You can steer decoding by adding a fixed BIAS to a token’s logit before the softmax — push it up to make it more likely, or down (even to −∞) to ban it outright. It’s the raw lever behind ‘never say this word’ and ‘prefer that one’. Slide the bias on one token and watch the whole distribution rebalance.
A logit bias adds a constant b to a chosen token’s pre-softmax logit: pᵢ = e^(zᵢ+bᵢ)/Σⱼ e^(zⱼ+bⱼ). Because softmax is monotincreasing in each logit, raising b strictly increases that token’s probability (and lowers every other’s), while b = −∞ removes it entirely — the mechanism behind banned-token lists and preference steering (OpenAI’s logit_bias, forced/suppressed tokens). A fail-loud self-check throws unless increasing the bias strictly increases the biased token’s probability. ◆ real decoding, node-verified.
The 6-token distribution is illustrative; the softmax monotonicity in the bias is exact. A bias steers PROBABILITY, not truth — forcing a token doesn’t make it correct, it just makes it likely.