A common way to build a cheap content-moderation model is to have a large language model (the "teacher") label conversations and then train a small classifier (the "student") on those labels. The teacher usually reads more of the conversation than the student will see at deployment. We propose that this gap is itself a source of inherited bias. When the teacher's label depends on a turn the student never sees, the student is trained to predict something it has no evidence for. It can only fit that label through shortcuts such as slurs, keywords or style. We expect this problem to be worst for fine-grained labels, for example telling counter-speech from hate or defamation from legal insult, because those distinctions depend most on earlier turns.
The core claim is testable. Take the items whose teacher label changes when the teacher's window shrinks to the student's window. A student trained on long-window labels but given short-window input should show higher error and worse calibration on those items. The damage should also spread to items elsewhere through learned shortcuts, and it should be larger for multi-class labels than for the binary collapse of the same labels. We also propose a cheap fix, flip-aware distillation. The teacher labels every item twice, once at its own window and once at the student's. Wherever the two labels disagree, the student's training target is made consistent with what the student can actually see.
What the literature establishes and leaves open
Three strands of prior work motivate the proposal.
The first is about context windows. Performance on multi-turn tasks rises and then falls as more history is added (an "inverted U"). Two causes are usually given: attention spreads across too many tokens, and old turns drift off topic. Context-Aware Toxicity Detection in Multiplayer Games trains a domain-adapted distilRoBERTa on game chat. It finds that 67–89% of toxicity is identifiable only with earlier interactions, yet full ten-player history hurt relative to the sender's own history, which the authors attribute to overfitting on small labeled data. So neither "shortest" nor "longest" is the right student window.
The second strand shows that hiding context flips judgments. Hate Speech and Counter Speech Detection: Conversational Context Does Matter annotated 6,846 Reddit parent–target pairs twice, once with the target alone and once with the parent visible. Labels differed in 38.3% of cases, and 55.1% of counter-hate was mislabeled neutral without the parent. This is a human version of the mechanism we study. That paper measures label flips; it does not ask what happens when flipped labels become training targets for a model that sees less.
The third strand documents bias inheritance in distillation. Students copy teacher hallucinations in rationales (Okay, Let's Do This!). They copy stylistic narrowness in synthetic dialogue (Conversation Chronicles) and cultural bias in figurative generation (ViPE). Synthetic judges are lenient and favor their own model family (Towards Understanding Bias in Synthetic Data for Evaluation). All of these treat inheritance as a property of teacher content. None treats it as a property of the mismatch between the evidence the teacher conditioned on and the evidence the student receives. That structural source is our target.
Two papers constrain the design. Social Bias Frames shows that detecting offensiveness (~80.5% F1) is much easier than recovering the implied stereotype, and that generators fall back on generic stereotypes. We therefore evaluate categorical fine-grained labels rather than free-text rationales, so that label quality can be measured directly. A Novel Dataset for Classifying German Hate Speech Comments with Criminal Relevance maps comments to German Criminal Code provisions. It finds LLAMA-70B with 10 examples in the prompt beats the same model given explicit statutory definitions. We therefore run the teacher with examples, not rulebooks.
Domain knowledge enters as the application regime. Game vernacular and statutory categories are where a general teacher is least reliable. They are also where a student needs domain-adaptive pretraining, meaning continued masked-language-model training on in-domain text.
Core Claim
Let $K_T$ be the number of preceding turns the teacher reads and $K_S$ the number the student reads. Let $y_T(x)$ and $y_S(x)$ be the teacher's labels for item $x$ at those two windows. The flip set is
$$F = \{x : y_T(x) \neq y_S(x)\}.$$
Our first prediction is that training a student at window $K_S$ on labels $y_T$ produces errors concentrated on $F$, measured against human context-aware gold labels. The second is that these errors come with overconfidence on $F$. The third is that the student's reliance on lexical cues rises on non-flip items too. The fourth is that $|F|$ and the resulting damage grow with label granularity: they should be larger for three-way hate, counter-speech and neutral labels than for hate versus not-hate on the same items.
Flip-aware distillation replaces $y_T$ on $F$ with one of three targets. The first is the short-window teacher label $y_S$. The second is a soft mixture of the two labels. The third keeps $y_T$ and appends the single history turn the teacher names as decisive. We claim that at least one of these beats both naive long-window distillation and matched-window distillation at equal student compute.
Research Program
The program has three stages.
The first stage is measurement. We run an open teacher, LLaMA-70B-class and prompted with examples as in the German dataset paper, at window sizes 0, 1, 2, 4 and 8 preceding turns wherever threads allow. We record the flip rate per label dimension and compare it against the human 38.3% flip rate on the Reddit data. This tells us whether the teacher's sensitivity to context resembles that of humans.
The second stage is the distillation matrix. We train students over the full grid of teacher window versus student window. Students are RoBERTa-base on Reddit, domain-adapted distilRoBERTa on game chat, and German BERT on the statutory data. We break error down into flip and non-flip items. We measure shortcut reliance with a counterfactual probe: history turns are swapped for turns from unrelated threads, and a student that ignores its context shows no change.
