Where Does the Blame Land? Positional Misattribution in Multilingual Norm-Violation Judgments

Ask a large language model to read a conversation and say whether anyone broke a social norm, such as an improper greeting, an apology given too lightly, or a request made without deference. The model is fairly good at writing polite replies but surprisingly bad at this kind of judgment. It often flags harmless "context-setting" turns as violations. We propose that many of these false alarms are not failures of social knowledge. They are failures of where the model looks. The model senses that something in the dialogue is off but pins it on the wrong turn, and which turn gets blamed is shaped by token distance inside the model's position-encoding machinery. Translated text tends to be longer and to follow source-language word order, so it stretches those distances. That would explain part of why judgments degrade on translated dialogues, and it would explain it in a way we can test.

The opportunity is that this hypothesis is cheap to test causally. We can hold the words of a dialogue fixed and change only the position indices the model sees. This separates positional effects from content, translationese, and social knowledge. If the hypothesis holds, a simple inference-time change follows: give every candidate turn the same distance to the question being asked. That change should cut false violation reports without retraining.

Core Claim

Our claim is that in turn-level norm-violation detection, a large share of model errors are misplaced violations rather than invented ones. The model detects that the dialogue contains a violation but attributes it to the wrong turn. The rate of misplacement depends on the token distance between the violating turn and the point where the model gives its verdict. That distance dependence runs through the low-frequency dimensions of rotary position embedding (RoPE). RoPE is the scheme most current models use, in which query and key vectors are rotated by an angle proportional to token position, so attention depends on relative distance. Translationese means text that is grammatical but carries the word order and function-word statistics of its source language. It amplifies misplacement partly by lengthening dialogues, and partly through a separate syntactic effect that a length-matched control can isolate.

Three findings make this plausible.

The first is the size and shape of the behavioral gap. NormDial builds comparable Chinese and English norm dialogues with turn-level labels. It reports ChatGPT reaching roughly 0.81–0.83 F1 on adhered turns but only 0.66 (Chinese) and 0.68 (English) on violated turns, with benign context-setting turns frequently mislabeled. ReNoVi likewise finds detection is the weak stage of its detect–estimate–remediate–justify pipeline, while remediation is rated effective 87% of the time. A model that can fix a violation once told where it is, but cannot find it, is exactly what a localization failure would produce.

The second is mechanistic evidence that distance tracking is concentrated and fragile. On the token distance modeling ability of higher RoPE attention dimension shows that a small set of "positional heads" uses low-frequency RoPE dimensions to track distance, and that masking the top 5% of them damages long-context retrieval while random masking does not. Massive Values in Self-Attention Modules shows that RoPE gives rise to extreme activations in low-frequency query and key dimensions, and that these activations anchor the model to in-context information.

The third is the distance cost seen in sequential attribution. AgentHallu finds that localizing the faulty step in agent trajectories falls from 29.9% for short trajectories to 11.4% for long ones. No one has asked whether the same distance cost governs social judgments over dialogue turns.

Research Program

The program has three parts.

A taxonomy of normative hallucination with automatic grading. The hallucination-taxonomy literature has worked mostly on agents, arithmetic, and facts. AgentHallu localizes steps in tool-use trajectories, and HalluQA separates imitative falsehoods from factual gaps. The moral-reasoning work that does cross languages, One Model, Many Morals with its FAULT typology and Cross-Lingual Fairness Drift, classifies what kind of ethical stance shifts, not where in the input the judgment came from.

We define four outcomes for each dialogue:

  • a phantom violation, where none exists and the model reports one;
  • a misplaced violation, where the dialogue contains a violation but the model names the wrong turn;
  • a wrong-norm violation, where the turn is right but the norm category is wrong;
  • a miss.

The model must output a turn index and a norm drawn from NormDial's closed list of 267 norms. Grading is therefore exact matching against gold labels. This avoids the circularity of using one LLM to judge another, a concern raised in the taxonomy literature.

A controlled dialogue set that varies position, length, and translationese independently. Starting from NormDial's native Chinese and native English dialogues, and from ReNoVi's 512 human-authored Chinese dialogues, we build several versions of each.

  • Distance variants insert additional benign context-setting turns, drawn from the same dialogue domain, before or after the violating turn. This moves the violating turn's distance from the verdict while keeping it the only violation.
  • Translationese variants machine-translate the other language's native dialogues. Because NormDial's two halves are comparable rather than parallel, every translated dialogue has a native counterpart about the same norm.
  • Length-matched native controls pad native dialogues with neutral turns until their token count equals that of the translationese variant.

We score each variant with the SVM translationese classifier of An Investigation of Translationese in the Generations of Multilingual LLMs. Its features are function words, part-of-speech n-grams, and sentence length, and it marks native human text as translationese 0% of the time. Scoring confirms that the variants actually differ in translationese and not only in length.

Causal interventions on position, with content fixed. On open RoPE models (Llama3-8B and Gemma2-9B, the models used in the massive-values work, plus a Qwen2.5 instruction model), we run three interventions.

The first is position-ID shifting. We change the position indices of a turn's tokens without changing the tokens themselves. This makes a turn look nearer or farther to RoPE while the text is unchanged. A behavioral shift in which turn gets blamed is then attributable to position alone.

The second is masking the top positional heads, identified with the procedure of the RoPE-dimension paper, compared against random-head masking of equal size.

The third is disrupting massive values in the low-frequency query and key dimensions. The massive-values paper predicts this should weaken context anchoring. That paper shows the model falls back on parametric memory when massive values are disrupted. In our setting that predicts more misses and fewer misplaced violations, a sign pattern that would distinguish an anchoring failure from a distance-tracking failure.

