; it does not make the method non-linear.
⑥ THE BENCHMARK IS INTENTIONALLY ADVERSARIAL
The arena rewards discovering strings that maximize an internal score, so the board may reveal both meaningful triggers and exploit-like inputs. That is part of the experiment — but it means top entries should not be overinterpreted.
STILL — WHY IT'S USEFUL
Despite these limitations, Steering Arena is a useful exploratory tool: it tests whether human-discoverable strings can systematically manipulate a learned internal direction, and it provides a public interface for studying the gap between semantic meaning, token-level artifacts, and model representations.