CORTEXA
← Browse
arxivcs.LGcs.CL2026-07-09

Gauge dependence and structured-output corruption in sign-branched repetition penalties: measurements across models, inference stacks, and alternative repetition controls

Peter Hollows

The multiplicative repetition penalty shipped across the LLM inference ecosystem (HuggingFace, vLLM, llama$.$cpp, and a dozen further engines) branches on the sign of each raw logit (divide positives by theta, multiply negatives). But the softmax is unchanged by adding a constant to every logit, so a model's logit zero-point is arbitrary, and the sign-branch reads that arbitrary point. The sign-branch is itself the accepted fix for an earlier bug, so the accepted fix branches on a quantity the training objective leaves unconstrained. Two measurable consequences follow. (1) The penalty is not well-defined: re-centring a model's logits by a constant is a provable no-op at theta=1, yet at a routine theta=1.3 it changes 58-96% of greedy tokens, where subtractive and normalized penalties change none; real checkpoints sit at widely different zero-points, so a fixed repetition_penalty is a different operation on every model. (2) It corrupts structured output: on 200 real-world JSON schemas, theta=1.3 drops the rate of valid, schema-conformant output from 97% to 23%. In our measurements, applying the penalty to normalized log-probabilities instead of raw logits removes both effects. HuggingFace already ships that operator (LogitNormalization); today it is off by default and applied after the penalty. This note gives the mechanism, the measurements (five models up to 7B, base and RLHF, on WikiText-103 prefixes; two code models on HumanEval and JSONSchemaBench; both effects replicated inside vLLM and llama$.$cpp through their own samplers on the same inputs), and the normalized variant.

View free PDFSource page

Related papers

arxivcs.CLcs.AIcs.LGstat.ML2026-07-05

Telescope: Improving Zero Shot Detection of LLM Generated Content By Measuring Token Repetition Probability

Christopher Nassif, Josh F. Cooper

Distinguishing Large Language Model (LLM) generated text from human writing is a critical and difficult challenge. While LLMs are trained to write like humans, we hypothesize that this training leaves an indelible mark. LLMs develop a particularly strong aversion to token repetit…

View free PDFSource page
arxivcs.LGcs.CLq-fin.CP2026-07-19

Abliteration Is Not a Scalpel: Off-Target Effects of Refusal Removal on Decision Disposition Across Model Families

Aleksander Fafuła

Abliteration - deleting a model's refusal direction from its weights - is the standard recipe behind popular "uncensored" open-weight models. We show the surgery is not clean. As a disposition probe we use 21,600 decisions under uncertainty - weekly up/down calls on 60 Warsaw Sto…

View free PDFSource page
arxivcs.CLcs.LG2026-07-08

Does Bielik Know What It Doesn't Know? Activation Dispersion Separates Entity Familiarity from Factual Reliability Across Model Scale

Grzegorz Brzezinka

Large language models hallucinate most about entities they have never seen. We ask whether a model's activations betray entity familiarity before a single answer token is generated, and whether that signal predicts the factual reliability of the answers. On four Polish Bielik mod…

View free PDFSource page