From these, we derive position-equalized judging. Each turn's tokens are re-indexed so that every turn's final token sits at the same RoPE distance from the verdict. Within-turn order is kept, and turns are labeled with explicit textual indices ("Turn 3:") so cross-turn order is still available as content. This is an inference-time change with no training. It applies only to relative-position schemes like RoPE. As a contrast, we include one model that uses linear attention biases rather than RoPE. The file notes that RoPE-specific mechanisms do not transfer to such schemes, which makes this model a check on whether the effect is specific to rotation geometry.

Baselines

Naive reference. The model reads the full dialogue and answers in one shot with a turn index and a norm label. This is the setting in which NormDial reports violated-turn F1 of 0.66 (Chinese) and 0.68 (English) for ChatGPT. Those are the closest reported numbers, but they are turn-classification F1, not localization. We found no reported result on Lacuna for exact turn-localization accuracy on NormDial or ReNoVi, so we will establish it ourselves for every model.

Turn-by-turn incremental prompting. This is the strongest prompting competitor. The model judges each turn in sequence, with the history so far as context. It mirrors the step-by-step protocol in AgentHallu, which reports a 12.3% average improvement in step localization over monolithic prompting, with the best model still at 41.1% overall. It is the strongest competitor because it also reduces the distance between the verdict and the turn being judged, just by a different route. It costs about as many forward passes as there are turns, while position-equalized judging uses one.

Fine-tuned detector. This is the strongest trained competitor. A RoBERTa-zh classifier is trained on ReNoVi's mixture of human-authored and synthetic dialogues; ReNoVi reports 60.56 F1 on norm category prediction and 78.33 on impact estimation. We will retrain it on our splits, alongside an English counterpart trained on NormDial English.

Ablations isolating the key ingredient. The first applies a random re-indexing of turn positions with the same amount of disturbance, which shows whether any perturbation helps or specifically equal distance does. The second runs position equalization without textual turn labels, which shows how much cross-turn order matters.

Fairness of comparison. All prompting methods use the same model weights, the same prompt template apart from the manipulation, greedy decoding, and the same small development set (about 10% of dialogues) for any prompt tuning. Each method gets an equal tuning budget of prompt variants.

What counts as beating a baseline. Position-equalized judging beats turn-by-turn prompting if it matches or exceeds its turn-localization accuracy at one forward pass, or exceeds it on the long-distance and translationese slices. It counts as a meaningful result only if it also lowers the misplaced-violation rate specifically, with bootstrap confidence intervals over dialogues that exclude zero.

Why not just…?

Why not just use turn-by-turn prompting and skip the mechanism? This is the real threat. If incremental judging removes the distance dependence and the misplacement errors, a practitioner needs nothing more. The proposal is built to settle this. We will plot misplacement rate against violating-turn distance for both methods. Incremental judging gives up the later turns that sometimes reveal whether an earlier turn was a violation, for example a reply that shows a greeting was taken as rude. So we predict it will trade misplacements for misses on dialogues where the violation is only identifiable from later turns, which we will tag in the data. If it wins on every slice, we will report that the positional account explains the error but the prompting fix is enough.

Why not just avoid translated data and evaluate only on native text? Native-only evaluation hides the problem rather than explaining it. The circular evaluation trap shows translated benchmarks can inflate scores for models trained on translated data. Meanwhile, the translationese study finds that models' own outputs in German are classified as translationese up to 60.2% of the time. Translationese is in the deployment distribution whether or not benchmarks contain it. The length-matched control tells us which part of the translationese penalty is distance, and so fixable by position equalization, and which part is syntactic unnaturalness, the "translation penalty," which would need rewriting the input.

Why not just inject the relevant norm rule into the prompt? The commonsense study ChatGPT Is a Knowledgeable but Inexperienced Solver reports that injecting gold knowledge can lower Social IQA accuracy. EgoNormia's retrieval of normative precedents helps in its video setting, but there over 70% of errors were reasoning rather than perception. Injection may help with wrong-norm errors. It does not obviously tell the model which turn violated the norm, and adding text lengthens the very distances we hypothesize cause misplacement. We will include gold-norm injection as a secondary condition and report its effect on each error type separately.

Why not just fine-tune a detector? ReNoVi's best fine-tuned model reaches only 60.56 category F1. ReNoVi also reports that training on synthetic data alone degraded detection because of domain shift. A fine-tuned detector also tells us nothing about why general-purpose models misplace blame. We will still run it. If it beats every inference-time method on the translationese slice, that is evidence the residual problem is lexical rather than positional.

Proposed Experiments and Decisive Checks

The first check is distributional: are misplaced violations a large share of all false alarms on NormDial and ReNoVi? If phantom violations dominate instead, the positional account is wrong at the root and we stop there.

The second check is causal. With tokens fixed, shifting a turn's position indices closer to the verdict should raise the probability that it is blamed. Masking positional heads should flatten the dependence on distance, and random-head masking should not. If position shifts with fixed content do not move attributions, the effect is not positional.

The third check decomposes translationese. We compare the drop in accuracy from native to translationese dialogues with the drop from native to length-matched native dialogues. The share explained by length is the positional component. The rest, correlated with the SVM translationese score, is the syntactic component.

The fourth check is the payoff. Position-equalized judging should reduce misplaced violations against the naive baseline. It should match or exceed turn-by-turn prompting at a fraction of the forward passes. And it should help most on long and translated dialogues. The contrast model without RoPE should show a different, and possibly weaker, distance dependence.

Results will be reported per language, per norm category, and per error type. A negative result on the second check is just as publishable: it would show that the normative-detection gap in these benchmarks lives in social reasoning rather than in attention to position, which would redirect effort toward the knowledge-side fixes the norm literature has pursued.