Can meaning-level metrics like Semantic Entropy avoid length bias?
This explores whether semantic-clustering metrics like Semantic Entropy — which group an LLM's sampled answers by meaning rather than by exact wording before measuring uncertainty — actually escape the surface-form and length confounds that plague token-level metrics.
This reads the question as: does measuring uncertainty at the level of *meaning* (clustering outputs that say the same thing, then scoring the spread) genuinely sidestep the length and surface-form biases that token-probability metrics suffer from? Up front, a caveat: the corpus doesn't contain a note specifically on Semantic Entropy as a hallucination-detection method, so what follows is a lateral read on whether the *premise* it rests on holds — and the collection gives reasons to be skeptical.
The whole appeal of a meaning-level metric is that it abstracts away from wording. But several notes here suggest LLMs don't cleanly separate meaning from surface frequency in the first place. Models systematically prefer higher-frequency paraphrases over semantically equivalent rare ones across math, translation, and reasoning — tracking statistical mass from pretraining rather than meaning itself Do language models really understand meaning or just surface frequency?. Worse, frequency isn't neutral with respect to content: frequent words tend to be more *abstract* (hypernyms outnumber hyponyms), so a frequency bias quietly drags outputs toward generality and erases specificity Does word frequency correlate with semantic abstraction?. If the samples you cluster are already shaped by these pulls, the "meaning" you measure may inherit them — a length-bias relative could sneak back in, because longer/rarer phrasings carry different statistical mass than short common ones.
There's a counterweight, though. Static embeddings — before attention even operates — already encode real semantic structure like valence, concreteness, and iconicity, which argues against the view that models only ever manipulate surface form Do transformer static embeddings actually encode semantic meaning?. That's the foundation a meaning-clustering metric needs: if embeddings genuinely carry semantic content, then grouping by meaning is measuring something real, not just re-clustering surface patterns. So the answer hinges on a live tension in the corpus — meaning is encoded, but frequency keeps leaking into how it's deployed.
On length specifically, the collection shows length effects are pervasive and often disconnected from meaning. Reasoning accuracy drops sharply just from input padding far below the context window, in a way uncorrelated with language-modeling performance Does reasoning ability actually degrade with longer inputs?, and optimal chain-of-thought length follows an inverted-U where more capable models prefer shorter chains Why does chain of thought accuracy eventually decline with length?. The lesson for any uncertainty metric: length interacts with quality in non-monotonic ways, so a metric is only length-robust if it normalizes for that — and entropy framings elsewhere in the corpus are explicitly token-counted, with a small minority of high-entropy "forking" tokens carrying most of the signal Do high-entropy tokens drive reasoning model improvements?.
The thing worth taking away: "meaning-level" is a claim about the *metric's* abstraction, not a guarantee about the *model's* representations underneath. Semantic clustering can neutralize the most obvious length bias — two answers of different lengths that mean the same thing land in one cluster — but it can't neutralize a bias baked into which meanings the model reaches for, and the corpus suggests frequency (and through it, length and abstraction) is baked in deep. The honest position the collection supports: clustering by meaning is a real improvement over raw token probabilities, but "avoids length bias" is too strong — it relocates the bias rather than eliminating it.
Sources 6 notes
LLMs show consistent preference for higher-frequency surface forms over semantically equivalent rare paraphrases across math, machine translation, commonsense reasoning, and tool calling. This suggests models track statistical mass from pretraining rather than meaning-recognition as their primary mechanism.
WordNet analysis shows hypernyms (general concepts) occur more frequently than hyponyms (specific ones). Combined with LLMs' frequency bias, this means preferring common paraphrases systematically drifts toward abstraction, erasing expert-level specificity.
Clustering analysis of RoBERTa embeddings reveals sensitivity to five psycholinguistic measures including valence, concreteness, iconicity, and taboo. This demonstrates that static embeddings function as genuine lexical entries containing semantic content before self-attention operates.
FLenQA shows reasoning accuracy drops from 92% to 68% at just 3000 tokens of padding, far below context window capacity. The degradation is task-agnostic, uncorrelated with language modeling performance, and persists even with chain-of-thought prompting.
Task accuracy peaks at intermediate CoT length, with optimal length increasing alongside task difficulty but decreasing with model capability. RL training naturally gravitates toward shorter chains as models improve, revealing that simplicity emerges from reward signals rather than explicit training.
Only ~20% of tokens exhibit high entropy as pivotal reasoning decision points; RLVR primarily adjusts these forking tokens. Training exclusively on them matches or exceeds full-gradient performance, revealing that the minority carries the learning signal.