The third stage is the fix. We compare the three flip-aware variants against the baselines below.
Baselines
The naive reference point is a student trained on human gold labels with the target comment alone. It shows what context and distillation add at all.
The strongest existing student method is the parent–target concatenated RoBERTa of Hate Speech and Counter Speech Detection. It reports weighted F1 of 0.64, up from 0.58 without context. On game chat, the strongest existing method is domain-adapted distilRoBERTa with player and team metadata tokens from Context-Aware Toxicity Detection in Multiplayer Games. It reports balanced accuracy of 0.77, versus 0.65 for a vanilla model, with metadata adding about 0.06. These are strong because they already use context and domain adaptation and are trained on human labels. We re-run both on our splits with the same architecture, the same number of training steps, and an equal hyperparameter search budget of 20 trials each.
The teacher itself is the accuracy ceiling. On the German statutory data, LLAMA-70B with 10 examples reports Macro F1 of 0.85230 (German criminal-relevance paper). We report how much of the gap to this ceiling each student closes.
The two distillation baselines are naive distillation, with teacher labels at the teacher's best window and student input at the student's window, and matched-window distillation, where both use the same window. Both use the same teacher calls as our method.
Two ablations isolate the key ingredient. In the first, we remove labels from a random subset the same size as $F$. In the second, we remove the items where the teacher is least confident, again matched in size to $F$. If flip-aware distillation does no better than confidence filtering, the flip set carries no special information and the claim fails.
Beating the baselines means two things at the same student window. Flip-aware distillation must exceed naive distillation on gold macro F1 over $F$, with bootstrap 95% intervals across 5 seeds excluding zero. It must also match or exceed naive distillation on non-flip items.
Why not just…?
Why not just grid-search a single window for the student? A single window tuned on validation data gives one window for all label dimensions and all items. It cannot remove the conflict between labels that needed the parent and inputs that lack it. The matched-window cells of our matrix are exactly this search. If the best matched cell equals flip-aware distillation, we will report that the mismatch story is unnecessary.
Why not just give the student the full history? The game-chat paper found that full multi-player history hurt a small classifier relative to sender-only history. The inverted-U evidence points the same way. Our matrix includes a student with the largest window, so this is tested directly rather than assumed.
Why not just train on human labels? Where human labels exist in quantity, this is the right answer, and our first strong baseline does exactly this. Distillation matters for the domains in the file where expert labels are scarce: the statutory dataset has 1,090 comments and required prosecutors to label them. We therefore run a learning-curve comparison of human labels against distilled labels at 10%, 30% and 100% of the gold data. This locates the point where distillation stops paying off.
Why not just filter low-confidence teacher labels? This is the most serious competitor, because flipped items may simply be the items the teacher is uncertain about. The confidence-filtering ablation settles it. We also report how much the flip set and the low-confidence set overlap.
Proposed Experiments
The primary testbed is the Reddit context dataset from Hate Speech and Counter Speech Detection. It has human labels under both windows, which lets us validate teacher flips against human flips. Where the original threads can be retrieved, we extend it past one parent turn to test deeper windows. Gold labels for those deeper windows would still be the parent-level labels, which we will state as a limitation.
The second testbed is the DOTA 2 chat of Context-Aware Toxicity Detection in Multiplayer Games, if it is accessible. Here the windows are sender-only history versus all-player history, which tests mismatch across speakers rather than across time. Its gold labels are binary, so this testbed serves as the binary comparison.
The third testbed is the German statutory dataset. For its DeTox-derived comments, we will reconstruct thread context from DeTox if that corpus retains conversation structure. This yields the first statutory classification experiment that uses context. The paper's weak Class 3 (defamation and insult) is where we most expect flips.
SBIC provides a negative control. Its posts are isolated, so there is no history to mismatch, and flip-aware distillation should give no gain there.
Four results would be decisive. The first is whether student error on $F$ exceeds error on non-flip items under naive distillation. The second is whether that excess is larger for the three-way labels than for their binary collapse. The third is whether the counterfactual history-swap probe shows reduced use of context in naively distilled students. The fourth is whether flip-aware distillation beats both the matched-window baseline and the confidence-filtering ablation. A null result on the first test would falsify the central claim.
References
- Context-Aware Toxicity Detection in Multiplayer Games: Integrating Domain-Adaptive Pretraining and Match Metadata
- Hate Speech and Counter Speech Detection: Conversational Context Does Matter
- Okay, Let's Do This! Modeling Event Coreference with Generated Rationales and Knowledge Distillation
- Conversation Chronicles: Towards Diverse Temporal and Relational Dynamics in Multi-Session Conversations
- ViPE: Visualise Pretty-much Everything
- Towards Understanding Bias in Synthetic Data for Evaluation
- Social Bias Frames: Reasoning about Social and Power Implications of Language
- A Novel Dataset for Classifying German Hate Speech Comments with Criminal Relevance
- DeTox: A Comprehensive Dataset for German Offensive Language and Conversation Analysis