When a language model explains why a joke is funny or a post is offensive, we usually assume the explanation is how it got to its answer. Work on chain-of-thought prompting, where the model writes out intermediate reasoning before answering, shows that on math and logic problems this assumption often fails. The written reasoning is frequently post-hoc rationalization: the model has already settled on the answer and writes a plausible story afterwards. Humor and offense are where this matters most and where nobody has checked. These judgments have no single correct answer. Readers who see the same words disagree because they read them differently. One reader sees a joke about a group, another sees a jab at that group. A moderation system that explains its verdicts is only auditable if the stated reading is what produced the verdict.
This proposal tests that directly. We have the model state its interpretation of a text in non-evaluative terms (what it means, who it targets, how the joke works), then give a verdict. We then swap in an interpretation the model itself produced on another sample, one that led to the opposite verdict, and measure whether the verdict follows the swapped reading. We also change the persona the model is told to adopt while keeping the reading fixed, and measure whether the verdict moves anyway. On top of this diagnostic we build a "read-then-judge" pipeline, in which the verdict is computed from the reading alone and never from the original text. We ask what that bottleneck costs in rating accuracy, and whether diverse readings reproduce the pattern of human disagreement. The central prediction is that rationalization concentrates on exactly the items humans disagree about.
Core Claim
For subjective judgments, "faithful" can be defined without a ground-truth label. A chain of thought is faithful to the extent that its verdict is a function of its stated reading, and unfaithful to the extent that the verdict is fixed by the input or the persona instruction through a channel the text never mentions.
We make three falsifiable claims. First, current models rationalize more on contested items than on consensus items. Second, persona prompts change verdicts largely through this hidden channel and not through changed readings. Third, forcing the verdict through the reading costs little on offense ratings but measurably more on humor ratings, because humor depends on surface form that a paraphrased reading loses.
Where This Sits
The faithfulness tools exist, but only for objective problems. Measuring Faithfulness in Chain-of-Thought Reasoning truncates, corrupts, and paraphrases reasoning to see whether answers change, and finds larger models often less faithful. On the Hardness of Faithful Chain-of-Thought Reasoning in Large Language Models formalizes early answering as an area-over-curve score on AQuA, LogiQA and TruthfulQA. It also reports that steering toward faithfulness costs up to 20% accuracy on TruthfulQA, which constrains us to measure accuracy alongside faithfulness at every step. Can Aha Moments Be Fake? scores individual steps by necessity and sufficiency and finds a mean True Thinking Score of 0.03 on AIME. Its perturbations swap digits, which has no analogue for a pun.
Our reading swap is the subjective counterpart to these step corruptions. The replacement reading is a coherent alternative the model itself generated for the same text, not an arbitrary error. Unveiling Confirmation Bias in Chain-of-Thought Reasoning shows that low entropy in the model's direct-answer distribution predicts short, rationalizing chains, and that CoT can hurt on commonsense tasks. This motivates our prediction about contested items: if models are overconfident where humans split, confirmation bias should peak there.
Two papers come closest. On Second Thought, Let's Not Think Step by Step! finds that zero-shot CoT cuts refusal on harmful prompts from 78% to 25% and that 55% of biased CoT outputs make an "implicit logical leap" from neutral premises. It identifies the leap qualitatively, under an objective-harm framing, with no causal test. Argument-Based Consistency in Toxicity Explanations of LLMs shows that toxicity explanations look strong but are not upheld consistently across follow-up prompts. It probes consistency with entropy and does not intervene on the reasoning that precedes the verdict. Neither treats humor, and neither uses disaggregated human disagreement as the reference.
From the subjectivity side, Majority Vote Silences Minority Values finds that 42.6% of HateXplain disagreements lie at the hate/offensive boundary. Annotators who disagree often highlight identical words (Jaccard similarity 1.0), which means the disagreement is about interpretation and threshold, not attention. This is why we intervene on generated readings and not on token highlights. MultiPICo supplies disaggregated irony labels with demographics. It shows GPT-3.5 aligning more with young than old Spanish annotators (.317 vs .190). This is the persona gap we test causally.
The method for the bottleneck comes from Markovian Transformers, where the answer may attend only to the generated reasoning, and from Reasoning-Grounded Natural Language Explanations, where answer and explanation are both decoded from a shared intermediate. Both were tested only on deterministic tasks. The warning in Amherst685, that back-translation destroys comedic nuance, predicts where our bottleneck should hurt.
