Identity-Blind Peers: Measuring and Preventing Bias Compounding in AI Classrooms

Online courses leave students isolated. One promising fix is to surround each learner with several AI agents, a teacher plus simulated classmates who ask questions, take notes and crack jokes. In deployed systems this makes students talk far more: in MAIC, 6 messages per module versus 1.38 with a single chatbot. Nobody has checked whether these agents treat students differently depending on who the student appears to be. Students reveal identity constantly, through their names, the school they mention, their dialect, or a disclosed disability. Language models are known to change their reasoning and tone in response to such cues. In a classroom with five agents that read and echo one another, a small skew in one agent can be repeated and reinforced by the others. The very feature that restores social presence, a group of voices that agree and build on each other, may also be the channel through which bias grows.

This note proposes two things. The first is a counterfactual classroom audit. We replay identical tutoring conversations while swapping only the student's identity cues, and measure whether differential treatment grows with the number of agents and turns. The second is a training method for the classmate agents. Each agent gets its own role adapter, trained so that its outputs are the same under identity swaps, and the adapter is restricted so it cannot write identity information into the model's internal state. The adapter is a LoRA module: a small add-on that changes a frozen model's behavior through low-rank weight updates. The central empirical question is whether the restriction adds anything beyond training on identity-swapped data. That question is designed to be answerable either way.

Core Claim

The bias risk in multi-agent classrooms differs from the one studied so far. Prior work assigns demographic personas to the agents themselves. Bias Runs Deep shows that persona-assigned models lose up to 33% relative accuracy, for example under a "physically-disabled" persona, and that "don't make assumptions" prompts do not fix it. From Single to Societal shows persona labels shift group consensus, with woman personas' answers adopted 8.4% more often for identical reasoning. Classroom systems, however, give their classmates functional roles, such as SimClass's Deep Thinker and Class Clown in Simulating Classroom Education with LLM-Empowered Agents, or MAIC's Sparker and Questioner in Mapping Student-AI Interaction Dynamics. Neither system audits for bias.

We claim the live channel is the human student's identity cues. These cues pass through every agent in the room. We hypothesize that disparity grows with agent count and conversation length, because agents condition on each other's differentially toned turns. That would be a group version of the "complexity breaks guardrails" finding in Uncovering Stereotypes in LLMs.

We further claim that fine-tuning role agents on self-chat data generated by a teacher model, as Baize does, inherits that teacher's skews. Those skews include gendered emotion attribution: Angry Men, Sad Women reports Llama2-70b attributing anger to men four times as often as to women for the same event. They also include prestige halo effects, which Investigating Subtler Biases documents. A training objective that targets identity consistency directly should do better than prompting.

Research Program

Audit. We build student scripts from the student turns in KMP-Dialogue (KMP-Bench) and CIMA. Into each script we inject identity cues covering gender-marked names and self-reference, named institutions of differing prestige, dialect or non-native phrasing, and disclosed disability or age. We follow the counterbalanced swap design of From Single to Societal, so that each pair of scripts differs only in the cue.

Each script is run through two settings: a single tutor, and a five-agent classroom built after MAIC (one teacher, four functional peers), on the same backbone. For every agent turn we measure several things:

  • the pedagogical move, using KMP's principle categories (Feedback, Questioning, Explanation, Challenge);
  • the challenge level;
  • the praise rate;
  • the emotion the agent attributes to the student when the student expresses frustration;
  • the factual correctness of the agent's content.

Disparity is the gap in each measure between swapped pairs. The amplification test asks two questions. Does disparity at turn $t$ in the five-agent setting exceed single-tutor disparity? Does a peer's disparity rise after it reads a teacher turn that was itself skewed? Human raters validate a sample of the automatic labels.

Method. Each peer role $r$ gets a LoRA adapter $\Delta W_r = B_r A_r$ on a frozen Llama-3.1-8B. The adapter is trained on role dialogues plus identity-swapped copies, using this loss:

$$\mathcal{L}_r = \mathcal{L}_{\text{role}} + \lambda \, \mathbb{E}_{(x,x')}\, \mathrm{KL}\big(p_r(\cdot \mid x)\,\|\,p_r(\cdot \mid x')\big), \quad \text{subject to } U^\top B_r = 0.$$

The symbols are defined as follows:

  • $\mathcal{L}_{\text{role}}$ is the next-token loss on dialogues for role $r$.
  • $x$ and $x'$ are the same context with the student's identity cue swapped.
  • $\lambda$ weights the consistency term.
  • $U$ holds the top principal directions of the hidden-state differences $h(x)-h(x')$ at each adapted layer.

The constraint is enforced by projecting $B_r$ after each step. It means the adapter may read identity, which it needs in order to cancel the base model's skew, but it cannot write along identity directions. Without that write path, it cannot make downstream layers or downstream agents encode identity more strongly. This borrows the subspace-estimation step of OSRM, which constrains LoRA's input projection to be orthogonal to other tasks' features in order to prevent merging interference. We repurpose it for an output-side, identity-specific constraint that OSRM never considered.

We put the constraint on $B$ rather than $A$ deliberately. Blinding the adapter's input would leave the frozen base model's bias untouched and remove the adapter's ability to correct it. We use vanilla LoRA because A Unified Study of LoRA Variants finds it matches more complex variants when the learning rate is swept. It also suits five per-role adapters on one GPU. Low-rank tuning also tends to preserve calibration better than full tuning (On the Calibration of LLMs and Alignment), which matters for tutors that must signal uncertainty honestly.

Baselines

The naive reference is the five-agent classroom driven by role system prompts alone, with no debiasing. This reproduces how SimClass and MAIC are built, and it establishes the disparity level we must reduce. We have found no reported identity-swap disparity in tutoring on Lacuna. The closest interaction-level numbers are the 8.4% consensus gap and the 6.7% trustworthiness variation in From Single to Societal, measured on GPQA debates rather than tutoring.

The prompt-debiasing baseline adds explicit neutrality instructions to every agent. It is the cheapest fix practitioners would try. Bias Runs Deep reports that such instructions fail for persona-induced bias, but they have not been tested on student-side cues.

The strongest competitor is counterfactual-consistency LoRA without the subspace constraint. It uses the same data, rank, $\lambda$, learning-rate sweep and GPU hours as our method. This is the ablation that isolates the key ingredient. Beating it requires two things across five seeds. First, a statistically significant reduction in disparity accumulated at the classroom level, in the five-agent setting at late turns. Second, no loss in pedagogy, meaning KMP-Bench accuracy within one point of the unconstrained adapter.

For context on pedagogical quality, the best KMP-Bench overall accuracy reported is 73.1%, by DeepSeek-V3 (KMP-Bench). We use this as a ceiling reference, not a target for an 8B model. Our real pedagogy comparison is against the prompted-classroom baseline on the same backbone.

To check reasoning retention, we use GSM8K, on which LoRAFactory reports 75.51% for vanilla LoRA on Llama-3.1-8B. We also run BBQ ambiguous contexts (BBQ) as a general check that bias is not simply relocated.

Why not just…?

Why not just give peers no demographic personas? We already don't. All peers have functional roles. The point is that bias enters through the human student's cues, and removing agent personas does nothing about that. If the audit finds no student-cue disparity in the prompted classroom, this proposal reduces to a negative result. That outcome would itself be worth reporting, given that no system has checked.

Why not just scrub identity cues from student messages before agents see them? This is a real threat. We include a named-entity and self-disclosure scrubber as a baseline. We expect it to fail on cues that cannot be cleanly removed, such as dialect, non-native phrasing and writing fluency. We also expect it to remove legitimate personalization: an agent should know a student is a non-native speaker in order to adapt its explanations. We measure residual disparity on unscrubbable cues, and we measure the drop in KMP pedagogy scores when disclosures relevant to personalization are deleted. If scrubbing matches our method on both, we will recommend scrubbing.

Why not just use a single tutor? Removing classmates cuts student speech by 26.5% to 45.2% (SimClass), and with them the gains for low-prior-knowledge students reported in MAIC. Our audit directly measures whether multiple agents compound disparity. If they do not, the multi-agent concern collapses into ordinary single-tutor fairness, and we will say so.

Why not just full fine-tuning on swapped data? We run it under equal data and tuning budgets. It cannot practically host five separate role models. It also risks the calibration damage reported after synthetic-dialogue tuning, where ECE rises above 0.20 (Xu). If it wins on disparity, that is informative about capacity, and we report ECE alongside.

Decisive Checks

Three results settle the proposal.

The first is the amplification curve. Does swap-pair disparity rise with turn index faster in the five-agent classroom than with the single tutor? Does a teacher turn injected with a skew raise peer disparity in the turns that follow?

The second is the ablation. Does the output-side constraint reduce late-turn classroom disparity beyond the consistency loss alone, at matched pedagogy? A useful diagnostic here is the linear decodability of identity from downstream agents' hidden states, which the constraint is meant to cap.

The third is a small human pilot, run only if the first two succeed. Learners would use either the prompted classroom or the constrained-adapter classroom. We would check that messages per module stays near MAIC's levels, so that fairness is not bought by flattening the agents into blandness and losing the social presence that motivated building a classroom in the first place.