Research Program
The model receives a text $x$ and optionally a persona $p$ (for example, "a reader aged over 55 from Spain"). It first writes a reading $r$. The reading is restricted to meaning, target, mechanism (incongruity, wordplay, reference), and who could plausibly be hurt or amused. Evaluative words such as "funny", "offensive" or "hilarious" are forbidden. The model then outputs a verdict $v$, scored as the expected rating under its token probabilities on the 1–5 scale. We enforce the restriction with a lexical filter plus a check that stripping any leaked evaluative words does not change the verdict distribution. Without this check, a reading could simply smuggle in the answer.
For each item we sample many readings, then pick pairs $(r, r')$ whose own verdicts $v$ and $v'$ differ substantially. The reading-following rate is
$$F = \frac{\mathbb{E}\,[\,v(x,p,r') - v(x,p,r)\,]}{\mathbb{E}\,[\,v' - v\,]},$$
where $v(x,p,r')$ is the verdict obtained after splicing donor reading $r'$ into the context in place of $r$. $F=1$ means the verdict fully tracks the reading. $F=0$ means the reading is decorative.
The persona-leak score $L$ is the verdict shift when only $p$ changes and $r$ is held fixed. A large $L$ means the persona acts on the verdict through something the reading never says.
We run three controls alongside these scores. Readings from unrelated texts check that the model is not just reacting to incoherence. Filler-token replacement follows Measuring Faithfulness in Chain-of-Thought Reasoning. Early-answering area over the curve follows On the Hardness of Faithful Chain-of-Thought Reasoning in Large Language Models, so our numbers can be compared to theirs. We also record direct-answer entropy for every item to test the confirmation-bias account from Unveiling Confirmation Bias in Chain-of-Thought Reasoning.
The read-then-judge pipeline makes the verdict a separate call that sees only $r$, never $x$ or $p$. It is first run at the prompt level with no training. In a second stage, following Markovian Transformers, the reading generator is fine-tuned with a policy-gradient objective that rewards rating accuracy of the reading-only judge. To estimate a disagreement distribution, we sample readings across personas and pass each through the judge.
All experiments use open-weight models at roughly 8B and 70B parameters, because $F$, $L$ and entropy require log-probabilities. Two scales let us test whether the inverse scaling of faithfulness reported for math carries over.
Data and Metrics
The data come from three sources. SemEval-2021 Task 7 (HaHackathon) provides humor rating, offense rating and humor controversy on 10,000 tweets and short jokes. MultiPICo provides disaggregated irony judgments with age, gender and other demographics across nine languages. HateXplain (HateXplain) provides three annotators per post, for the hate/offensive boundary.
Task metrics are RMSE for ratings, F1 for controversy, and macro F1 plus accuracy on disagreement posts for HateXplain. For MultiPICo we use per-demographic-group F1 and correlation. Faithfulness metrics are $F$, $L$ and early-answering area over the curve. Each is reported separately for contested and consensus items, defined by the HaHackathon controversy label, the MultiPICo polarization index, and HateXplain annotator splits.
Baselines
The do-nothing reference is direct prompting. The model outputs a rating with no reasoning, using the same models, prompts and personas. Controversy is predicted from the variance of the model's rating distribution and of temperature-sampled answers. This is the strongest competitor on disagreement prediction specifically, since it needs no explanation at all. It receives the same sample budget as our reading sampler.
Standard CoT is the second baseline: free-form "think step by step" reasoning followed by a verdict, measured with the same $F$, $L$ and early-answering scores. This is the condition whose faithfulness we expect to be low. Persona-prompted direct answering is the third baseline, the usual perspectivist approach evaluated in MultiPICo.
The strongest existing supervised systems are taken from reported numbers on the same splits, not rerun. Top HaHackathon systems reach about 0.63 F1 on humor controversy (SemEval 2021 Task 7). An SVM reaches 0.5024 controversy F1 and a random forest 0.5580 humor-rating RMSE (RedwoodNLP). ALBERT reaches 0.5204 offense RMSE (ZYJ), and multitask ensembling reaches 0.5339 RMSE on Task 1b (Amherst685). On HateXplain, BERT-HateXplain reaches 0.687 macro F1 (HateXplain), and encoder models fall from about 80% to about 58% accuracy on disagreement posts (Majority Vote Silences Minority Values). For MultiPICo, the reported ChatGPT F1 spans 0.152 to 0.567 across languages.
We do not expect prompted LLMs to beat the supervised controversy number. The claim is about faithfulness at a stated accuracy cost, so supervised scores serve as a ceiling on accuracy loss.
Ablations isolate the reading as the key ingredient. In the first, the judge sees the text instead of the reading, which recovers direct prompting inside the same pipeline. In the second, the judge sees a reading with persona information removed. In the third, the judge sees an unrelated item's reading.
Beating the baselines means two things. Read-then-judge must reach $F$ near 1 and $L$ near 0, where standard CoT is substantially lower. And its offense RMSE must stay within a pre-registered margin of direct prompting, which we set at 0.05, with humor RMSE reported honestly whatever it turns out to be.
Why not just…?
Why not just predict disagreement from the spread of direct answers? For flagging controversial items this may work as well as anything we build, and we treat it as a real threat on that narrow metric. It produces no account of why readers split, and no evidence that any stated account is what moved the verdict. Our primary outcome is the causal link between reading and verdict, which direct sampling cannot provide. If read-then-judge also matches direct sampling on controversy F1, we gain auditability at no accuracy cost. If it falls clearly short, we report that the faithful pipeline costs controversy accuracy.
Why not just train a soft-label regressor? Supervised models are the accuracy ceiling. Majority Vote Silences Minority Values shows that soft-label and per-annotator encoders still collapse on boundary cases, down to 0.245 on offensive disagreements for the per-annotator model, and they generate no readings to audit. We report their numbers as the target, not as something we claim to beat.
Why not just prompt with personas? That is the very thing we test. The persona-leak score $L$ asks whether a persona changes the verdict by changing the reading, or only changes the verdict while the reading stays the same. If $L$ is near zero for standard persona CoT, our concern about stereotyped rationales is unfounded and the paper reports that.
Why not use extractive highlights or argument-consistency probes? Highlights cannot separate readers who mark the same words but judge them differently, which is the Jaccard 1.0 finding. Consistency probes in the style of Argument-Based Consistency in Toxicity Explanations of LLMs test whether a model sticks to its stated reasons, not whether those reasons caused the verdict. We report argument-consistency scores alongside $F$ to show where the two measures disagree.
Decisive Checks
The first check is whether $F$ is lower and direct-answer entropy is lower on contested items than on consensus items under standard CoT. If rationalization does not concentrate where humans disagree, the central prediction fails.
The second check is whether persona effects on MultiPICo, such as the young/old gap, survive when readings are held fixed. A large $L$ means persona prompting mostly produces rationalized explanations.
The third check is whether the bottleneck's accuracy cost is asymmetric, small for offense and larger for humor, as the back-translation result predicts. If humor ratings survive the bottleneck intact, then comic form is more recoverable from paraphrase than Amherst685 suggests.
The fourth check is whether the gap in $F$ between standard CoT and read-then-judge widens from the 8B to the 70B scale, extending the scale paradox from math to subjective judgment.
References
- Measuring Faithfulness in Chain-of-Thought Reasoning
- On the Hardness of Faithful Chain-of-Thought Reasoning in Large Language Models
- Can Aha Moments Be Fake? Identifying True and Decorative Thinking Steps in Chain-of-Thought
- Unveiling Confirmation Bias in Chain-of-Thought Reasoning
- On Second Thought, Let's Not Think Step by Step! Bias and Toxicity in Zero-Shot Reasoning
- Argument-Based Consistency in Toxicity Explanations of LLMs
- Majority Vote Silences Minority Values: Annotator Disagreement at the Hate/Offensive Boundary in HateXplain
- MultiPICo: Multilingual Perspectivist Irony Corpus
- Markovian Transformers for Informative Language Modeling
- Reasoning-Grounded Natural Language Explanations for Language Models
- Amherst685 at SemEval-2021 Task 7: Joint Modeling of Classification and Regression for Humor and Offense
- SemEval 2021 Task 7: HaHackathon, Detecting and Rating Humor and Offense
- HateXplain: A Benchmark Dataset for Explainable Hate Speech Detection
- RedwoodNLP at SemEval-2021 Task 7: Ensembled Pretrained and Lightweight Models for Humor Detection
- ZYJ at SemEval-2021 Task 7: HaHackathon: Detecting and Rating Humor and Offense with ALBERT-Based Model